USE CASE · AI AGENT DELIVERY · AGENTFORCE AND BEYOND
AI agents in production, not in pilot. Because yours pass an exam first.
Anyone can demo an agent. Hapie delivers them: requirements and guardrails written down, an exam of hundreds of realistic scenarios it must pass, a named person approving the launch, every action logged in production, and a re-exam every time the underlying model changes. Into Agentforce, into your platforms, on your terms.
THE FOUR THINGS EVERY AGENT PROGRAMME RECOGNISES
Why agents wow in March and are still "two weeks away" in September.
If your organisation has piloted an AI agent, you will recognise these.
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
01
the-lost-weekend
"Three senior architects, two weekends, one deadline. The people who bill the most, writing documents that usually lose."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
02
the-copy-paste-bid.
"Last year's response with the client's name changed. Except the one place it wasn't."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
03
the-promise-delivery-discovered-later.
"Somewhere on page forty, we committed to something. Delivery found out after we won."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
04
the-loss-with-no-lesson.
"We lost. Procurement said 'scoring'. Nobody knows which answer cost us, so the next bid repeats it."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
01
the-specialist-queue.
"Good MuleSoft people are scarce and expensive. Your whole programme moves at the speed of the three you have."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
02
the-api-that-already-existed.
"Nobody checks the catalogue under deadline pressure. Three teams have now built the same customer lookup, three ways."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
03
the-rules-that-live-in-a-wiki.
"The three-layer rules and naming standards are written down beautifully. On a deadline, they lose."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
04
the-handover-that-stops-at-go-live.
"The delivery partner rolled off. The documentation stops at go-live, and the people who knew why things were built that way went with them."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
01
the-quote-that-starts-with-discovery.
"The vendor's plan begins with a nine-month discovery phase, before anything moves, because nobody can say what is actually on the old platform."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
02
the-list-nobody-has.
"Ask three people how many connections run on the old platform and you get three numbers, hundreds apart."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
03
the-ones-that-might-be-dead.
"Some connections run every night for nobody. You keep paying for them, because switching one off is scarier than keeping it alive."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
04
the-last-attempt.
"The previous migration stalled partway. The people who ran it have moved on, and nobody is sure which connections were finished."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
01
lost-in-translation.
"What the business asked for and what got built are cousins, not twins. Nobody can say where it drifted, because nothing was written down with sources."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
02
the-prototype-that-shipped.
"The demo impressed everyone, so it quietly became the product. No error handling, no tests, no documentation. Three years later, everyone is afraid of it."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
03
the-creeping-scope.
"Every sprint adds a little. By month six the budget is spent and the core still is not finished."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
04
the-documentation-debt.
"The documentation was written in the last week, from memory, by whoever was left."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
01
pilot-purgatory.
"The demo wowed the board in March. It is September, and launch is still two weeks away, because nobody can say what would make it safe to say yes."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
02
the-improviser.
"In the demo it answered perfectly. Nobody knows what it says when a customer swears at it, lies to it, or asks it for a discount."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
03
the-3x-estimate.-copy
'What is our AI agent strategy?' Every vendor answers 'buy ours'. Nobody answers the real question: how would we trust one?
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
04
the-unowned-promise.
"If the agent promises a customer a refund, whose promise is it? Today, nobody has signed for that."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
01
the-out-of-date-chart.
"The chart showing how your systems connect was last correct in 2021. Three of the people on it have left the company."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
02
the-connection-nobody-dares-touch.
"One old connection moves payroll data. The person who built it left. Nobody touches it, so every project builds an expensive detour around it."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
03
the-3x-estimate.
"A two-week change is quoted at ten weeks, because eight of them go on working out what it might break."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
04
the-shelf.
"The improvements you wanted in 2024 are still a slide in a deck. Not for lack of value, but because changing running systems feels too risky and too expensive."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
01
the-deck-that-never-became-delivery.
"The last assessment cost a fortune, took five months, and lives in a drawer. Beautiful slides. Nothing shipped."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
02
the-interview-truth
"It was built from forty interviews. People told what they remembered, defended what they built, and nobody mentioned the system nobody owns."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
03
the-assessment-that-expired.
By the time decisions were made, the estate had moved on. The document described last year."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
04
the-roadmap-with-no-numbers-behind-it
"The plan said eighteen months and a big figure. Asked how it was calculated, the consultants said: experience."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
01
the-object-nobody-can-read.
Hundreds of fields added over a decade. Nobody knows which are used, so nobody dares remove them.
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
02
the-friday-release-that-broke-monday.
A new automation collided with an old one on the same records.
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
03
the-custom-code-nobody-remembers-why.
Custom code now does what a standard feature can do out of the box.
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
04
the-queue-behind-three-admins.
Every department’s asks funnel into one backlog behind the same few people.
flip card
THE EXAM
A demo shows what an agent can do. The exam shows what it will do.
Before any agent meets a customer, Hapie builds its exam: hundreds of realistic scenarios drawn from your actual work: the routine asks, the ambiguous ones, the angry customer, the customer who asks for something the agent must not give, and the one who tries to trick it. Every scenario has a defined right behaviour, agreed with your team. The agent runs the whole exam, is scored on accuracy, safety, staying inside its guardrails, knowing when to hand over, and cost per interaction, and the results go to a named person. No pass, no licence. And the exam is kept: it runs again on every change.
"You would not let a new employee near customers without training and supervision. An agent gets the same, in writing."
THE TO-BE, AS OPTIONS WITH TRADE-OFFS
Not one recommendation. Options, priced in trade-offs, for you to choose.
From the as-is, Hapie lays out target options: typically a steady path (fix the riskiest things, keep the shape), a balanced path (consolidate and modernise in waves), and a bold path (replatform what matters most). Each option states, with its evidence: what changes, what it makes possible, what it risks, roughly what it takes, drawn from measured past deliveries, and what it deliberately leaves alone. The choice is yours; the evidence is the same for every option.
Genuinely custom. Built to the integration pattern, documented, and added to the map for the next team.
Genuinely custom. Built to the integration pattern, documented, and added to the map for the next team.
Genuinely custom. Built to the integration pattern, documented, and added to the map for the next team.
Genuinely custom. Built to the integration pattern, documented, and added to the map for the next team.
Genuinely custom. Built to the integration pattern, documented, and added to the map for the next team.
Genuinely custom. Built to the integration pattern, documented, and added to the map for the next team.
Genuinely custom. Built to the integration pattern, documented, and added to the map for the next team.
Genuinely custom. Built to the integration pattern, documented, and added to the map for the next team.
Genuinely custom. Built to the integration pattern, documented, and added to the map for the next team.
Genuinely custom. Built to the integration pattern, documented, and added to the map for the next team.
ONE AGENT, DELIVERED
From job description to working agent, with proof at every step.
One real agent, walked end to end: an order-status agent in Agentforce.
The agent gets a job description.
What it is for, who it serves, what good looks like, and its guardrails, written and approved before anything is built, like any role worth hiring for.
The exam is built with your team.
Bring one request from your backlog. Leave with it scoped.
The agent is built into your platform.
Into Agentforce here, connected only to the systems its guardrails allow, with its logging wired in from the start.
It sits the exam.
The full scenario set, scored. Weaknesses are fixed and it resits. No pass, no licence: see the checkpoint moment below.
A named person signs the launch.
The licence: exam results, guardrails, monitoring plan and rollback path, approved in one place. It launches to a limited group first.
It works, watched, and gets better.
Every action logged, hand-overs reviewed, the exam grows from real conversations, and every model change triggers a re-exam before the agent runs on it.
THE EXAM
A demo shows what an agent can do. The exam shows what it will do.
Before any agent meets a customer, Hapie builds its exam: hundreds of realistic scenarios drawn from your actual work: the routine asks, the ambiguous ones, the angry customer, the customer who asks for something the agent must not give, and the one who tries to trick it. Every scenario has a defined right behaviour, agreed with your team. The agent runs the whole exam, is scored on accuracy, safety, staying inside its guardrails, knowing when to hand over, and cost per interaction, and the results go to a named person. No pass, no licence. And the exam is kept: it runs again on every change.
"You would not let a new employee near customers without training and supervision. An agent gets the same, in writing."
Start
Bring one request from your backlog. Leave with it scoped.
Fifteen minutes on an org like yours: see the map, make the standard-before-custom call, and see the collision check in action.
Read-only to start · Ships through your own release process · Reply within two working days
The three doubts, answered straight
What Salesforce leaders ask.
Is the 95% failure statistic real?
The study is real, the number is loose. MIT's Project NANDA published The GenAI Divide in July 2025 based on 52 interviews and 153 survey responses. Its traceable finding is that 60% of organisations evaluated these tools, 20% piloted and 5% reached production. The "95% get zero return" framing has been publicly questioned by academics for not being derivable from the document.
If agents only succeed 35% of the time on multi-turn tasks, why use them?
Some of them genuinely do, and their judgement matters: that is the context step, and it is welcome. What no person can do is hold two hundred systems' connections in their head, current, at once. Evidence from the systems plus judgement from people beats either alone, and this page is the first half done properly.
How long does it take to get a first agent into production properly?
Someone accountable for the outcome in the estate, not someone accountable for the platform. In practice that is usually the person who would be asked to explain the change to a regulator. If nobody can be named in a sentence, that is the finding, and it is worth surfacing before an auditor surfaces it.
Does the EU AI Act apply to us if we only use agents internally?
Yes. The map is built from the org as it is, debt included, turning cleanup into ordinary evidenced changes.
Is Salesforce Shield enough on its own?
Some of them genuinely do, and their judgement matters: that is the context step, and it is welcome. What no person can do is hold two hundred systems' connections in their head, current, at once. Evidence from the systems plus judgement from people beats either alone, and this page is the first half done properly.
Who should be the named approver for agent work?
Someone accountable for the outcome in the estate, not someone accountable for the platform. In practice that is usually the person who would be asked to explain the change to a regulator. If nobody can be named in a sentence, that is the finding, and it is worth surfacing before an auditor surfaces it.
Is this not just a traceability matrix?
Yes. The map is built from the org as it is, debt included, turning cleanup into ordinary evidenced changes.
Does this slow delivery down by adding approval steps?
Some of them genuinely do, and their judgement matters: that is the context step, and it is welcome. What no person can do is hold two hundred systems' connections in their head, current, at once. Evidence from the systems plus judgement from people beats either alone, and this page is the first half done properly.
How does this work on a brownfield estate with ten years of history?
That is the harder and more common case. The map is built from what already exists, including the parts nobody documented. Long-lived estates usually see the biggest difference, because reconstruction cost per transition is highest where the decisions are oldest.
"Our bids are our secret sauce."
Yes. The map is built from the org as it is, debt included, turning cleanup into ordinary evidenced changes.
"Every RFP is different."
Some of them genuinely do, and their judgement matters: that is the context step, and it is welcome. What no person can do is hold two hundred systems' connections in their head, current, at once. Evidence from the systems plus judgement from people beats either alone, and this page is the first half done properly.
"Will it read like AI wrote it?"
The opposite, and here is why: what makes bids read machine-made is generic filler, and filler exists because humans padding forty pages at midnight reach for it. Here, every sentence must attach to your record or it does not survive the checkpoint, so what remains is specific, in your voice, and signed by your people. The blandest bid your buyer will read this quarter was written by exhausted humans.
"Our MuleSoft estate is already a mess."
Yes. The map is built from the org as it is, debt included, turning cleanup into ordinary evidenced changes.
"We already have accelerators and templates."
Some of them genuinely do, and their judgement matters: that is the context step, and it is welcome. What no person can do is hold two hundred systems' connections in their head, current, at once. Evidence from the systems plus judgement from people beats either alone, and this page is the first half done properly.
"Will it fit our Anypoint setup and our release process?"
Hapie reads your Anypoint estate to build the map and the catalogue view, and everything ships through your own pipelines and approvals. Your setup stays your setup; Hapie supplies the method and the proof.
"We tried this before and it stalled."
Yes. The map is built from the org as it is, debt included, turning cleanup into ordinary evidenced changes.
"We cannot freeze the business for the duration."
Some of them genuinely do, and their judgement matters: that is the context step, and it is welcome. What no person can do is hold two hundred systems' connections in their head, current, at once. Evidence from the systems plus judgement from people beats either alone, and this page is the first half done properly.
"What if the new platform behaves differently?"
That is what the side-by-side run exists to catch. Every connection processes the same real inputs on both platforms until the outputs match, line by line, and the quiet edge cases, rounding, empty fields, odd dates, are exactly what it makes loud.
"AI tools already build apps from a prompt. Why is this different?"
Yes. The map is built from the org as it is, debt included, turning cleanup into ordinary evidenced changes.
"Our requirements always change mid-build."
Some of them genuinely do, and their judgement matters: that is the context step, and it is welcome. What no person can do is hold two hundred systems' connections in their head, current, at once. Evidence from the systems plus judgement from people beats either alone, and this page is the first half done properly.
"We do not have people free to write requirements for months."
You do not write them; you approve them. The agents draft requirements from the workshops and documents you already have, and your people answer a short open-questions list and sign the scope. The heavy documentation lifting is exactly what is automated.
"We are already building in Agentforce."
Yes. The map is built from the org as it is, debt included, turning cleanup into ordinary evidenced changes.
"Models change constantly. Won't the agent drift?"
Some of them genuinely do, and their judgement matters: that is the context step, and it is welcome. What no person can do is hold two hundred systems' connections in their head, current, at once. Evidence from the systems plus judgement from people beats either alone, and this page is the first half done properly.
"Our legal and compliance teams will never sign this."
They will not sign vibes, and they are right not to. What they can sign is what this process manufactures: written limits with named owners, exam results against agreed behaviours, complete action logs, and a rollback path. Several sections of this site, Security, and Documents you share with us, are written to be sent to them first. Bring them to the demo; they are the easiest audience we have.
"Our systems are a mess. No tool can read them."
Yes. The map is built from the org as it is, debt included, turning cleanup into ordinary evidenced changes.
"We cannot let an AI change systems that are running.
Some of them genuinely do, and their judgement matters: that is the context step, and it is welcome. What no person can do is hold two hundred systems' connections in their head, current, at once. Evidence from the systems plus judgement from people beats either alone, and this page is the first half done properly.
"The people who understood the old parts have left."
Much of what they knew is still in what they left behind: the code, the tickets, the history of changes. Hapie recovers that part. What cannot be recovered is confirmed once by your team, then kept on the map, so it is never lost again.
"We did an assessment last year."
Yes. The map is built from the org as it is, debt included, turning cleanup into ordinary evidenced changes.
"Our consultants already know our estate"
Some of them genuinely do, and their judgement matters: that is the context step, and it is welcome. What no person can do is hold two hundred systems' connections in their head, current, at once. Evidence from the systems plus judgement from people beats either alone, and this page is the first half done properly.
"We cannot open our systems to an outside tool."
It is read-only, it runs in your environment, and building the as-is changes nothing. The access conversation is with your security team on your terms, and the Documents and Security pages on this site are written to be sent to them first.
“Our org is a decade of debt. Start there?”
Yes. The map is built from the org as it is, debt included, turning cleanup into ordinary evidenced changes.
“We already have admins and a partner.”
Keep them. They become the approvers and standard-setters while the agents do the building.
“Will this fight our release process?”
No. It ships through your sandboxes, pipelines and release calendar. What it adds is the map and checkpoint.
USE CASE · AI AGENT DELIVERY · AGENTFORCE AND BEYOND
AI agents in production, not in pilot. Because yours pass an exam first.
Anyone can demo an agent. Hapie delivers them: requirements and guardrails written down, an exam of hundreds of realistic scenarios it must pass, a named person approving the launch, every action logged in production, and a re-exam every time the underlying model changes. Into Agentforce, into your platforms, on your terms.
THE FOUR THINGS EVERY AGENT PROGRAMME RECOGNISES
Why agents wow in March and are still "two weeks away" in September.
If your organisation has piloted an AI agent, you will recognise these.
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
01
the-lost-weekend
"Three senior architects, two weekends, one deadline. The people who bill the most, writing documents that usually lose."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
02
the-copy-paste-bid.
"Last year's response with the client's name changed. Except the one place it wasn't."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
03
the-promise-delivery-discovered-later.
"Somewhere on page forty, we committed to something. Delivery found out after we won."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
04
the-loss-with-no-lesson.
"We lost. Procurement said 'scoring'. Nobody knows which answer cost us, so the next bid repeats it."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
01
the-specialist-queue.
"Good MuleSoft people are scarce and expensive. Your whole programme moves at the speed of the three you have."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
02
the-api-that-already-existed.
"Nobody checks the catalogue under deadline pressure. Three teams have now built the same customer lookup, three ways."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
03
the-rules-that-live-in-a-wiki.
"The three-layer rules and naming standards are written down beautifully. On a deadline, they lose."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
04
the-handover-that-stops-at-go-live.
"The delivery partner rolled off. The documentation stops at go-live, and the people who knew why things were built that way went with them."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
01
the-quote-that-starts-with-discovery.
"The vendor's plan begins with a nine-month discovery phase, before anything moves, because nobody can say what is actually on the old platform."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
02
the-list-nobody-has.
"Ask three people how many connections run on the old platform and you get three numbers, hundreds apart."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
03
the-ones-that-might-be-dead.
"Some connections run every night for nobody. You keep paying for them, because switching one off is scarier than keeping it alive."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
04
the-last-attempt.
"The previous migration stalled partway. The people who ran it have moved on, and nobody is sure which connections were finished."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
01
lost-in-translation.
"What the business asked for and what got built are cousins, not twins. Nobody can say where it drifted, because nothing was written down with sources."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
02
the-prototype-that-shipped.
"The demo impressed everyone, so it quietly became the product. No error handling, no tests, no documentation. Three years later, everyone is afraid of it."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
03
the-creeping-scope.
"Every sprint adds a little. By month six the budget is spent and the core still is not finished."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
04
the-documentation-debt.
"The documentation was written in the last week, from memory, by whoever was left."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
01
pilot-purgatory.
"The demo wowed the board in March. It is September, and launch is still two weeks away, because nobody can say what would make it safe to say yes."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
02
the-improviser.
"In the demo it answered perfectly. Nobody knows what it says when a customer swears at it, lies to it, or asks it for a discount."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
03
the-3x-estimate.-copy
'What is our AI agent strategy?' Every vendor answers 'buy ours'. Nobody answers the real question: how would we trust one?
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
04
the-unowned-promise.
"If the agent promises a customer a refund, whose promise is it? Today, nobody has signed for that."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
01
the-out-of-date-chart.
"The chart showing how your systems connect was last correct in 2021. Three of the people on it have left the company."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
02
the-connection-nobody-dares-touch.
"One old connection moves payroll data. The person who built it left. Nobody touches it, so every project builds an expensive detour around it."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
03
the-3x-estimate.
"A two-week change is quoted at ten weeks, because eight of them go on working out what it might break."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
04
the-shelf.
"The improvements you wanted in 2024 are still a slide in a deck. Not for lack of value, but because changing running systems feels too risky and too expensive."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
01
the-deck-that-never-became-delivery.
"The last assessment cost a fortune, took five months, and lives in a drawer. Beautiful slides. Nothing shipped."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
02
the-interview-truth
"It was built from forty interviews. People told what they remembered, defended what they built, and nobody mentioned the system nobody owns."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
03
the-assessment-that-expired.
By the time decisions were made, the estate had moved on. The document described last year."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
04
the-roadmap-with-no-numbers-behind-it
"The plan said eighteen months and a big figure. Asked how it was calculated, the consultants said: experience."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
01
the-object-nobody-can-read.
Hundreds of fields added over a decade. Nobody knows which are used, so nobody dares remove them.
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
02
the-friday-release-that-broke-monday.
A new automation collided with an old one on the same records.
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
03
the-custom-code-nobody-remembers-why.
Custom code now does what a standard feature can do out of the box.
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
04
the-queue-behind-three-admins.
Every department’s asks funnel into one backlog behind the same few people.
flip card
THE EXAM
A demo shows what an agent can do. The exam shows what it will do.
Before any agent meets a customer, Hapie builds its exam: hundreds of realistic scenarios drawn from your actual work: the routine asks, the ambiguous ones, the angry customer, the customer who asks for something the agent must not give, and the one who tries to trick it. Every scenario has a defined right behaviour, agreed with your team. The agent runs the whole exam, is scored on accuracy, safety, staying inside its guardrails, knowing when to hand over, and cost per interaction, and the results go to a named person. No pass, no licence. And the exam is kept: it runs again on every change.
"You would not let a new employee near customers without training and supervision. An agent gets the same, in writing."
THE TO-BE, AS OPTIONS WITH TRADE-OFFS
Not one recommendation. Options, priced in trade-offs, for you to choose.
From the as-is, Hapie lays out target options: typically a steady path (fix the riskiest things, keep the shape), a balanced path (consolidate and modernise in waves), and a bold path (replatform what matters most). Each option states, with its evidence: what changes, what it makes possible, what it risks, roughly what it takes, drawn from measured past deliveries, and what it deliberately leaves alone. The choice is yours; the evidence is the same for every option.
Genuinely custom. Built to the integration pattern, documented, and added to the map for the next team.
Genuinely custom. Built to the integration pattern, documented, and added to the map for the next team.
Genuinely custom. Built to the integration pattern, documented, and added to the map for the next team.
Genuinely custom. Built to the integration pattern, documented, and added to the map for the next team.
Genuinely custom. Built to the integration pattern, documented, and added to the map for the next team.
Genuinely custom. Built to the integration pattern, documented, and added to the map for the next team.
Genuinely custom. Built to the integration pattern, documented, and added to the map for the next team.
Genuinely custom. Built to the integration pattern, documented, and added to the map for the next team.
Genuinely custom. Built to the integration pattern, documented, and added to the map for the next team.
Genuinely custom. Built to the integration pattern, documented, and added to the map for the next team.
ONE AGENT, DELIVERED
From job description to working agent, with proof at every step.
One real agent, walked end to end: an order-status agent in Agentforce.
The agent gets a job description.
What it is for, who it serves, what good looks like, and its guardrails, written and approved before anything is built, like any role worth hiring for.
The exam is built with your team.
Bring one request from your backlog. Leave with it scoped.
The agent is built into your platform.
Into Agentforce here, connected only to the systems its guardrails allow, with its logging wired in from the start.
It sits the exam.
The full scenario set, scored. Weaknesses are fixed and it resits. No pass, no licence: see the checkpoint moment below.
A named person signs the launch.
The licence: exam results, guardrails, monitoring plan and rollback path, approved in one place. It launches to a limited group first.
It works, watched, and gets better.
Every action logged, hand-overs reviewed, the exam grows from real conversations, and every model change triggers a re-exam before the agent runs on it.
THE EXAM
A demo shows what an agent can do. The exam shows what it will do.
Before any agent meets a customer, Hapie builds its exam: hundreds of realistic scenarios drawn from your actual work: the routine asks, the ambiguous ones, the angry customer, the customer who asks for something the agent must not give, and the one who tries to trick it. Every scenario has a defined right behaviour, agreed with your team. The agent runs the whole exam, is scored on accuracy, safety, staying inside its guardrails, knowing when to hand over, and cost per interaction, and the results go to a named person. No pass, no licence. And the exam is kept: it runs again on every change.
"You would not let a new employee near customers without training and supervision. An agent gets the same, in writing."
Start
Bring one request from your backlog. Leave with it scoped.
Fifteen minutes on an org like yours: see the map, make the standard-before-custom call, and see the collision check in action.
Read-only to start · Ships through your own release process · Reply within two working days
The three doubts, answered straight
What Salesforce leaders ask.
Is the 95% failure statistic real?
The study is real, the number is loose. MIT's Project NANDA published The GenAI Divide in July 2025 based on 52 interviews and 153 survey responses. Its traceable finding is that 60% of organisations evaluated these tools, 20% piloted and 5% reached production. The "95% get zero return" framing has been publicly questioned by academics for not being derivable from the document.
If agents only succeed 35% of the time on multi-turn tasks, why use them?
Some of them genuinely do, and their judgement matters: that is the context step, and it is welcome. What no person can do is hold two hundred systems' connections in their head, current, at once. Evidence from the systems plus judgement from people beats either alone, and this page is the first half done properly.
How long does it take to get a first agent into production properly?
Someone accountable for the outcome in the estate, not someone accountable for the platform. In practice that is usually the person who would be asked to explain the change to a regulator. If nobody can be named in a sentence, that is the finding, and it is worth surfacing before an auditor surfaces it.
Does the EU AI Act apply to us if we only use agents internally?
Yes. The map is built from the org as it is, debt included, turning cleanup into ordinary evidenced changes.
Is Salesforce Shield enough on its own?
Some of them genuinely do, and their judgement matters: that is the context step, and it is welcome. What no person can do is hold two hundred systems' connections in their head, current, at once. Evidence from the systems plus judgement from people beats either alone, and this page is the first half done properly.
Who should be the named approver for agent work?
Someone accountable for the outcome in the estate, not someone accountable for the platform. In practice that is usually the person who would be asked to explain the change to a regulator. If nobody can be named in a sentence, that is the finding, and it is worth surfacing before an auditor surfaces it.
Is this not just a traceability matrix?
Yes. The map is built from the org as it is, debt included, turning cleanup into ordinary evidenced changes.
Does this slow delivery down by adding approval steps?
Some of them genuinely do, and their judgement matters: that is the context step, and it is welcome. What no person can do is hold two hundred systems' connections in their head, current, at once. Evidence from the systems plus judgement from people beats either alone, and this page is the first half done properly.
How does this work on a brownfield estate with ten years of history?
That is the harder and more common case. The map is built from what already exists, including the parts nobody documented. Long-lived estates usually see the biggest difference, because reconstruction cost per transition is highest where the decisions are oldest.
"Our bids are our secret sauce."
Yes. The map is built from the org as it is, debt included, turning cleanup into ordinary evidenced changes.
"Every RFP is different."
Some of them genuinely do, and their judgement matters: that is the context step, and it is welcome. What no person can do is hold two hundred systems' connections in their head, current, at once. Evidence from the systems plus judgement from people beats either alone, and this page is the first half done properly.
"Will it read like AI wrote it?"
The opposite, and here is why: what makes bids read machine-made is generic filler, and filler exists because humans padding forty pages at midnight reach for it. Here, every sentence must attach to your record or it does not survive the checkpoint, so what remains is specific, in your voice, and signed by your people. The blandest bid your buyer will read this quarter was written by exhausted humans.
"Our MuleSoft estate is already a mess."
Yes. The map is built from the org as it is, debt included, turning cleanup into ordinary evidenced changes.
"We already have accelerators and templates."
Some of them genuinely do, and their judgement matters: that is the context step, and it is welcome. What no person can do is hold two hundred systems' connections in their head, current, at once. Evidence from the systems plus judgement from people beats either alone, and this page is the first half done properly.
"Will it fit our Anypoint setup and our release process?"
Hapie reads your Anypoint estate to build the map and the catalogue view, and everything ships through your own pipelines and approvals. Your setup stays your setup; Hapie supplies the method and the proof.
"We tried this before and it stalled."
Yes. The map is built from the org as it is, debt included, turning cleanup into ordinary evidenced changes.
"We cannot freeze the business for the duration."
Some of them genuinely do, and their judgement matters: that is the context step, and it is welcome. What no person can do is hold two hundred systems' connections in their head, current, at once. Evidence from the systems plus judgement from people beats either alone, and this page is the first half done properly.
"What if the new platform behaves differently?"
That is what the side-by-side run exists to catch. Every connection processes the same real inputs on both platforms until the outputs match, line by line, and the quiet edge cases, rounding, empty fields, odd dates, are exactly what it makes loud.
"AI tools already build apps from a prompt. Why is this different?"
Yes. The map is built from the org as it is, debt included, turning cleanup into ordinary evidenced changes.
"Our requirements always change mid-build."
Some of them genuinely do, and their judgement matters: that is the context step, and it is welcome. What no person can do is hold two hundred systems' connections in their head, current, at once. Evidence from the systems plus judgement from people beats either alone, and this page is the first half done properly.
"We do not have people free to write requirements for months."
You do not write them; you approve them. The agents draft requirements from the workshops and documents you already have, and your people answer a short open-questions list and sign the scope. The heavy documentation lifting is exactly what is automated.
"We are already building in Agentforce."
Yes. The map is built from the org as it is, debt included, turning cleanup into ordinary evidenced changes.
"Models change constantly. Won't the agent drift?"
Some of them genuinely do, and their judgement matters: that is the context step, and it is welcome. What no person can do is hold two hundred systems' connections in their head, current, at once. Evidence from the systems plus judgement from people beats either alone, and this page is the first half done properly.
"Our legal and compliance teams will never sign this."
They will not sign vibes, and they are right not to. What they can sign is what this process manufactures: written limits with named owners, exam results against agreed behaviours, complete action logs, and a rollback path. Several sections of this site, Security, and Documents you share with us, are written to be sent to them first. Bring them to the demo; they are the easiest audience we have.
"Our systems are a mess. No tool can read them."
Yes. The map is built from the org as it is, debt included, turning cleanup into ordinary evidenced changes.
"We cannot let an AI change systems that are running.
Some of them genuinely do, and their judgement matters: that is the context step, and it is welcome. What no person can do is hold two hundred systems' connections in their head, current, at once. Evidence from the systems plus judgement from people beats either alone, and this page is the first half done properly.
"The people who understood the old parts have left."
Much of what they knew is still in what they left behind: the code, the tickets, the history of changes. Hapie recovers that part. What cannot be recovered is confirmed once by your team, then kept on the map, so it is never lost again.
"We did an assessment last year."
Yes. The map is built from the org as it is, debt included, turning cleanup into ordinary evidenced changes.
"Our consultants already know our estate"
Some of them genuinely do, and their judgement matters: that is the context step, and it is welcome. What no person can do is hold two hundred systems' connections in their head, current, at once. Evidence from the systems plus judgement from people beats either alone, and this page is the first half done properly.
"We cannot open our systems to an outside tool."
It is read-only, it runs in your environment, and building the as-is changes nothing. The access conversation is with your security team on your terms, and the Documents and Security pages on this site are written to be sent to them first.
“Our org is a decade of debt. Start there?”
Yes. The map is built from the org as it is, debt included, turning cleanup into ordinary evidenced changes.
“We already have admins and a partner.”
Keep them. They become the approvers and standard-setters while the agents do the building.
“Will this fight our release process?”
No. It ships through your sandboxes, pipelines and release calendar. What it adds is the map and checkpoint.
USE CASE · AI AGENT DELIVERY · AGENTFORCE AND BEYOND
AI agents in production, not in pilot. Because yours pass an exam first.
Anyone can demo an agent. Hapie delivers them: requirements and guardrails written down, an exam of hundreds of realistic scenarios it must pass, a named person approving the launch, every action logged in production, and a re-exam every time the underlying model changes. Into Agentforce, into your platforms, on your terms.
THE FOUR THINGS EVERY AGENT PROGRAMME RECOGNISES
Why agents wow in March and are still "two weeks away" in September.
If your organisation has piloted an AI agent, you will recognise these.
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
01
the-lost-weekend
"Three senior architects, two weekends, one deadline. The people who bill the most, writing documents that usually lose."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
02
the-copy-paste-bid.
"Last year's response with the client's name changed. Except the one place it wasn't."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
03
the-promise-delivery-discovered-later.
"Somewhere on page forty, we committed to something. Delivery found out after we won."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
04
the-loss-with-no-lesson.
"We lost. Procurement said 'scoring'. Nobody knows which answer cost us, so the next bid repeats it."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
01
the-specialist-queue.
"Good MuleSoft people are scarce and expensive. Your whole programme moves at the speed of the three you have."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
02
the-api-that-already-existed.
"Nobody checks the catalogue under deadline pressure. Three teams have now built the same customer lookup, three ways."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
03
the-rules-that-live-in-a-wiki.
"The three-layer rules and naming standards are written down beautifully. On a deadline, they lose."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
04
the-handover-that-stops-at-go-live.
"The delivery partner rolled off. The documentation stops at go-live, and the people who knew why things were built that way went with them."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
01
the-quote-that-starts-with-discovery.
"The vendor's plan begins with a nine-month discovery phase, before anything moves, because nobody can say what is actually on the old platform."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
02
the-list-nobody-has.
"Ask three people how many connections run on the old platform and you get three numbers, hundreds apart."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
03
the-ones-that-might-be-dead.
"Some connections run every night for nobody. You keep paying for them, because switching one off is scarier than keeping it alive."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
04
the-last-attempt.
"The previous migration stalled partway. The people who ran it have moved on, and nobody is sure which connections were finished."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
01
lost-in-translation.
"What the business asked for and what got built are cousins, not twins. Nobody can say where it drifted, because nothing was written down with sources."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
02
the-prototype-that-shipped.
"The demo impressed everyone, so it quietly became the product. No error handling, no tests, no documentation. Three years later, everyone is afraid of it."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
03
the-creeping-scope.
"Every sprint adds a little. By month six the budget is spent and the core still is not finished."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
04
the-documentation-debt.
"The documentation was written in the last week, from memory, by whoever was left."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
01
pilot-purgatory.
"The demo wowed the board in March. It is September, and launch is still two weeks away, because nobody can say what would make it safe to say yes."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
02
the-improviser.
"In the demo it answered perfectly. Nobody knows what it says when a customer swears at it, lies to it, or asks it for a discount."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
03
the-3x-estimate.-copy
'What is our AI agent strategy?' Every vendor answers 'buy ours'. Nobody answers the real question: how would we trust one?
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
04
the-unowned-promise.
"If the agent promises a customer a refund, whose promise is it? Today, nobody has signed for that."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
01
the-out-of-date-chart.
"The chart showing how your systems connect was last correct in 2021. Three of the people on it have left the company."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
02
the-connection-nobody-dares-touch.
"One old connection moves payroll data. The person who built it left. Nobody touches it, so every project builds an expensive detour around it."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
03
the-3x-estimate.
"A two-week change is quoted at ten weeks, because eight of them go on working out what it might break."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
04
the-shelf.
"The improvements you wanted in 2024 are still a slide in a deck. Not for lack of value, but because changing running systems feels too risky and too expensive."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
01
the-deck-that-never-became-delivery.
"The last assessment cost a fortune, took five months, and lives in a drawer. Beautiful slides. Nothing shipped."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
02
the-interview-truth
"It was built from forty interviews. People told what they remembered, defended what they built, and nobody mentioned the system nobody owns."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
03
the-assessment-that-expired.
By the time decisions were made, the estate had moved on. The document described last year."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
04
the-roadmap-with-no-numbers-behind-it
"The plan said eighteen months and a big figure. Asked how it was calculated, the consultants said: experience."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
01
the-object-nobody-can-read.
Hundreds of fields added over a decade. Nobody knows which are used, so nobody dares remove them.
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
02
the-friday-release-that-broke-monday.
A new automation collided with an old one on the same records.
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
03
the-custom-code-nobody-remembers-why.
Custom code now does what a standard feature can do out of the box.
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
04
the-queue-behind-three-admins.
Every department’s asks funnel into one backlog behind the same few people.
flip card
THE EXAM
A demo shows what an agent can do. The exam shows what it will do.
Before any agent meets a customer, Hapie builds its exam: hundreds of realistic scenarios drawn from your actual work: the routine asks, the ambiguous ones, the angry customer, the customer who asks for something the agent must not give, and the one who tries to trick it. Every scenario has a defined right behaviour, agreed with your team. The agent runs the whole exam, is scored on accuracy, safety, staying inside its guardrails, knowing when to hand over, and cost per interaction, and the results go to a named person. No pass, no licence. And the exam is kept: it runs again on every change.
"You would not let a new employee near customers without training and supervision. An agent gets the same, in writing."
THE TO-BE, AS OPTIONS WITH TRADE-OFFS
Not one recommendation. Options, priced in trade-offs, for you to choose.
From the as-is, Hapie lays out target options: typically a steady path (fix the riskiest things, keep the shape), a balanced path (consolidate and modernise in waves), and a bold path (replatform what matters most). Each option states, with its evidence: what changes, what it makes possible, what it risks, roughly what it takes, drawn from measured past deliveries, and what it deliberately leaves alone. The choice is yours; the evidence is the same for every option.
Genuinely custom. Built to the integration pattern, documented, and added to the map for the next team.
Genuinely custom. Built to the integration pattern, documented, and added to the map for the next team.
Genuinely custom. Built to the integration pattern, documented, and added to the map for the next team.
Genuinely custom. Built to the integration pattern, documented, and added to the map for the next team.
Genuinely custom. Built to the integration pattern, documented, and added to the map for the next team.
Genuinely custom. Built to the integration pattern, documented, and added to the map for the next team.
Genuinely custom. Built to the integration pattern, documented, and added to the map for the next team.
Genuinely custom. Built to the integration pattern, documented, and added to the map for the next team.
Genuinely custom. Built to the integration pattern, documented, and added to the map for the next team.
Genuinely custom. Built to the integration pattern, documented, and added to the map for the next team.
ONE AGENT, DELIVERED
From job description to working agent, with proof at every step.
One real agent, walked end to end: an order-status agent in Agentforce.
The agent gets a job description.
What it is for, who it serves, what good looks like, and its guardrails, written and approved before anything is built, like any role worth hiring for.
The exam is built with your team.
Bring one request from your backlog. Leave with it scoped.
The agent is built into your platform.
Into Agentforce here, connected only to the systems its guardrails allow, with its logging wired in from the start.
It sits the exam.
The full scenario set, scored. Weaknesses are fixed and it resits. No pass, no licence: see the checkpoint moment below.
A named person signs the launch.
The licence: exam results, guardrails, monitoring plan and rollback path, approved in one place. It launches to a limited group first.
It works, watched, and gets better.
Every action logged, hand-overs reviewed, the exam grows from real conversations, and every model change triggers a re-exam before the agent runs on it.
THE EXAM
A demo shows what an agent can do. The exam shows what it will do.
Before any agent meets a customer, Hapie builds its exam: hundreds of realistic scenarios drawn from your actual work: the routine asks, the ambiguous ones, the angry customer, the customer who asks for something the agent must not give, and the one who tries to trick it. Every scenario has a defined right behaviour, agreed with your team. The agent runs the whole exam, is scored on accuracy, safety, staying inside its guardrails, knowing when to hand over, and cost per interaction, and the results go to a named person. No pass, no licence. And the exam is kept: it runs again on every change.
"You would not let a new employee near customers without training and supervision. An agent gets the same, in writing."
Start
Bring one request from your backlog. Leave with it scoped.
Fifteen minutes on an org like yours: see the map, make the standard-before-custom call, and see the collision check in action.
Read-only to start · Ships through your own release process · Reply within two working days
The three doubts, answered straight
What Salesforce leaders ask.
Is the 95% failure statistic real?
The study is real, the number is loose. MIT's Project NANDA published The GenAI Divide in July 2025 based on 52 interviews and 153 survey responses. Its traceable finding is that 60% of organisations evaluated these tools, 20% piloted and 5% reached production. The "95% get zero return" framing has been publicly questioned by academics for not being derivable from the document.
If agents only succeed 35% of the time on multi-turn tasks, why use them?
Some of them genuinely do, and their judgement matters: that is the context step, and it is welcome. What no person can do is hold two hundred systems' connections in their head, current, at once. Evidence from the systems plus judgement from people beats either alone, and this page is the first half done properly.
How long does it take to get a first agent into production properly?
Someone accountable for the outcome in the estate, not someone accountable for the platform. In practice that is usually the person who would be asked to explain the change to a regulator. If nobody can be named in a sentence, that is the finding, and it is worth surfacing before an auditor surfaces it.
Does the EU AI Act apply to us if we only use agents internally?
Yes. The map is built from the org as it is, debt included, turning cleanup into ordinary evidenced changes.
Is Salesforce Shield enough on its own?
Some of them genuinely do, and their judgement matters: that is the context step, and it is welcome. What no person can do is hold two hundred systems' connections in their head, current, at once. Evidence from the systems plus judgement from people beats either alone, and this page is the first half done properly.
Who should be the named approver for agent work?
Someone accountable for the outcome in the estate, not someone accountable for the platform. In practice that is usually the person who would be asked to explain the change to a regulator. If nobody can be named in a sentence, that is the finding, and it is worth surfacing before an auditor surfaces it.
Is this not just a traceability matrix?
Yes. The map is built from the org as it is, debt included, turning cleanup into ordinary evidenced changes.
Does this slow delivery down by adding approval steps?
Some of them genuinely do, and their judgement matters: that is the context step, and it is welcome. What no person can do is hold two hundred systems' connections in their head, current, at once. Evidence from the systems plus judgement from people beats either alone, and this page is the first half done properly.
How does this work on a brownfield estate with ten years of history?
That is the harder and more common case. The map is built from what already exists, including the parts nobody documented. Long-lived estates usually see the biggest difference, because reconstruction cost per transition is highest where the decisions are oldest.
"Our bids are our secret sauce."
Yes. The map is built from the org as it is, debt included, turning cleanup into ordinary evidenced changes.
"Every RFP is different."
Some of them genuinely do, and their judgement matters: that is the context step, and it is welcome. What no person can do is hold two hundred systems' connections in their head, current, at once. Evidence from the systems plus judgement from people beats either alone, and this page is the first half done properly.
"Will it read like AI wrote it?"
The opposite, and here is why: what makes bids read machine-made is generic filler, and filler exists because humans padding forty pages at midnight reach for it. Here, every sentence must attach to your record or it does not survive the checkpoint, so what remains is specific, in your voice, and signed by your people. The blandest bid your buyer will read this quarter was written by exhausted humans.
"Our MuleSoft estate is already a mess."
Yes. The map is built from the org as it is, debt included, turning cleanup into ordinary evidenced changes.
"We already have accelerators and templates."
Some of them genuinely do, and their judgement matters: that is the context step, and it is welcome. What no person can do is hold two hundred systems' connections in their head, current, at once. Evidence from the systems plus judgement from people beats either alone, and this page is the first half done properly.
"Will it fit our Anypoint setup and our release process?"
Hapie reads your Anypoint estate to build the map and the catalogue view, and everything ships through your own pipelines and approvals. Your setup stays your setup; Hapie supplies the method and the proof.
"We tried this before and it stalled."
Yes. The map is built from the org as it is, debt included, turning cleanup into ordinary evidenced changes.
"We cannot freeze the business for the duration."
Some of them genuinely do, and their judgement matters: that is the context step, and it is welcome. What no person can do is hold two hundred systems' connections in their head, current, at once. Evidence from the systems plus judgement from people beats either alone, and this page is the first half done properly.
"What if the new platform behaves differently?"
That is what the side-by-side run exists to catch. Every connection processes the same real inputs on both platforms until the outputs match, line by line, and the quiet edge cases, rounding, empty fields, odd dates, are exactly what it makes loud.
"AI tools already build apps from a prompt. Why is this different?"
Yes. The map is built from the org as it is, debt included, turning cleanup into ordinary evidenced changes.
"Our requirements always change mid-build."
Some of them genuinely do, and their judgement matters: that is the context step, and it is welcome. What no person can do is hold two hundred systems' connections in their head, current, at once. Evidence from the systems plus judgement from people beats either alone, and this page is the first half done properly.
"We do not have people free to write requirements for months."
You do not write them; you approve them. The agents draft requirements from the workshops and documents you already have, and your people answer a short open-questions list and sign the scope. The heavy documentation lifting is exactly what is automated.
"We are already building in Agentforce."
Yes. The map is built from the org as it is, debt included, turning cleanup into ordinary evidenced changes.
"Models change constantly. Won't the agent drift?"
Some of them genuinely do, and their judgement matters: that is the context step, and it is welcome. What no person can do is hold two hundred systems' connections in their head, current, at once. Evidence from the systems plus judgement from people beats either alone, and this page is the first half done properly.
"Our legal and compliance teams will never sign this."
They will not sign vibes, and they are right not to. What they can sign is what this process manufactures: written limits with named owners, exam results against agreed behaviours, complete action logs, and a rollback path. Several sections of this site, Security, and Documents you share with us, are written to be sent to them first. Bring them to the demo; they are the easiest audience we have.
"Our systems are a mess. No tool can read them."
Yes. The map is built from the org as it is, debt included, turning cleanup into ordinary evidenced changes.
"We cannot let an AI change systems that are running.
Some of them genuinely do, and their judgement matters: that is the context step, and it is welcome. What no person can do is hold two hundred systems' connections in their head, current, at once. Evidence from the systems plus judgement from people beats either alone, and this page is the first half done properly.
"The people who understood the old parts have left."
Much of what they knew is still in what they left behind: the code, the tickets, the history of changes. Hapie recovers that part. What cannot be recovered is confirmed once by your team, then kept on the map, so it is never lost again.
"We did an assessment last year."
Yes. The map is built from the org as it is, debt included, turning cleanup into ordinary evidenced changes.
"Our consultants already know our estate"
Some of them genuinely do, and their judgement matters: that is the context step, and it is welcome. What no person can do is hold two hundred systems' connections in their head, current, at once. Evidence from the systems plus judgement from people beats either alone, and this page is the first half done properly.
"We cannot open our systems to an outside tool."
It is read-only, it runs in your environment, and building the as-is changes nothing. The access conversation is with your security team on your terms, and the Documents and Security pages on this site are written to be sent to them first.
“Our org is a decade of debt. Start there?”
Yes. The map is built from the org as it is, debt included, turning cleanup into ordinary evidenced changes.
“We already have admins and a partner.”
Keep them. They become the approvers and standard-setters while the agents do the building.
“Will this fight our release process?”
No. It ships through your sandboxes, pipelines and release calendar. What it adds is the map and checkpoint.
USE CASE · AI AGENT DELIVERY · AGENTFORCE AND BEYOND
AI agents in production, not in pilot. Because yours pass an exam first.
Anyone can demo an agent. Hapie delivers them: requirements and guardrails written down, an exam of hundreds of realistic scenarios it must pass, a named person approving the launch, every action logged in production, and a re-exam every time the underlying model changes. Into Agentforce, into your platforms, on your terms.
THE FOUR THINGS EVERY AGENT PROGRAMME RECOGNISES
Why agents wow in March and are still "two weeks away" in September.
If your organisation has piloted an AI agent, you will recognise these.
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
01
the-lost-weekend
"Three senior architects, two weekends, one deadline. The people who bill the most, writing documents that usually lose."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
02
the-copy-paste-bid.
"Last year's response with the client's name changed. Except the one place it wasn't."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
03
the-promise-delivery-discovered-later.
"Somewhere on page forty, we committed to something. Delivery found out after we won."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
04
the-loss-with-no-lesson.
"We lost. Procurement said 'scoring'. Nobody knows which answer cost us, so the next bid repeats it."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
01
the-specialist-queue.
"Good MuleSoft people are scarce and expensive. Your whole programme moves at the speed of the three you have."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
02
the-api-that-already-existed.
"Nobody checks the catalogue under deadline pressure. Three teams have now built the same customer lookup, three ways."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
03
the-rules-that-live-in-a-wiki.
"The three-layer rules and naming standards are written down beautifully. On a deadline, they lose."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
04
the-handover-that-stops-at-go-live.
"The delivery partner rolled off. The documentation stops at go-live, and the people who knew why things were built that way went with them."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
01
the-quote-that-starts-with-discovery.
"The vendor's plan begins with a nine-month discovery phase, before anything moves, because nobody can say what is actually on the old platform."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
02
the-list-nobody-has.
"Ask three people how many connections run on the old platform and you get three numbers, hundreds apart."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
03
the-ones-that-might-be-dead.
"Some connections run every night for nobody. You keep paying for them, because switching one off is scarier than keeping it alive."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
04
the-last-attempt.
"The previous migration stalled partway. The people who ran it have moved on, and nobody is sure which connections were finished."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
01
lost-in-translation.
"What the business asked for and what got built are cousins, not twins. Nobody can say where it drifted, because nothing was written down with sources."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
02
the-prototype-that-shipped.
"The demo impressed everyone, so it quietly became the product. No error handling, no tests, no documentation. Three years later, everyone is afraid of it."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
03
the-creeping-scope.
"Every sprint adds a little. By month six the budget is spent and the core still is not finished."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
04
the-documentation-debt.
"The documentation was written in the last week, from memory, by whoever was left."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
01
pilot-purgatory.
"The demo wowed the board in March. It is September, and launch is still two weeks away, because nobody can say what would make it safe to say yes."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
02
the-improviser.
"In the demo it answered perfectly. Nobody knows what it says when a customer swears at it, lies to it, or asks it for a discount."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
03
the-3x-estimate.-copy
'What is our AI agent strategy?' Every vendor answers 'buy ours'. Nobody answers the real question: how would we trust one?
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
04
the-unowned-promise.
"If the agent promises a customer a refund, whose promise is it? Today, nobody has signed for that."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
01
the-out-of-date-chart.
"The chart showing how your systems connect was last correct in 2021. Three of the people on it have left the company."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
02
the-connection-nobody-dares-touch.
"One old connection moves payroll data. The person who built it left. Nobody touches it, so every project builds an expensive detour around it."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
03
the-3x-estimate.
"A two-week change is quoted at ten weeks, because eight of them go on working out what it might break."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
04
the-shelf.
"The improvements you wanted in 2024 are still a slide in a deck. Not for lack of value, but because changing running systems feels too risky and too expensive."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
01
the-deck-that-never-became-delivery.
"The last assessment cost a fortune, took five months, and lives in a drawer. Beautiful slides. Nothing shipped."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
02
the-interview-truth
"It was built from forty interviews. People told what they remembered, defended what they built, and nobody mentioned the system nobody owns."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
03
the-assessment-that-expired.
By the time decisions were made, the estate had moved on. The document described last year."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
04
the-roadmap-with-no-numbers-behind-it
"The plan said eighteen months and a big figure. Asked how it was calculated, the consultants said: experience."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
01
the-object-nobody-can-read.
Hundreds of fields added over a decade. Nobody knows which are used, so nobody dares remove them.
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
02
the-friday-release-that-broke-monday.
A new automation collided with an old one on the same records.
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
03
the-custom-code-nobody-remembers-why.
Custom code now does what a standard feature can do out of the box.
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
04
the-queue-behind-three-admins.
Every department’s asks funnel into one backlog behind the same few people.
flip card
THE EXAM
A demo shows what an agent can do. The exam shows what it will do.
Before any agent meets a customer, Hapie builds its exam: hundreds of realistic scenarios drawn from your actual work: the routine asks, the ambiguous ones, the angry customer, the customer who asks for something the agent must not give, and the one who tries to trick it. Every scenario has a defined right behaviour, agreed with your team. The agent runs the whole exam, is scored on accuracy, safety, staying inside its guardrails, knowing when to hand over, and cost per interaction, and the results go to a named person. No pass, no licence. And the exam is kept: it runs again on every change.
"You would not let a new employee near customers without training and supervision. An agent gets the same, in writing."
THE TO-BE, AS OPTIONS WITH TRADE-OFFS
Not one recommendation. Options, priced in trade-offs, for you to choose.
From the as-is, Hapie lays out target options: typically a steady path (fix the riskiest things, keep the shape), a balanced path (consolidate and modernise in waves), and a bold path (replatform what matters most). Each option states, with its evidence: what changes, what it makes possible, what it risks, roughly what it takes, drawn from measured past deliveries, and what it deliberately leaves alone. The choice is yours; the evidence is the same for every option.
Genuinely custom. Built to the integration pattern, documented, and added to the map for the next team.
Genuinely custom. Built to the integration pattern, documented, and added to the map for the next team.
Genuinely custom. Built to the integration pattern, documented, and added to the map for the next team.
Genuinely custom. Built to the integration pattern, documented, and added to the map for the next team.
Genuinely custom. Built to the integration pattern, documented, and added to the map for the next team.
Genuinely custom. Built to the integration pattern, documented, and added to the map for the next team.
Genuinely custom. Built to the integration pattern, documented, and added to the map for the next team.
Genuinely custom. Built to the integration pattern, documented, and added to the map for the next team.
Genuinely custom. Built to the integration pattern, documented, and added to the map for the next team.
Genuinely custom. Built to the integration pattern, documented, and added to the map for the next team.
ONE AGENT, DELIVERED
From job description to working agent, with proof at every step.
One real agent, walked end to end: an order-status agent in Agentforce.
The agent gets a job description.
What it is for, who it serves, what good looks like, and its guardrails, written and approved before anything is built, like any role worth hiring for.
The exam is built with your team.
Bring one request from your backlog. Leave with it scoped.
The agent is built into your platform.
Into Agentforce here, connected only to the systems its guardrails allow, with its logging wired in from the start.
It sits the exam.
The full scenario set, scored. Weaknesses are fixed and it resits. No pass, no licence: see the checkpoint moment below.
A named person signs the launch.
The licence: exam results, guardrails, monitoring plan and rollback path, approved in one place. It launches to a limited group first.
It works, watched, and gets better.
Every action logged, hand-overs reviewed, the exam grows from real conversations, and every model change triggers a re-exam before the agent runs on it.
THE EXAM
A demo shows what an agent can do. The exam shows what it will do.
Before any agent meets a customer, Hapie builds its exam: hundreds of realistic scenarios drawn from your actual work: the routine asks, the ambiguous ones, the angry customer, the customer who asks for something the agent must not give, and the one who tries to trick it. Every scenario has a defined right behaviour, agreed with your team. The agent runs the whole exam, is scored on accuracy, safety, staying inside its guardrails, knowing when to hand over, and cost per interaction, and the results go to a named person. No pass, no licence. And the exam is kept: it runs again on every change.
"You would not let a new employee near customers without training and supervision. An agent gets the same, in writing."
Start
Bring one request from your backlog. Leave with it scoped.
Fifteen minutes on an org like yours: see the map, make the standard-before-custom call, and see the collision check in action.
Read-only to start · Ships through your own release process · Reply within two working days
The three doubts, answered straight
What Salesforce leaders ask.
Is the 95% failure statistic real?
The study is real, the number is loose. MIT's Project NANDA published The GenAI Divide in July 2025 based on 52 interviews and 153 survey responses. Its traceable finding is that 60% of organisations evaluated these tools, 20% piloted and 5% reached production. The "95% get zero return" framing has been publicly questioned by academics for not being derivable from the document.
If agents only succeed 35% of the time on multi-turn tasks, why use them?
Some of them genuinely do, and their judgement matters: that is the context step, and it is welcome. What no person can do is hold two hundred systems' connections in their head, current, at once. Evidence from the systems plus judgement from people beats either alone, and this page is the first half done properly.
How long does it take to get a first agent into production properly?
Someone accountable for the outcome in the estate, not someone accountable for the platform. In practice that is usually the person who would be asked to explain the change to a regulator. If nobody can be named in a sentence, that is the finding, and it is worth surfacing before an auditor surfaces it.
Does the EU AI Act apply to us if we only use agents internally?
Yes. The map is built from the org as it is, debt included, turning cleanup into ordinary evidenced changes.
Is Salesforce Shield enough on its own?
Some of them genuinely do, and their judgement matters: that is the context step, and it is welcome. What no person can do is hold two hundred systems' connections in their head, current, at once. Evidence from the systems plus judgement from people beats either alone, and this page is the first half done properly.
Who should be the named approver for agent work?
Someone accountable for the outcome in the estate, not someone accountable for the platform. In practice that is usually the person who would be asked to explain the change to a regulator. If nobody can be named in a sentence, that is the finding, and it is worth surfacing before an auditor surfaces it.
Is this not just a traceability matrix?
Yes. The map is built from the org as it is, debt included, turning cleanup into ordinary evidenced changes.
Does this slow delivery down by adding approval steps?
Some of them genuinely do, and their judgement matters: that is the context step, and it is welcome. What no person can do is hold two hundred systems' connections in their head, current, at once. Evidence from the systems plus judgement from people beats either alone, and this page is the first half done properly.
How does this work on a brownfield estate with ten years of history?
That is the harder and more common case. The map is built from what already exists, including the parts nobody documented. Long-lived estates usually see the biggest difference, because reconstruction cost per transition is highest where the decisions are oldest.
"Our bids are our secret sauce."
Yes. The map is built from the org as it is, debt included, turning cleanup into ordinary evidenced changes.
"Every RFP is different."
Some of them genuinely do, and their judgement matters: that is the context step, and it is welcome. What no person can do is hold two hundred systems' connections in their head, current, at once. Evidence from the systems plus judgement from people beats either alone, and this page is the first half done properly.
"Will it read like AI wrote it?"
The opposite, and here is why: what makes bids read machine-made is generic filler, and filler exists because humans padding forty pages at midnight reach for it. Here, every sentence must attach to your record or it does not survive the checkpoint, so what remains is specific, in your voice, and signed by your people. The blandest bid your buyer will read this quarter was written by exhausted humans.
"Our MuleSoft estate is already a mess."
Yes. The map is built from the org as it is, debt included, turning cleanup into ordinary evidenced changes.
"We already have accelerators and templates."
Some of them genuinely do, and their judgement matters: that is the context step, and it is welcome. What no person can do is hold two hundred systems' connections in their head, current, at once. Evidence from the systems plus judgement from people beats either alone, and this page is the first half done properly.
"Will it fit our Anypoint setup and our release process?"
Hapie reads your Anypoint estate to build the map and the catalogue view, and everything ships through your own pipelines and approvals. Your setup stays your setup; Hapie supplies the method and the proof.
"We tried this before and it stalled."
Yes. The map is built from the org as it is, debt included, turning cleanup into ordinary evidenced changes.
"We cannot freeze the business for the duration."
Some of them genuinely do, and their judgement matters: that is the context step, and it is welcome. What no person can do is hold two hundred systems' connections in their head, current, at once. Evidence from the systems plus judgement from people beats either alone, and this page is the first half done properly.
"What if the new platform behaves differently?"
That is what the side-by-side run exists to catch. Every connection processes the same real inputs on both platforms until the outputs match, line by line, and the quiet edge cases, rounding, empty fields, odd dates, are exactly what it makes loud.
"AI tools already build apps from a prompt. Why is this different?"
Yes. The map is built from the org as it is, debt included, turning cleanup into ordinary evidenced changes.
"Our requirements always change mid-build."
Some of them genuinely do, and their judgement matters: that is the context step, and it is welcome. What no person can do is hold two hundred systems' connections in their head, current, at once. Evidence from the systems plus judgement from people beats either alone, and this page is the first half done properly.
"We do not have people free to write requirements for months."
You do not write them; you approve them. The agents draft requirements from the workshops and documents you already have, and your people answer a short open-questions list and sign the scope. The heavy documentation lifting is exactly what is automated.
"We are already building in Agentforce."
Yes. The map is built from the org as it is, debt included, turning cleanup into ordinary evidenced changes.
"Models change constantly. Won't the agent drift?"
Some of them genuinely do, and their judgement matters: that is the context step, and it is welcome. What no person can do is hold two hundred systems' connections in their head, current, at once. Evidence from the systems plus judgement from people beats either alone, and this page is the first half done properly.
"Our legal and compliance teams will never sign this."
They will not sign vibes, and they are right not to. What they can sign is what this process manufactures: written limits with named owners, exam results against agreed behaviours, complete action logs, and a rollback path. Several sections of this site, Security, and Documents you share with us, are written to be sent to them first. Bring them to the demo; they are the easiest audience we have.
"Our systems are a mess. No tool can read them."
Yes. The map is built from the org as it is, debt included, turning cleanup into ordinary evidenced changes.
"We cannot let an AI change systems that are running.
Some of them genuinely do, and their judgement matters: that is the context step, and it is welcome. What no person can do is hold two hundred systems' connections in their head, current, at once. Evidence from the systems plus judgement from people beats either alone, and this page is the first half done properly.
"The people who understood the old parts have left."
Much of what they knew is still in what they left behind: the code, the tickets, the history of changes. Hapie recovers that part. What cannot be recovered is confirmed once by your team, then kept on the map, so it is never lost again.
"We did an assessment last year."
Yes. The map is built from the org as it is, debt included, turning cleanup into ordinary evidenced changes.
"Our consultants already know our estate"
Some of them genuinely do, and their judgement matters: that is the context step, and it is welcome. What no person can do is hold two hundred systems' connections in their head, current, at once. Evidence from the systems plus judgement from people beats either alone, and this page is the first half done properly.
"We cannot open our systems to an outside tool."
It is read-only, it runs in your environment, and building the as-is changes nothing. The access conversation is with your security team on your terms, and the Documents and Security pages on this site are written to be sent to them first.
“Our org is a decade of debt. Start there?”
Yes. The map is built from the org as it is, debt included, turning cleanup into ordinary evidenced changes.
“We already have admins and a partner.”
Keep them. They become the approvers and standard-setters while the agents do the building.
“Will this fight our release process?”
No. It ships through your sandboxes, pipelines and release calendar. What it adds is the map and checkpoint.
USE CASE · AI AGENT DELIVERY · AGENTFORCE AND BEYOND
AI agents in production, not in pilot. Because yours pass an exam first.
Anyone can demo an agent. Hapie delivers them: requirements and guardrails written down, an exam of hundreds of realistic scenarios it must pass, a named person approving the launch, every action logged in production, and a re-exam every time the underlying model changes. Into Agentforce, into your platforms, on your terms.
THE FOUR THINGS EVERY AGENT PROGRAMME RECOGNISES
Why agents wow in March and are still "two weeks away" in September.
If your organisation has piloted an AI agent, you will recognise these.
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
01
the-lost-weekend
"Three senior architects, two weekends, one deadline. The people who bill the most, writing documents that usually lose."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
02
the-copy-paste-bid.
"Last year's response with the client's name changed. Except the one place it wasn't."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
03
the-promise-delivery-discovered-later.
"Somewhere on page forty, we committed to something. Delivery found out after we won."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
04
the-loss-with-no-lesson.
"We lost. Procurement said 'scoring'. Nobody knows which answer cost us, so the next bid repeats it."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
01
the-specialist-queue.
"Good MuleSoft people are scarce and expensive. Your whole programme moves at the speed of the three you have."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
02
the-api-that-already-existed.
"Nobody checks the catalogue under deadline pressure. Three teams have now built the same customer lookup, three ways."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
03
the-rules-that-live-in-a-wiki.
"The three-layer rules and naming standards are written down beautifully. On a deadline, they lose."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
04
the-handover-that-stops-at-go-live.
"The delivery partner rolled off. The documentation stops at go-live, and the people who knew why things were built that way went with them."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
01
the-quote-that-starts-with-discovery.
"The vendor's plan begins with a nine-month discovery phase, before anything moves, because nobody can say what is actually on the old platform."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
02
the-list-nobody-has.
"Ask three people how many connections run on the old platform and you get three numbers, hundreds apart."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
03
the-ones-that-might-be-dead.
"Some connections run every night for nobody. You keep paying for them, because switching one off is scarier than keeping it alive."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
04
the-last-attempt.
"The previous migration stalled partway. The people who ran it have moved on, and nobody is sure which connections were finished."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
01
lost-in-translation.
"What the business asked for and what got built are cousins, not twins. Nobody can say where it drifted, because nothing was written down with sources."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
02
the-prototype-that-shipped.
"The demo impressed everyone, so it quietly became the product. No error handling, no tests, no documentation. Three years later, everyone is afraid of it."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
03
the-creeping-scope.
"Every sprint adds a little. By month six the budget is spent and the core still is not finished."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
04
the-documentation-debt.
"The documentation was written in the last week, from memory, by whoever was left."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
01
pilot-purgatory.
"The demo wowed the board in March. It is September, and launch is still two weeks away, because nobody can say what would make it safe to say yes."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
02
the-improviser.
"In the demo it answered perfectly. Nobody knows what it says when a customer swears at it, lies to it, or asks it for a discount."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
03
the-3x-estimate.-copy
'What is our AI agent strategy?' Every vendor answers 'buy ours'. Nobody answers the real question: how would we trust one?
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
04
the-unowned-promise.
"If the agent promises a customer a refund, whose promise is it? Today, nobody has signed for that."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
01
the-out-of-date-chart.
"The chart showing how your systems connect was last correct in 2021. Three of the people on it have left the company."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
02
the-connection-nobody-dares-touch.
"One old connection moves payroll data. The person who built it left. Nobody touches it, so every project builds an expensive detour around it."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
03
the-3x-estimate.
"A two-week change is quoted at ten weeks, because eight of them go on working out what it might break."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
04
the-shelf.
"The improvements you wanted in 2024 are still a slide in a deck. Not for lack of value, but because changing running systems feels too risky and too expensive."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
01
the-deck-that-never-became-delivery.
"The last assessment cost a fortune, took five months, and lives in a drawer. Beautiful slides. Nothing shipped."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
02
the-interview-truth
"It was built from forty interviews. People told what they remembered, defended what they built, and nobody mentioned the system nobody owns."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
03
the-assessment-that-expired.
By the time decisions were made, the estate had moved on. The document described last year."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
04
the-roadmap-with-no-numbers-behind-it
"The plan said eighteen months and a big figure. Asked how it was calculated, the consultants said: experience."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
01
the-object-nobody-can-read.
Hundreds of fields added over a decade. Nobody knows which are used, so nobody dares remove them.
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
02
the-friday-release-that-broke-monday.
A new automation collided with an old one on the same records.
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
03
the-custom-code-nobody-remembers-why.
Custom code now does what a standard feature can do out of the box.
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
04
the-queue-behind-three-admins.
Every department’s asks funnel into one backlog behind the same few people.
flip card
THE EXAM
A demo shows what an agent can do. The exam shows what it will do.
Before any agent meets a customer, Hapie builds its exam: hundreds of realistic scenarios drawn from your actual work: the routine asks, the ambiguous ones, the angry customer, the customer who asks for something the agent must not give, and the one who tries to trick it. Every scenario has a defined right behaviour, agreed with your team. The agent runs the whole exam, is scored on accuracy, safety, staying inside its guardrails, knowing when to hand over, and cost per interaction, and the results go to a named person. No pass, no licence. And the exam is kept: it runs again on every change.
"You would not let a new employee near customers without training and supervision. An agent gets the same, in writing."
THE TO-BE, AS OPTIONS WITH TRADE-OFFS
Not one recommendation. Options, priced in trade-offs, for you to choose.
From the as-is, Hapie lays out target options: typically a steady path (fix the riskiest things, keep the shape), a balanced path (consolidate and modernise in waves), and a bold path (replatform what matters most). Each option states, with its evidence: what changes, what it makes possible, what it risks, roughly what it takes, drawn from measured past deliveries, and what it deliberately leaves alone. The choice is yours; the evidence is the same for every option.
Genuinely custom. Built to the integration pattern, documented, and added to the map for the next team.
Genuinely custom. Built to the integration pattern, documented, and added to the map for the next team.
Genuinely custom. Built to the integration pattern, documented, and added to the map for the next team.
Genuinely custom. Built to the integration pattern, documented, and added to the map for the next team.
Genuinely custom. Built to the integration pattern, documented, and added to the map for the next team.
Genuinely custom. Built to the integration pattern, documented, and added to the map for the next team.
Genuinely custom. Built to the integration pattern, documented, and added to the map for the next team.
Genuinely custom. Built to the integration pattern, documented, and added to the map for the next team.
Genuinely custom. Built to the integration pattern, documented, and added to the map for the next team.
Genuinely custom. Built to the integration pattern, documented, and added to the map for the next team.
ONE AGENT, DELIVERED
From job description to working agent, with proof at every step.
One real agent, walked end to end: an order-status agent in Agentforce.
The agent gets a job description.
What it is for, who it serves, what good looks like, and its guardrails, written and approved before anything is built, like any role worth hiring for.
The exam is built with your team.
Bring one request from your backlog. Leave with it scoped.
The agent is built into your platform.
Into Agentforce here, connected only to the systems its guardrails allow, with its logging wired in from the start.
It sits the exam.
The full scenario set, scored. Weaknesses are fixed and it resits. No pass, no licence: see the checkpoint moment below.
A named person signs the launch.
The licence: exam results, guardrails, monitoring plan and rollback path, approved in one place. It launches to a limited group first.
It works, watched, and gets better.
Every action logged, hand-overs reviewed, the exam grows from real conversations, and every model change triggers a re-exam before the agent runs on it.
THE EXAM
A demo shows what an agent can do. The exam shows what it will do.
Before any agent meets a customer, Hapie builds its exam: hundreds of realistic scenarios drawn from your actual work: the routine asks, the ambiguous ones, the angry customer, the customer who asks for something the agent must not give, and the one who tries to trick it. Every scenario has a defined right behaviour, agreed with your team. The agent runs the whole exam, is scored on accuracy, safety, staying inside its guardrails, knowing when to hand over, and cost per interaction, and the results go to a named person. No pass, no licence. And the exam is kept: it runs again on every change.
"You would not let a new employee near customers without training and supervision. An agent gets the same, in writing."
Start
Bring one request from your backlog. Leave with it scoped.
Fifteen minutes on an org like yours: see the map, make the standard-before-custom call, and see the collision check in action.
Read-only to start · Ships through your own release process · Reply within two working days
The three doubts, answered straight
What Salesforce leaders ask.
Is the 95% failure statistic real?
The study is real, the number is loose. MIT's Project NANDA published The GenAI Divide in July 2025 based on 52 interviews and 153 survey responses. Its traceable finding is that 60% of organisations evaluated these tools, 20% piloted and 5% reached production. The "95% get zero return" framing has been publicly questioned by academics for not being derivable from the document.
If agents only succeed 35% of the time on multi-turn tasks, why use them?
Some of them genuinely do, and their judgement matters: that is the context step, and it is welcome. What no person can do is hold two hundred systems' connections in their head, current, at once. Evidence from the systems plus judgement from people beats either alone, and this page is the first half done properly.
How long does it take to get a first agent into production properly?
Someone accountable for the outcome in the estate, not someone accountable for the platform. In practice that is usually the person who would be asked to explain the change to a regulator. If nobody can be named in a sentence, that is the finding, and it is worth surfacing before an auditor surfaces it.
Does the EU AI Act apply to us if we only use agents internally?
Yes. The map is built from the org as it is, debt included, turning cleanup into ordinary evidenced changes.
Is Salesforce Shield enough on its own?
Some of them genuinely do, and their judgement matters: that is the context step, and it is welcome. What no person can do is hold two hundred systems' connections in their head, current, at once. Evidence from the systems plus judgement from people beats either alone, and this page is the first half done properly.
Who should be the named approver for agent work?
Someone accountable for the outcome in the estate, not someone accountable for the platform. In practice that is usually the person who would be asked to explain the change to a regulator. If nobody can be named in a sentence, that is the finding, and it is worth surfacing before an auditor surfaces it.
Is this not just a traceability matrix?
Yes. The map is built from the org as it is, debt included, turning cleanup into ordinary evidenced changes.
Does this slow delivery down by adding approval steps?
Some of them genuinely do, and their judgement matters: that is the context step, and it is welcome. What no person can do is hold two hundred systems' connections in their head, current, at once. Evidence from the systems plus judgement from people beats either alone, and this page is the first half done properly.
How does this work on a brownfield estate with ten years of history?
That is the harder and more common case. The map is built from what already exists, including the parts nobody documented. Long-lived estates usually see the biggest difference, because reconstruction cost per transition is highest where the decisions are oldest.
"Our bids are our secret sauce."
Yes. The map is built from the org as it is, debt included, turning cleanup into ordinary evidenced changes.
"Every RFP is different."
Some of them genuinely do, and their judgement matters: that is the context step, and it is welcome. What no person can do is hold two hundred systems' connections in their head, current, at once. Evidence from the systems plus judgement from people beats either alone, and this page is the first half done properly.
"Will it read like AI wrote it?"
The opposite, and here is why: what makes bids read machine-made is generic filler, and filler exists because humans padding forty pages at midnight reach for it. Here, every sentence must attach to your record or it does not survive the checkpoint, so what remains is specific, in your voice, and signed by your people. The blandest bid your buyer will read this quarter was written by exhausted humans.
"Our MuleSoft estate is already a mess."
Yes. The map is built from the org as it is, debt included, turning cleanup into ordinary evidenced changes.
"We already have accelerators and templates."
Some of them genuinely do, and their judgement matters: that is the context step, and it is welcome. What no person can do is hold two hundred systems' connections in their head, current, at once. Evidence from the systems plus judgement from people beats either alone, and this page is the first half done properly.
"Will it fit our Anypoint setup and our release process?"
Hapie reads your Anypoint estate to build the map and the catalogue view, and everything ships through your own pipelines and approvals. Your setup stays your setup; Hapie supplies the method and the proof.
"We tried this before and it stalled."
Yes. The map is built from the org as it is, debt included, turning cleanup into ordinary evidenced changes.
"We cannot freeze the business for the duration."
Some of them genuinely do, and their judgement matters: that is the context step, and it is welcome. What no person can do is hold two hundred systems' connections in their head, current, at once. Evidence from the systems plus judgement from people beats either alone, and this page is the first half done properly.
"What if the new platform behaves differently?"
That is what the side-by-side run exists to catch. Every connection processes the same real inputs on both platforms until the outputs match, line by line, and the quiet edge cases, rounding, empty fields, odd dates, are exactly what it makes loud.
"AI tools already build apps from a prompt. Why is this different?"
Yes. The map is built from the org as it is, debt included, turning cleanup into ordinary evidenced changes.
"Our requirements always change mid-build."
Some of them genuinely do, and their judgement matters: that is the context step, and it is welcome. What no person can do is hold two hundred systems' connections in their head, current, at once. Evidence from the systems plus judgement from people beats either alone, and this page is the first half done properly.
"We do not have people free to write requirements for months."
You do not write them; you approve them. The agents draft requirements from the workshops and documents you already have, and your people answer a short open-questions list and sign the scope. The heavy documentation lifting is exactly what is automated.
"We are already building in Agentforce."
Yes. The map is built from the org as it is, debt included, turning cleanup into ordinary evidenced changes.
"Models change constantly. Won't the agent drift?"
Some of them genuinely do, and their judgement matters: that is the context step, and it is welcome. What no person can do is hold two hundred systems' connections in their head, current, at once. Evidence from the systems plus judgement from people beats either alone, and this page is the first half done properly.
"Our legal and compliance teams will never sign this."
They will not sign vibes, and they are right not to. What they can sign is what this process manufactures: written limits with named owners, exam results against agreed behaviours, complete action logs, and a rollback path. Several sections of this site, Security, and Documents you share with us, are written to be sent to them first. Bring them to the demo; they are the easiest audience we have.
"Our systems are a mess. No tool can read them."
Yes. The map is built from the org as it is, debt included, turning cleanup into ordinary evidenced changes.
"We cannot let an AI change systems that are running.
Some of them genuinely do, and their judgement matters: that is the context step, and it is welcome. What no person can do is hold two hundred systems' connections in their head, current, at once. Evidence from the systems plus judgement from people beats either alone, and this page is the first half done properly.
"The people who understood the old parts have left."
Much of what they knew is still in what they left behind: the code, the tickets, the history of changes. Hapie recovers that part. What cannot be recovered is confirmed once by your team, then kept on the map, so it is never lost again.
"We did an assessment last year."
Yes. The map is built from the org as it is, debt included, turning cleanup into ordinary evidenced changes.
"Our consultants already know our estate"
Some of them genuinely do, and their judgement matters: that is the context step, and it is welcome. What no person can do is hold two hundred systems' connections in their head, current, at once. Evidence from the systems plus judgement from people beats either alone, and this page is the first half done properly.
"We cannot open our systems to an outside tool."
It is read-only, it runs in your environment, and building the as-is changes nothing. The access conversation is with your security team on your terms, and the Documents and Security pages on this site are written to be sent to them first.
“Our org is a decade of debt. Start there?”
Yes. The map is built from the org as it is, debt included, turning cleanup into ordinary evidenced changes.
“We already have admins and a partner.”
Keep them. They become the approvers and standard-setters while the agents do the building.
“Will this fight our release process?”
No. It ships through your sandboxes, pipelines and release calendar. What it adds is the map and checkpoint.
USE CASE · AI AGENT DELIVERY · AGENTFORCE AND BEYOND
AI agents in production, not in pilot. Because yours pass an exam first.
Anyone can demo an agent. Hapie delivers them: requirements and guardrails written down, an exam of hundreds of realistic scenarios it must pass, a named person approving the launch, every action logged in production, and a re-exam every time the underlying model changes. Into Agentforce, into your platforms, on your terms.
THE FOUR THINGS EVERY AGENT PROGRAMME RECOGNISES
Why agents wow in March and are still "two weeks away" in September.
If your organisation has piloted an AI agent, you will recognise these.
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
01
the-lost-weekend
"Three senior architects, two weekends, one deadline. The people who bill the most, writing documents that usually lose."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
02
the-copy-paste-bid.
"Last year's response with the client's name changed. Except the one place it wasn't."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
03
the-promise-delivery-discovered-later.
"Somewhere on page forty, we committed to something. Delivery found out after we won."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
04
the-loss-with-no-lesson.
"We lost. Procurement said 'scoring'. Nobody knows which answer cost us, so the next bid repeats it."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
01
the-specialist-queue.
"Good MuleSoft people are scarce and expensive. Your whole programme moves at the speed of the three you have."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
02
the-api-that-already-existed.
"Nobody checks the catalogue under deadline pressure. Three teams have now built the same customer lookup, three ways."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
03
the-rules-that-live-in-a-wiki.
"The three-layer rules and naming standards are written down beautifully. On a deadline, they lose."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
04
the-handover-that-stops-at-go-live.
"The delivery partner rolled off. The documentation stops at go-live, and the people who knew why things were built that way went with them."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
01
the-quote-that-starts-with-discovery.
"The vendor's plan begins with a nine-month discovery phase, before anything moves, because nobody can say what is actually on the old platform."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
02
the-list-nobody-has.
"Ask three people how many connections run on the old platform and you get three numbers, hundreds apart."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
03
the-ones-that-might-be-dead.
"Some connections run every night for nobody. You keep paying for them, because switching one off is scarier than keeping it alive."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
04
the-last-attempt.
"The previous migration stalled partway. The people who ran it have moved on, and nobody is sure which connections were finished."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
01
lost-in-translation.
"What the business asked for and what got built are cousins, not twins. Nobody can say where it drifted, because nothing was written down with sources."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
02
the-prototype-that-shipped.
"The demo impressed everyone, so it quietly became the product. No error handling, no tests, no documentation. Three years later, everyone is afraid of it."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
03
the-creeping-scope.
"Every sprint adds a little. By month six the budget is spent and the core still is not finished."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
04
the-documentation-debt.
"The documentation was written in the last week, from memory, by whoever was left."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
01
pilot-purgatory.
"The demo wowed the board in March. It is September, and launch is still two weeks away, because nobody can say what would make it safe to say yes."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
02
the-improviser.
"In the demo it answered perfectly. Nobody knows what it says when a customer swears at it, lies to it, or asks it for a discount."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
03
the-3x-estimate.-copy
'What is our AI agent strategy?' Every vendor answers 'buy ours'. Nobody answers the real question: how would we trust one?
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
04
the-unowned-promise.
"If the agent promises a customer a refund, whose promise is it? Today, nobody has signed for that."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
01
the-out-of-date-chart.
"The chart showing how your systems connect was last correct in 2021. Three of the people on it have left the company."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
02
the-connection-nobody-dares-touch.
"One old connection moves payroll data. The person who built it left. Nobody touches it, so every project builds an expensive detour around it."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
03
the-3x-estimate.
"A two-week change is quoted at ten weeks, because eight of them go on working out what it might break."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
04
the-shelf.
"The improvements you wanted in 2024 are still a slide in a deck. Not for lack of value, but because changing running systems feels too risky and too expensive."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
01
the-deck-that-never-became-delivery.
"The last assessment cost a fortune, took five months, and lives in a drawer. Beautiful slides. Nothing shipped."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
02
the-interview-truth
"It was built from forty interviews. People told what they remembered, defended what they built, and nobody mentioned the system nobody owns."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
03
the-assessment-that-expired.
By the time decisions were made, the estate had moved on. The document described last year."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
04
the-roadmap-with-no-numbers-behind-it
"The plan said eighteen months and a big figure. Asked how it was calculated, the consultants said: experience."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
01
the-object-nobody-can-read.
Hundreds of fields added over a decade. Nobody knows which are used, so nobody dares remove them.
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
02
the-friday-release-that-broke-monday.
A new automation collided with an old one on the same records.
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
03
the-custom-code-nobody-remembers-why.
Custom code now does what a standard feature can do out of the box.
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
04
the-queue-behind-three-admins.
Every department’s asks funnel into one backlog behind the same few people.
flip card
THE EXAM
A demo shows what an agent can do. The exam shows what it will do.
Before any agent meets a customer, Hapie builds its exam: hundreds of realistic scenarios drawn from your actual work: the routine asks, the ambiguous ones, the angry customer, the customer who asks for something the agent must not give, and the one who tries to trick it. Every scenario has a defined right behaviour, agreed with your team. The agent runs the whole exam, is scored on accuracy, safety, staying inside its guardrails, knowing when to hand over, and cost per interaction, and the results go to a named person. No pass, no licence. And the exam is kept: it runs again on every change.
"You would not let a new employee near customers without training and supervision. An agent gets the same, in writing."
THE TO-BE, AS OPTIONS WITH TRADE-OFFS
Not one recommendation. Options, priced in trade-offs, for you to choose.
From the as-is, Hapie lays out target options: typically a steady path (fix the riskiest things, keep the shape), a balanced path (consolidate and modernise in waves), and a bold path (replatform what matters most). Each option states, with its evidence: what changes, what it makes possible, what it risks, roughly what it takes, drawn from measured past deliveries, and what it deliberately leaves alone. The choice is yours; the evidence is the same for every option.
Genuinely custom. Built to the integration pattern, documented, and added to the map for the next team.
Genuinely custom. Built to the integration pattern, documented, and added to the map for the next team.
Genuinely custom. Built to the integration pattern, documented, and added to the map for the next team.
Genuinely custom. Built to the integration pattern, documented, and added to the map for the next team.
Genuinely custom. Built to the integration pattern, documented, and added to the map for the next team.
Genuinely custom. Built to the integration pattern, documented, and added to the map for the next team.
Genuinely custom. Built to the integration pattern, documented, and added to the map for the next team.
Genuinely custom. Built to the integration pattern, documented, and added to the map for the next team.
Genuinely custom. Built to the integration pattern, documented, and added to the map for the next team.
Genuinely custom. Built to the integration pattern, documented, and added to the map for the next team.
ONE AGENT, DELIVERED
From job description to working agent, with proof at every step.
One real agent, walked end to end: an order-status agent in Agentforce.
The agent gets a job description.
What it is for, who it serves, what good looks like, and its guardrails, written and approved before anything is built, like any role worth hiring for.
The exam is built with your team.
Bring one request from your backlog. Leave with it scoped.
The agent is built into your platform.
Into Agentforce here, connected only to the systems its guardrails allow, with its logging wired in from the start.
It sits the exam.
The full scenario set, scored. Weaknesses are fixed and it resits. No pass, no licence: see the checkpoint moment below.
A named person signs the launch.
The licence: exam results, guardrails, monitoring plan and rollback path, approved in one place. It launches to a limited group first.
It works, watched, and gets better.
Every action logged, hand-overs reviewed, the exam grows from real conversations, and every model change triggers a re-exam before the agent runs on it.
THE EXAM
A demo shows what an agent can do. The exam shows what it will do.
Before any agent meets a customer, Hapie builds its exam: hundreds of realistic scenarios drawn from your actual work: the routine asks, the ambiguous ones, the angry customer, the customer who asks for something the agent must not give, and the one who tries to trick it. Every scenario has a defined right behaviour, agreed with your team. The agent runs the whole exam, is scored on accuracy, safety, staying inside its guardrails, knowing when to hand over, and cost per interaction, and the results go to a named person. No pass, no licence. And the exam is kept: it runs again on every change.
"You would not let a new employee near customers without training and supervision. An agent gets the same, in writing."
Start
Bring one request from your backlog. Leave with it scoped.
Fifteen minutes on an org like yours: see the map, make the standard-before-custom call, and see the collision check in action.
Read-only to start · Ships through your own release process · Reply within two working days
The three doubts, answered straight
What Salesforce leaders ask.
Is the 95% failure statistic real?
The study is real, the number is loose. MIT's Project NANDA published The GenAI Divide in July 2025 based on 52 interviews and 153 survey responses. Its traceable finding is that 60% of organisations evaluated these tools, 20% piloted and 5% reached production. The "95% get zero return" framing has been publicly questioned by academics for not being derivable from the document.
If agents only succeed 35% of the time on multi-turn tasks, why use them?
Some of them genuinely do, and their judgement matters: that is the context step, and it is welcome. What no person can do is hold two hundred systems' connections in their head, current, at once. Evidence from the systems plus judgement from people beats either alone, and this page is the first half done properly.
How long does it take to get a first agent into production properly?
Someone accountable for the outcome in the estate, not someone accountable for the platform. In practice that is usually the person who would be asked to explain the change to a regulator. If nobody can be named in a sentence, that is the finding, and it is worth surfacing before an auditor surfaces it.
Does the EU AI Act apply to us if we only use agents internally?
Yes. The map is built from the org as it is, debt included, turning cleanup into ordinary evidenced changes.
Is Salesforce Shield enough on its own?
Some of them genuinely do, and their judgement matters: that is the context step, and it is welcome. What no person can do is hold two hundred systems' connections in their head, current, at once. Evidence from the systems plus judgement from people beats either alone, and this page is the first half done properly.
Who should be the named approver for agent work?
Someone accountable for the outcome in the estate, not someone accountable for the platform. In practice that is usually the person who would be asked to explain the change to a regulator. If nobody can be named in a sentence, that is the finding, and it is worth surfacing before an auditor surfaces it.
Is this not just a traceability matrix?
Yes. The map is built from the org as it is, debt included, turning cleanup into ordinary evidenced changes.
Does this slow delivery down by adding approval steps?
Some of them genuinely do, and their judgement matters: that is the context step, and it is welcome. What no person can do is hold two hundred systems' connections in their head, current, at once. Evidence from the systems plus judgement from people beats either alone, and this page is the first half done properly.
How does this work on a brownfield estate with ten years of history?
That is the harder and more common case. The map is built from what already exists, including the parts nobody documented. Long-lived estates usually see the biggest difference, because reconstruction cost per transition is highest where the decisions are oldest.
"Our bids are our secret sauce."
Yes. The map is built from the org as it is, debt included, turning cleanup into ordinary evidenced changes.
"Every RFP is different."
Some of them genuinely do, and their judgement matters: that is the context step, and it is welcome. What no person can do is hold two hundred systems' connections in their head, current, at once. Evidence from the systems plus judgement from people beats either alone, and this page is the first half done properly.
"Will it read like AI wrote it?"
The opposite, and here is why: what makes bids read machine-made is generic filler, and filler exists because humans padding forty pages at midnight reach for it. Here, every sentence must attach to your record or it does not survive the checkpoint, so what remains is specific, in your voice, and signed by your people. The blandest bid your buyer will read this quarter was written by exhausted humans.
"Our MuleSoft estate is already a mess."
Yes. The map is built from the org as it is, debt included, turning cleanup into ordinary evidenced changes.
"We already have accelerators and templates."
Some of them genuinely do, and their judgement matters: that is the context step, and it is welcome. What no person can do is hold two hundred systems' connections in their head, current, at once. Evidence from the systems plus judgement from people beats either alone, and this page is the first half done properly.
"Will it fit our Anypoint setup and our release process?"
Hapie reads your Anypoint estate to build the map and the catalogue view, and everything ships through your own pipelines and approvals. Your setup stays your setup; Hapie supplies the method and the proof.
"We tried this before and it stalled."
Yes. The map is built from the org as it is, debt included, turning cleanup into ordinary evidenced changes.
"We cannot freeze the business for the duration."
Some of them genuinely do, and their judgement matters: that is the context step, and it is welcome. What no person can do is hold two hundred systems' connections in their head, current, at once. Evidence from the systems plus judgement from people beats either alone, and this page is the first half done properly.
"What if the new platform behaves differently?"
That is what the side-by-side run exists to catch. Every connection processes the same real inputs on both platforms until the outputs match, line by line, and the quiet edge cases, rounding, empty fields, odd dates, are exactly what it makes loud.
"AI tools already build apps from a prompt. Why is this different?"
Yes. The map is built from the org as it is, debt included, turning cleanup into ordinary evidenced changes.
"Our requirements always change mid-build."
Some of them genuinely do, and their judgement matters: that is the context step, and it is welcome. What no person can do is hold two hundred systems' connections in their head, current, at once. Evidence from the systems plus judgement from people beats either alone, and this page is the first half done properly.
"We do not have people free to write requirements for months."
You do not write them; you approve them. The agents draft requirements from the workshops and documents you already have, and your people answer a short open-questions list and sign the scope. The heavy documentation lifting is exactly what is automated.
"We are already building in Agentforce."
Yes. The map is built from the org as it is, debt included, turning cleanup into ordinary evidenced changes.
"Models change constantly. Won't the agent drift?"
Some of them genuinely do, and their judgement matters: that is the context step, and it is welcome. What no person can do is hold two hundred systems' connections in their head, current, at once. Evidence from the systems plus judgement from people beats either alone, and this page is the first half done properly.
"Our legal and compliance teams will never sign this."
They will not sign vibes, and they are right not to. What they can sign is what this process manufactures: written limits with named owners, exam results against agreed behaviours, complete action logs, and a rollback path. Several sections of this site, Security, and Documents you share with us, are written to be sent to them first. Bring them to the demo; they are the easiest audience we have.
"Our systems are a mess. No tool can read them."
Yes. The map is built from the org as it is, debt included, turning cleanup into ordinary evidenced changes.
"We cannot let an AI change systems that are running.
Some of them genuinely do, and their judgement matters: that is the context step, and it is welcome. What no person can do is hold two hundred systems' connections in their head, current, at once. Evidence from the systems plus judgement from people beats either alone, and this page is the first half done properly.
"The people who understood the old parts have left."
Much of what they knew is still in what they left behind: the code, the tickets, the history of changes. Hapie recovers that part. What cannot be recovered is confirmed once by your team, then kept on the map, so it is never lost again.
"We did an assessment last year."
Yes. The map is built from the org as it is, debt included, turning cleanup into ordinary evidenced changes.
"Our consultants already know our estate"
Some of them genuinely do, and their judgement matters: that is the context step, and it is welcome. What no person can do is hold two hundred systems' connections in their head, current, at once. Evidence from the systems plus judgement from people beats either alone, and this page is the first half done properly.
"We cannot open our systems to an outside tool."
It is read-only, it runs in your environment, and building the as-is changes nothing. The access conversation is with your security team on your terms, and the Documents and Security pages on this site are written to be sent to them first.
“Our org is a decade of debt. Start there?”
Yes. The map is built from the org as it is, debt included, turning cleanup into ordinary evidenced changes.
“We already have admins and a partner.”
Keep them. They become the approvers and standard-setters while the agents do the building.
“Will this fight our release process?”
No. It ships through your sandboxes, pipelines and release calendar. What it adds is the map and checkpoint.
USE CASE · AI AGENT DELIVERY · AGENTFORCE AND BEYOND
AI agents in production, not in pilot. Because yours pass an exam first.
Anyone can demo an agent. Hapie delivers them: requirements and guardrails written down, an exam of hundreds of realistic scenarios it must pass, a named person approving the launch, every action logged in production, and a re-exam every time the underlying model changes. Into Agentforce, into your platforms, on your terms.
THE FOUR THINGS EVERY AGENT PROGRAMME RECOGNISES
Why agents wow in March and are still "two weeks away" in September.
If your organisation has piloted an AI agent, you will recognise these.
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
01
the-lost-weekend
"Three senior architects, two weekends, one deadline. The people who bill the most, writing documents that usually lose."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
02
the-copy-paste-bid.
"Last year's response with the client's name changed. Except the one place it wasn't."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
03
the-promise-delivery-discovered-later.
"Somewhere on page forty, we committed to something. Delivery found out after we won."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
04
the-loss-with-no-lesson.
"We lost. Procurement said 'scoring'. Nobody knows which answer cost us, so the next bid repeats it."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
01
the-specialist-queue.
"Good MuleSoft people are scarce and expensive. Your whole programme moves at the speed of the three you have."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
02
the-api-that-already-existed.
"Nobody checks the catalogue under deadline pressure. Three teams have now built the same customer lookup, three ways."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
03
the-rules-that-live-in-a-wiki.
"The three-layer rules and naming standards are written down beautifully. On a deadline, they lose."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
04
the-handover-that-stops-at-go-live.
"The delivery partner rolled off. The documentation stops at go-live, and the people who knew why things were built that way went with them."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
01
the-quote-that-starts-with-discovery.
"The vendor's plan begins with a nine-month discovery phase, before anything moves, because nobody can say what is actually on the old platform."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
02
the-list-nobody-has.
"Ask three people how many connections run on the old platform and you get three numbers, hundreds apart."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
03
the-ones-that-might-be-dead.
"Some connections run every night for nobody. You keep paying for them, because switching one off is scarier than keeping it alive."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
04
the-last-attempt.
"The previous migration stalled partway. The people who ran it have moved on, and nobody is sure which connections were finished."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
01
lost-in-translation.
"What the business asked for and what got built are cousins, not twins. Nobody can say where it drifted, because nothing was written down with sources."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
02
the-prototype-that-shipped.
"The demo impressed everyone, so it quietly became the product. No error handling, no tests, no documentation. Three years later, everyone is afraid of it."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
03
the-creeping-scope.
"Every sprint adds a little. By month six the budget is spent and the core still is not finished."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
04
the-documentation-debt.
"The documentation was written in the last week, from memory, by whoever was left."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
01
pilot-purgatory.
"The demo wowed the board in March. It is September, and launch is still two weeks away, because nobody can say what would make it safe to say yes."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
02
the-improviser.
"In the demo it answered perfectly. Nobody knows what it says when a customer swears at it, lies to it, or asks it for a discount."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
03
the-3x-estimate.-copy
'What is our AI agent strategy?' Every vendor answers 'buy ours'. Nobody answers the real question: how would we trust one?
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
04
the-unowned-promise.
"If the agent promises a customer a refund, whose promise is it? Today, nobody has signed for that."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
01
the-out-of-date-chart.
"The chart showing how your systems connect was last correct in 2021. Three of the people on it have left the company."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
02
the-connection-nobody-dares-touch.
"One old connection moves payroll data. The person who built it left. Nobody touches it, so every project builds an expensive detour around it."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
03
the-3x-estimate.
"A two-week change is quoted at ten weeks, because eight of them go on working out what it might break."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
04
the-shelf.
"The improvements you wanted in 2024 are still a slide in a deck. Not for lack of value, but because changing running systems feels too risky and too expensive."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
01
the-deck-that-never-became-delivery.
"The last assessment cost a fortune, took five months, and lives in a drawer. Beautiful slides. Nothing shipped."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
02
the-interview-truth
"It was built from forty interviews. People told what they remembered, defended what they built, and nobody mentioned the system nobody owns."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
03
the-assessment-that-expired.
By the time decisions were made, the estate had moved on. The document described last year."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
04
the-roadmap-with-no-numbers-behind-it
"The plan said eighteen months and a big figure. Asked how it was calculated, the consultants said: experience."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
01
the-object-nobody-can-read.
Hundreds of fields added over a decade. Nobody knows which are used, so nobody dares remove them.
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
02
the-friday-release-that-broke-monday.
A new automation collided with an old one on the same records.
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
03
the-custom-code-nobody-remembers-why.
Custom code now does what a standard feature can do out of the box.
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
04
the-queue-behind-three-admins.
Every department’s asks funnel into one backlog behind the same few people.
flip card
THE EXAM
A demo shows what an agent can do. The exam shows what it will do.
Before any agent meets a customer, Hapie builds its exam: hundreds of realistic scenarios drawn from your actual work: the routine asks, the ambiguous ones, the angry customer, the customer who asks for something the agent must not give, and the one who tries to trick it. Every scenario has a defined right behaviour, agreed with your team. The agent runs the whole exam, is scored on accuracy, safety, staying inside its guardrails, knowing when to hand over, and cost per interaction, and the results go to a named person. No pass, no licence. And the exam is kept: it runs again on every change.
"You would not let a new employee near customers without training and supervision. An agent gets the same, in writing."
THE TO-BE, AS OPTIONS WITH TRADE-OFFS
Not one recommendation. Options, priced in trade-offs, for you to choose.
From the as-is, Hapie lays out target options: typically a steady path (fix the riskiest things, keep the shape), a balanced path (consolidate and modernise in waves), and a bold path (replatform what matters most). Each option states, with its evidence: what changes, what it makes possible, what it risks, roughly what it takes, drawn from measured past deliveries, and what it deliberately leaves alone. The choice is yours; the evidence is the same for every option.
Genuinely custom. Built to the integration pattern, documented, and added to the map for the next team.
Genuinely custom. Built to the integration pattern, documented, and added to the map for the next team.
Genuinely custom. Built to the integration pattern, documented, and added to the map for the next team.
Genuinely custom. Built to the integration pattern, documented, and added to the map for the next team.
Genuinely custom. Built to the integration pattern, documented, and added to the map for the next team.
Genuinely custom. Built to the integration pattern, documented, and added to the map for the next team.
Genuinely custom. Built to the integration pattern, documented, and added to the map for the next team.
Genuinely custom. Built to the integration pattern, documented, and added to the map for the next team.
Genuinely custom. Built to the integration pattern, documented, and added to the map for the next team.
Genuinely custom. Built to the integration pattern, documented, and added to the map for the next team.
ONE AGENT, DELIVERED
From job description to working agent, with proof at every step.
One real agent, walked end to end: an order-status agent in Agentforce.
The agent gets a job description.
What it is for, who it serves, what good looks like, and its guardrails, written and approved before anything is built, like any role worth hiring for.
The exam is built with your team.
Bring one request from your backlog. Leave with it scoped.
The agent is built into your platform.
Into Agentforce here, connected only to the systems its guardrails allow, with its logging wired in from the start.
It sits the exam.
The full scenario set, scored. Weaknesses are fixed and it resits. No pass, no licence: see the checkpoint moment below.
A named person signs the launch.
The licence: exam results, guardrails, monitoring plan and rollback path, approved in one place. It launches to a limited group first.
It works, watched, and gets better.
Every action logged, hand-overs reviewed, the exam grows from real conversations, and every model change triggers a re-exam before the agent runs on it.
THE EXAM
A demo shows what an agent can do. The exam shows what it will do.
Before any agent meets a customer, Hapie builds its exam: hundreds of realistic scenarios drawn from your actual work: the routine asks, the ambiguous ones, the angry customer, the customer who asks for something the agent must not give, and the one who tries to trick it. Every scenario has a defined right behaviour, agreed with your team. The agent runs the whole exam, is scored on accuracy, safety, staying inside its guardrails, knowing when to hand over, and cost per interaction, and the results go to a named person. No pass, no licence. And the exam is kept: it runs again on every change.
"You would not let a new employee near customers without training and supervision. An agent gets the same, in writing."
Start
Bring one request from your backlog. Leave with it scoped.
Fifteen minutes on an org like yours: see the map, make the standard-before-custom call, and see the collision check in action.
Read-only to start · Ships through your own release process · Reply within two working days
The three doubts, answered straight
What Salesforce leaders ask.
Is the 95% failure statistic real?
The study is real, the number is loose. MIT's Project NANDA published The GenAI Divide in July 2025 based on 52 interviews and 153 survey responses. Its traceable finding is that 60% of organisations evaluated these tools, 20% piloted and 5% reached production. The "95% get zero return" framing has been publicly questioned by academics for not being derivable from the document.
If agents only succeed 35% of the time on multi-turn tasks, why use them?
Some of them genuinely do, and their judgement matters: that is the context step, and it is welcome. What no person can do is hold two hundred systems' connections in their head, current, at once. Evidence from the systems plus judgement from people beats either alone, and this page is the first half done properly.
How long does it take to get a first agent into production properly?
Someone accountable for the outcome in the estate, not someone accountable for the platform. In practice that is usually the person who would be asked to explain the change to a regulator. If nobody can be named in a sentence, that is the finding, and it is worth surfacing before an auditor surfaces it.
Does the EU AI Act apply to us if we only use agents internally?
Yes. The map is built from the org as it is, debt included, turning cleanup into ordinary evidenced changes.
Is Salesforce Shield enough on its own?
Some of them genuinely do, and their judgement matters: that is the context step, and it is welcome. What no person can do is hold two hundred systems' connections in their head, current, at once. Evidence from the systems plus judgement from people beats either alone, and this page is the first half done properly.
Who should be the named approver for agent work?
Someone accountable for the outcome in the estate, not someone accountable for the platform. In practice that is usually the person who would be asked to explain the change to a regulator. If nobody can be named in a sentence, that is the finding, and it is worth surfacing before an auditor surfaces it.
Is this not just a traceability matrix?
Yes. The map is built from the org as it is, debt included, turning cleanup into ordinary evidenced changes.
Does this slow delivery down by adding approval steps?
Some of them genuinely do, and their judgement matters: that is the context step, and it is welcome. What no person can do is hold two hundred systems' connections in their head, current, at once. Evidence from the systems plus judgement from people beats either alone, and this page is the first half done properly.
How does this work on a brownfield estate with ten years of history?
That is the harder and more common case. The map is built from what already exists, including the parts nobody documented. Long-lived estates usually see the biggest difference, because reconstruction cost per transition is highest where the decisions are oldest.
"Our bids are our secret sauce."
Yes. The map is built from the org as it is, debt included, turning cleanup into ordinary evidenced changes.
"Every RFP is different."
Some of them genuinely do, and their judgement matters: that is the context step, and it is welcome. What no person can do is hold two hundred systems' connections in their head, current, at once. Evidence from the systems plus judgement from people beats either alone, and this page is the first half done properly.
"Will it read like AI wrote it?"
The opposite, and here is why: what makes bids read machine-made is generic filler, and filler exists because humans padding forty pages at midnight reach for it. Here, every sentence must attach to your record or it does not survive the checkpoint, so what remains is specific, in your voice, and signed by your people. The blandest bid your buyer will read this quarter was written by exhausted humans.
"Our MuleSoft estate is already a mess."
Yes. The map is built from the org as it is, debt included, turning cleanup into ordinary evidenced changes.
"We already have accelerators and templates."
Some of them genuinely do, and their judgement matters: that is the context step, and it is welcome. What no person can do is hold two hundred systems' connections in their head, current, at once. Evidence from the systems plus judgement from people beats either alone, and this page is the first half done properly.
"Will it fit our Anypoint setup and our release process?"
Hapie reads your Anypoint estate to build the map and the catalogue view, and everything ships through your own pipelines and approvals. Your setup stays your setup; Hapie supplies the method and the proof.
"We tried this before and it stalled."
Yes. The map is built from the org as it is, debt included, turning cleanup into ordinary evidenced changes.
"We cannot freeze the business for the duration."
Some of them genuinely do, and their judgement matters: that is the context step, and it is welcome. What no person can do is hold two hundred systems' connections in their head, current, at once. Evidence from the systems plus judgement from people beats either alone, and this page is the first half done properly.
"What if the new platform behaves differently?"
That is what the side-by-side run exists to catch. Every connection processes the same real inputs on both platforms until the outputs match, line by line, and the quiet edge cases, rounding, empty fields, odd dates, are exactly what it makes loud.
"AI tools already build apps from a prompt. Why is this different?"
Yes. The map is built from the org as it is, debt included, turning cleanup into ordinary evidenced changes.
"Our requirements always change mid-build."
Some of them genuinely do, and their judgement matters: that is the context step, and it is welcome. What no person can do is hold two hundred systems' connections in their head, current, at once. Evidence from the systems plus judgement from people beats either alone, and this page is the first half done properly.
"We do not have people free to write requirements for months."
You do not write them; you approve them. The agents draft requirements from the workshops and documents you already have, and your people answer a short open-questions list and sign the scope. The heavy documentation lifting is exactly what is automated.
"We are already building in Agentforce."
Yes. The map is built from the org as it is, debt included, turning cleanup into ordinary evidenced changes.
"Models change constantly. Won't the agent drift?"
Some of them genuinely do, and their judgement matters: that is the context step, and it is welcome. What no person can do is hold two hundred systems' connections in their head, current, at once. Evidence from the systems plus judgement from people beats either alone, and this page is the first half done properly.
"Our legal and compliance teams will never sign this."
They will not sign vibes, and they are right not to. What they can sign is what this process manufactures: written limits with named owners, exam results against agreed behaviours, complete action logs, and a rollback path. Several sections of this site, Security, and Documents you share with us, are written to be sent to them first. Bring them to the demo; they are the easiest audience we have.
"Our systems are a mess. No tool can read them."
Yes. The map is built from the org as it is, debt included, turning cleanup into ordinary evidenced changes.
"We cannot let an AI change systems that are running.
Some of them genuinely do, and their judgement matters: that is the context step, and it is welcome. What no person can do is hold two hundred systems' connections in their head, current, at once. Evidence from the systems plus judgement from people beats either alone, and this page is the first half done properly.
"The people who understood the old parts have left."
Much of what they knew is still in what they left behind: the code, the tickets, the history of changes. Hapie recovers that part. What cannot be recovered is confirmed once by your team, then kept on the map, so it is never lost again.
"We did an assessment last year."
Yes. The map is built from the org as it is, debt included, turning cleanup into ordinary evidenced changes.
"Our consultants already know our estate"
Some of them genuinely do, and their judgement matters: that is the context step, and it is welcome. What no person can do is hold two hundred systems' connections in their head, current, at once. Evidence from the systems plus judgement from people beats either alone, and this page is the first half done properly.
"We cannot open our systems to an outside tool."
It is read-only, it runs in your environment, and building the as-is changes nothing. The access conversation is with your security team on your terms, and the Documents and Security pages on this site are written to be sent to them first.
“Our org is a decade of debt. Start there?”
Yes. The map is built from the org as it is, debt included, turning cleanup into ordinary evidenced changes.
“We already have admins and a partner.”
Keep them. They become the approvers and standard-setters while the agents do the building.
“Will this fight our release process?”
No. It ships through your sandboxes, pipelines and release calendar. What it adds is the map and checkpoint.
USE CASE · AI AGENT DELIVERY · AGENTFORCE AND BEYOND
AI agents in production, not in pilot. Because yours pass an exam first.
Anyone can demo an agent. Hapie delivers them: requirements and guardrails written down, an exam of hundreds of realistic scenarios it must pass, a named person approving the launch, every action logged in production, and a re-exam every time the underlying model changes. Into Agentforce, into your platforms, on your terms.
THE FOUR THINGS EVERY AGENT PROGRAMME RECOGNISES
Why agents wow in March and are still "two weeks away" in September.
If your organisation has piloted an AI agent, you will recognise these.
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
01
the-lost-weekend
"Three senior architects, two weekends, one deadline. The people who bill the most, writing documents that usually lose."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
02
the-copy-paste-bid.
"Last year's response with the client's name changed. Except the one place it wasn't."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
03
the-promise-delivery-discovered-later.
"Somewhere on page forty, we committed to something. Delivery found out after we won."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
04
the-loss-with-no-lesson.
"We lost. Procurement said 'scoring'. Nobody knows which answer cost us, so the next bid repeats it."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
01
the-specialist-queue.
"Good MuleSoft people are scarce and expensive. Your whole programme moves at the speed of the three you have."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
02
the-api-that-already-existed.
"Nobody checks the catalogue under deadline pressure. Three teams have now built the same customer lookup, three ways."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
03
the-rules-that-live-in-a-wiki.
"The three-layer rules and naming standards are written down beautifully. On a deadline, they lose."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
04
the-handover-that-stops-at-go-live.
"The delivery partner rolled off. The documentation stops at go-live, and the people who knew why things were built that way went with them."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
01
the-quote-that-starts-with-discovery.
"The vendor's plan begins with a nine-month discovery phase, before anything moves, because nobody can say what is actually on the old platform."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
02
the-list-nobody-has.
"Ask three people how many connections run on the old platform and you get three numbers, hundreds apart."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
03
the-ones-that-might-be-dead.
"Some connections run every night for nobody. You keep paying for them, because switching one off is scarier than keeping it alive."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
04
the-last-attempt.
"The previous migration stalled partway. The people who ran it have moved on, and nobody is sure which connections were finished."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
01
lost-in-translation.
"What the business asked for and what got built are cousins, not twins. Nobody can say where it drifted, because nothing was written down with sources."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
02
the-prototype-that-shipped.
"The demo impressed everyone, so it quietly became the product. No error handling, no tests, no documentation. Three years later, everyone is afraid of it."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
03
the-creeping-scope.
"Every sprint adds a little. By month six the budget is spent and the core still is not finished."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
04
the-documentation-debt.
"The documentation was written in the last week, from memory, by whoever was left."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
01
pilot-purgatory.
"The demo wowed the board in March. It is September, and launch is still two weeks away, because nobody can say what would make it safe to say yes."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
02
the-improviser.
"In the demo it answered perfectly. Nobody knows what it says when a customer swears at it, lies to it, or asks it for a discount."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
03
the-3x-estimate.-copy
'What is our AI agent strategy?' Every vendor answers 'buy ours'. Nobody answers the real question: how would we trust one?
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
04
the-unowned-promise.
"If the agent promises a customer a refund, whose promise is it? Today, nobody has signed for that."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
01
the-out-of-date-chart.
"The chart showing how your systems connect was last correct in 2021. Three of the people on it have left the company."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
02
the-connection-nobody-dares-touch.
"One old connection moves payroll data. The person who built it left. Nobody touches it, so every project builds an expensive detour around it."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
03
the-3x-estimate.
"A two-week change is quoted at ten weeks, because eight of them go on working out what it might break."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
04
the-shelf.
"The improvements you wanted in 2024 are still a slide in a deck. Not for lack of value, but because changing running systems feels too risky and too expensive."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
01
the-deck-that-never-became-delivery.
"The last assessment cost a fortune, took five months, and lives in a drawer. Beautiful slides. Nothing shipped."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
02
the-interview-truth
"It was built from forty interviews. People told what they remembered, defended what they built, and nobody mentioned the system nobody owns."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
03
the-assessment-that-expired.
By the time decisions were made, the estate had moved on. The document described last year."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
04
the-roadmap-with-no-numbers-behind-it
"The plan said eighteen months and a big figure. Asked how it was calculated, the consultants said: experience."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
01
the-object-nobody-can-read.
Hundreds of fields added over a decade. Nobody knows which are used, so nobody dares remove them.
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
02
the-friday-release-that-broke-monday.
A new automation collided with an old one on the same records.
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
03
the-custom-code-nobody-remembers-why.
Custom code now does what a standard feature can do out of the box.
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
04
the-queue-behind-three-admins.
Every department’s asks funnel into one backlog behind the same few people.
flip card
THE EXAM
A demo shows what an agent can do. The exam shows what it will do.
Before any agent meets a customer, Hapie builds its exam: hundreds of realistic scenarios drawn from your actual work: the routine asks, the ambiguous ones, the angry customer, the customer who asks for something the agent must not give, and the one who tries to trick it. Every scenario has a defined right behaviour, agreed with your team. The agent runs the whole exam, is scored on accuracy, safety, staying inside its guardrails, knowing when to hand over, and cost per interaction, and the results go to a named person. No pass, no licence. And the exam is kept: it runs again on every change.
"You would not let a new employee near customers without training and supervision. An agent gets the same, in writing."
THE TO-BE, AS OPTIONS WITH TRADE-OFFS
Not one recommendation. Options, priced in trade-offs, for you to choose.
From the as-is, Hapie lays out target options: typically a steady path (fix the riskiest things, keep the shape), a balanced path (consolidate and modernise in waves), and a bold path (replatform what matters most). Each option states, with its evidence: what changes, what it makes possible, what it risks, roughly what it takes, drawn from measured past deliveries, and what it deliberately leaves alone. The choice is yours; the evidence is the same for every option.
Genuinely custom. Built to the integration pattern, documented, and added to the map for the next team.
Genuinely custom. Built to the integration pattern, documented, and added to the map for the next team.
Genuinely custom. Built to the integration pattern, documented, and added to the map for the next team.
Genuinely custom. Built to the integration pattern, documented, and added to the map for the next team.
Genuinely custom. Built to the integration pattern, documented, and added to the map for the next team.
Genuinely custom. Built to the integration pattern, documented, and added to the map for the next team.
Genuinely custom. Built to the integration pattern, documented, and added to the map for the next team.
Genuinely custom. Built to the integration pattern, documented, and added to the map for the next team.
Genuinely custom. Built to the integration pattern, documented, and added to the map for the next team.
Genuinely custom. Built to the integration pattern, documented, and added to the map for the next team.
ONE AGENT, DELIVERED
From job description to working agent, with proof at every step.
One real agent, walked end to end: an order-status agent in Agentforce.
The agent gets a job description.
What it is for, who it serves, what good looks like, and its guardrails, written and approved before anything is built, like any role worth hiring for.
The exam is built with your team.
Bring one request from your backlog. Leave with it scoped.
The agent is built into your platform.
Into Agentforce here, connected only to the systems its guardrails allow, with its logging wired in from the start.
It sits the exam.
The full scenario set, scored. Weaknesses are fixed and it resits. No pass, no licence: see the checkpoint moment below.
A named person signs the launch.
The licence: exam results, guardrails, monitoring plan and rollback path, approved in one place. It launches to a limited group first.
It works, watched, and gets better.
Every action logged, hand-overs reviewed, the exam grows from real conversations, and every model change triggers a re-exam before the agent runs on it.
THE EXAM
A demo shows what an agent can do. The exam shows what it will do.
Before any agent meets a customer, Hapie builds its exam: hundreds of realistic scenarios drawn from your actual work: the routine asks, the ambiguous ones, the angry customer, the customer who asks for something the agent must not give, and the one who tries to trick it. Every scenario has a defined right behaviour, agreed with your team. The agent runs the whole exam, is scored on accuracy, safety, staying inside its guardrails, knowing when to hand over, and cost per interaction, and the results go to a named person. No pass, no licence. And the exam is kept: it runs again on every change.
"You would not let a new employee near customers without training and supervision. An agent gets the same, in writing."
Start
Bring one request from your backlog. Leave with it scoped.
Fifteen minutes on an org like yours: see the map, make the standard-before-custom call, and see the collision check in action.
Read-only to start · Ships through your own release process · Reply within two working days
The three doubts, answered straight
What Salesforce leaders ask.
Is the 95% failure statistic real?
The study is real, the number is loose. MIT's Project NANDA published The GenAI Divide in July 2025 based on 52 interviews and 153 survey responses. Its traceable finding is that 60% of organisations evaluated these tools, 20% piloted and 5% reached production. The "95% get zero return" framing has been publicly questioned by academics for not being derivable from the document.
If agents only succeed 35% of the time on multi-turn tasks, why use them?
Some of them genuinely do, and their judgement matters: that is the context step, and it is welcome. What no person can do is hold two hundred systems' connections in their head, current, at once. Evidence from the systems plus judgement from people beats either alone, and this page is the first half done properly.
How long does it take to get a first agent into production properly?
Someone accountable for the outcome in the estate, not someone accountable for the platform. In practice that is usually the person who would be asked to explain the change to a regulator. If nobody can be named in a sentence, that is the finding, and it is worth surfacing before an auditor surfaces it.
Does the EU AI Act apply to us if we only use agents internally?
Yes. The map is built from the org as it is, debt included, turning cleanup into ordinary evidenced changes.
Is Salesforce Shield enough on its own?
Some of them genuinely do, and their judgement matters: that is the context step, and it is welcome. What no person can do is hold two hundred systems' connections in their head, current, at once. Evidence from the systems plus judgement from people beats either alone, and this page is the first half done properly.
Who should be the named approver for agent work?
Someone accountable for the outcome in the estate, not someone accountable for the platform. In practice that is usually the person who would be asked to explain the change to a regulator. If nobody can be named in a sentence, that is the finding, and it is worth surfacing before an auditor surfaces it.
Is this not just a traceability matrix?
Yes. The map is built from the org as it is, debt included, turning cleanup into ordinary evidenced changes.
Does this slow delivery down by adding approval steps?
Some of them genuinely do, and their judgement matters: that is the context step, and it is welcome. What no person can do is hold two hundred systems' connections in their head, current, at once. Evidence from the systems plus judgement from people beats either alone, and this page is the first half done properly.
How does this work on a brownfield estate with ten years of history?
That is the harder and more common case. The map is built from what already exists, including the parts nobody documented. Long-lived estates usually see the biggest difference, because reconstruction cost per transition is highest where the decisions are oldest.
"Our bids are our secret sauce."
Yes. The map is built from the org as it is, debt included, turning cleanup into ordinary evidenced changes.
"Every RFP is different."
Some of them genuinely do, and their judgement matters: that is the context step, and it is welcome. What no person can do is hold two hundred systems' connections in their head, current, at once. Evidence from the systems plus judgement from people beats either alone, and this page is the first half done properly.
"Will it read like AI wrote it?"
The opposite, and here is why: what makes bids read machine-made is generic filler, and filler exists because humans padding forty pages at midnight reach for it. Here, every sentence must attach to your record or it does not survive the checkpoint, so what remains is specific, in your voice, and signed by your people. The blandest bid your buyer will read this quarter was written by exhausted humans.
"Our MuleSoft estate is already a mess."
Yes. The map is built from the org as it is, debt included, turning cleanup into ordinary evidenced changes.
"We already have accelerators and templates."
Some of them genuinely do, and their judgement matters: that is the context step, and it is welcome. What no person can do is hold two hundred systems' connections in their head, current, at once. Evidence from the systems plus judgement from people beats either alone, and this page is the first half done properly.
"Will it fit our Anypoint setup and our release process?"
Hapie reads your Anypoint estate to build the map and the catalogue view, and everything ships through your own pipelines and approvals. Your setup stays your setup; Hapie supplies the method and the proof.
"We tried this before and it stalled."
Yes. The map is built from the org as it is, debt included, turning cleanup into ordinary evidenced changes.
"We cannot freeze the business for the duration."
Some of them genuinely do, and their judgement matters: that is the context step, and it is welcome. What no person can do is hold two hundred systems' connections in their head, current, at once. Evidence from the systems plus judgement from people beats either alone, and this page is the first half done properly.
"What if the new platform behaves differently?"
That is what the side-by-side run exists to catch. Every connection processes the same real inputs on both platforms until the outputs match, line by line, and the quiet edge cases, rounding, empty fields, odd dates, are exactly what it makes loud.
"AI tools already build apps from a prompt. Why is this different?"
Yes. The map is built from the org as it is, debt included, turning cleanup into ordinary evidenced changes.
"Our requirements always change mid-build."
Some of them genuinely do, and their judgement matters: that is the context step, and it is welcome. What no person can do is hold two hundred systems' connections in their head, current, at once. Evidence from the systems plus judgement from people beats either alone, and this page is the first half done properly.
"We do not have people free to write requirements for months."
You do not write them; you approve them. The agents draft requirements from the workshops and documents you already have, and your people answer a short open-questions list and sign the scope. The heavy documentation lifting is exactly what is automated.
"We are already building in Agentforce."
Yes. The map is built from the org as it is, debt included, turning cleanup into ordinary evidenced changes.
"Models change constantly. Won't the agent drift?"
Some of them genuinely do, and their judgement matters: that is the context step, and it is welcome. What no person can do is hold two hundred systems' connections in their head, current, at once. Evidence from the systems plus judgement from people beats either alone, and this page is the first half done properly.
"Our legal and compliance teams will never sign this."
They will not sign vibes, and they are right not to. What they can sign is what this process manufactures: written limits with named owners, exam results against agreed behaviours, complete action logs, and a rollback path. Several sections of this site, Security, and Documents you share with us, are written to be sent to them first. Bring them to the demo; they are the easiest audience we have.
"Our systems are a mess. No tool can read them."
Yes. The map is built from the org as it is, debt included, turning cleanup into ordinary evidenced changes.
"We cannot let an AI change systems that are running.
Some of them genuinely do, and their judgement matters: that is the context step, and it is welcome. What no person can do is hold two hundred systems' connections in their head, current, at once. Evidence from the systems plus judgement from people beats either alone, and this page is the first half done properly.
"The people who understood the old parts have left."
Much of what they knew is still in what they left behind: the code, the tickets, the history of changes. Hapie recovers that part. What cannot be recovered is confirmed once by your team, then kept on the map, so it is never lost again.
"We did an assessment last year."
Yes. The map is built from the org as it is, debt included, turning cleanup into ordinary evidenced changes.
"Our consultants already know our estate"
Some of them genuinely do, and their judgement matters: that is the context step, and it is welcome. What no person can do is hold two hundred systems' connections in their head, current, at once. Evidence from the systems plus judgement from people beats either alone, and this page is the first half done properly.
"We cannot open our systems to an outside tool."
It is read-only, it runs in your environment, and building the as-is changes nothing. The access conversation is with your security team on your terms, and the Documents and Security pages on this site are written to be sent to them first.
“Our org is a decade of debt. Start there?”
Yes. The map is built from the org as it is, debt included, turning cleanup into ordinary evidenced changes.
“We already have admins and a partner.”
Keep them. They become the approvers and standard-setters while the agents do the building.
“Will this fight our release process?”
No. It ships through your sandboxes, pipelines and release calendar. What it adds is the map and checkpoint.
USE CASE · BROWNFIELD CHANG
Update the systems your business runs on. In weeks, not months.
Hapie first builds a map of your systems and how they really connect. Then, for every change, it shows exactly what will be affected, does the work, tests it against everything connected, and a named person on your team approves it before it goes live.
THE FOUR THINGS EVERY TEAM RECOGNISES
Why a two-week change gets quoted at ten weeks.
If your company has run its systems for more than a few years, you will recognise these.
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
01
the-lost-weekend
"Three senior architects, two weekends, one deadline. The people who bill the most, writing documents that usually lose."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
02
the-copy-paste-bid.
"Last year's response with the client's name changed. Except the one place it wasn't."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
03
the-promise-delivery-discovered-later.
"Somewhere on page forty, we committed to something. Delivery found out after we won."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
04
the-loss-with-no-lesson.
"We lost. Procurement said 'scoring'. Nobody knows which answer cost us, so the next bid repeats it."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
01
the-specialist-queue.
"Good MuleSoft people are scarce and expensive. Your whole programme moves at the speed of the three you have."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
02
the-api-that-already-existed.
"Nobody checks the catalogue under deadline pressure. Three teams have now built the same customer lookup, three ways."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
03
the-rules-that-live-in-a-wiki.
"The three-layer rules and naming standards are written down beautifully. On a deadline, they lose."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
04
the-handover-that-stops-at-go-live.
"The delivery partner rolled off. The documentation stops at go-live, and the people who knew why things were built that way went with them."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
01
the-quote-that-starts-with-discovery.
"The vendor's plan begins with a nine-month discovery phase, before anything moves, because nobody can say what is actually on the old platform."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
02
the-list-nobody-has.
"Ask three people how many connections run on the old platform and you get three numbers, hundreds apart."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
03
the-ones-that-might-be-dead.
"Some connections run every night for nobody. You keep paying for them, because switching one off is scarier than keeping it alive."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
04
the-last-attempt.
"The previous migration stalled partway. The people who ran it have moved on, and nobody is sure which connections were finished."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
01
lost-in-translation.
"What the business asked for and what got built are cousins, not twins. Nobody can say where it drifted, because nothing was written down with sources."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
02
the-prototype-that-shipped.
"The demo impressed everyone, so it quietly became the product. No error handling, no tests, no documentation. Three years later, everyone is afraid of it."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
03
the-creeping-scope.
"Every sprint adds a little. By month six the budget is spent and the core still is not finished."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
04
the-documentation-debt.
"The documentation was written in the last week, from memory, by whoever was left."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
01
pilot-purgatory.
"The demo wowed the board in March. It is September, and launch is still two weeks away, because nobody can say what would make it safe to say yes."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
02
the-improviser.
"In the demo it answered perfectly. Nobody knows what it says when a customer swears at it, lies to it, or asks it for a discount."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
03
the-3x-estimate.-copy
'What is our AI agent strategy?' Every vendor answers 'buy ours'. Nobody answers the real question: how would we trust one?
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
04
the-unowned-promise.
"If the agent promises a customer a refund, whose promise is it? Today, nobody has signed for that."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
01
the-out-of-date-chart.
"The chart showing how your systems connect was last correct in 2021. Three of the people on it have left the company."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
02
the-connection-nobody-dares-touch.
"One old connection moves payroll data. The person who built it left. Nobody touches it, so every project builds an expensive detour around it."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
03
the-3x-estimate.
"A two-week change is quoted at ten weeks, because eight of them go on working out what it might break."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
04
the-shelf.
"The improvements you wanted in 2024 are still a slide in a deck. Not for lack of value, but because changing running systems feels too risky and too expensive."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
01
the-deck-that-never-became-delivery.
"The last assessment cost a fortune, took five months, and lives in a drawer. Beautiful slides. Nothing shipped."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
02
the-interview-truth
"It was built from forty interviews. People told what they remembered, defended what they built, and nobody mentioned the system nobody owns."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
03
the-assessment-that-expired.
By the time decisions were made, the estate had moved on. The document described last year."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
04
the-roadmap-with-no-numbers-behind-it
"The plan said eighteen months and a big figure. Asked how it was calculated, the consultants said: experience."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
01
the-object-nobody-can-read.
Hundreds of fields added over a decade. Nobody knows which are used, so nobody dares remove them.
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
02
the-friday-release-that-broke-monday.
A new automation collided with an old one on the same records.
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
03
the-custom-code-nobody-remembers-why.
Custom code now does what a standard feature can do out of the box.
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
04
the-queue-behind-three-admins.
Every department’s asks funnel into one backlog behind the same few people.
flip card
THE MAP
First, Hapie builds a live map of your systems, from the systems themselves.
Hapie connects in read-only mode to the tools you already use: your project tracker (Jira or Azure DevOps), your documentation (Confluence), your code, and platforms such as MuleSoft and Salesforce. From what is actually there, it builds one map: every system, every connection between them, what data moves where, and which of your rules apply. Where something is unclear, it never guesses: it adds it to a short list of open questions for your team to confirm.
"Your documents can be out of date. Your running systems cannot."
THE TO-BE, AS OPTIONS WITH TRADE-OFFS
Not one recommendation. Options, priced in trade-offs, for you to choose.
From the as-is, Hapie lays out target options: typically a steady path (fix the riskiest things, keep the shape), a balanced path (consolidate and modernise in waves), and a bold path (replatform what matters most). Each option states, with its evidence: what changes, what it makes possible, what it risks, roughly what it takes, drawn from measured past deliveries, and what it deliberately leaves alone. The choice is yours; the evidence is the same for every option.
Genuinely custom. Built to the integration pattern, documented, and added to the map for the next team.
Genuinely custom. Built to the integration pattern, documented, and added to the map for the next team.
Genuinely custom. Built to the integration pattern, documented, and added to the map for the next team.
Genuinely custom. Built to the integration pattern, documented, and added to the map for the next team.
Genuinely custom. Built to the integration pattern, documented, and added to the map for the next team.
Genuinely custom. Built to the integration pattern, documented, and added to the map for the next team.
Genuinely custom. Built to the integration pattern, documented, and added to the map for the next team.
Genuinely custom. Built to the integration pattern, documented, and added to the map for the next team.
Genuinely custom. Built to the integration pattern, documented, and added to the map for the next team.
Genuinely custom. Built to the integration pattern, documented, and added to the map for the next team.
ONE POSTPONED CHANGE, TAKEN OFF THE SHELF AND DELIVERED
From the shelf to live, with proof at every step
One real change, walked end to end.
A postponed task comes off the shelf
"Show live stock levels on the customer portal", postponed twice. Hapie proposes the plan and an estimate based on measured past deliveries, not guesswork.
The affected-systems list is attached.
Bring one request from your backlog. Leave with it scoped.
The design names your real systems.
It refers to your actual connections and your team's rules by name; the unlisted system gets a protection step. Your architect approves.
The work is done inside guardrails
AI agents build the change using your approved patterns, and automatic checks run on every piece as it is made.
The checkpoint proves it safe.
The change is tested against every connected system on the map, every claim must show its evidence, and a named person signs.
It goes live with a full record, and the map gets better.
Release evidence, refreshed documentation, and everything learned is kept, so the next change starts warmer.
THE MAP
First, Hapie builds a live map of your systems, from the systems themselves.
Hapie connects in read-only mode to the tools you already use: your project tracker (Jira or Azure DevOps), your documentation (Confluence), your code, and platforms such as MuleSoft and Salesforce. From what is actually there, it builds one map: every system, every connection between them, what data moves where, and which of your rules apply. Where something is unclear, it never guesses: it adds it to a short list of open questions for your team to confirm.
"Your documents can be out of date. Your running systems cannot."
Start
Bring one request from your backlog. Leave with it scoped.
Fifteen minutes on an org like yours: see the map, make the standard-before-custom call, and see the collision check in action.
Read-only to start · Ships through your own release process · Reply within two working days
The three doubts, answered straight
What Salesforce leaders ask.
Is the 95% failure statistic real?
The study is real, the number is loose. MIT's Project NANDA published The GenAI Divide in July 2025 based on 52 interviews and 153 survey responses. Its traceable finding is that 60% of organisations evaluated these tools, 20% piloted and 5% reached production. The "95% get zero return" framing has been publicly questioned by academics for not being derivable from the document.
If agents only succeed 35% of the time on multi-turn tasks, why use them?
Some of them genuinely do, and their judgement matters: that is the context step, and it is welcome. What no person can do is hold two hundred systems' connections in their head, current, at once. Evidence from the systems plus judgement from people beats either alone, and this page is the first half done properly.
How long does it take to get a first agent into production properly?
Someone accountable for the outcome in the estate, not someone accountable for the platform. In practice that is usually the person who would be asked to explain the change to a regulator. If nobody can be named in a sentence, that is the finding, and it is worth surfacing before an auditor surfaces it.
Does the EU AI Act apply to us if we only use agents internally?
Yes. The map is built from the org as it is, debt included, turning cleanup into ordinary evidenced changes.
Is Salesforce Shield enough on its own?
Some of them genuinely do, and their judgement matters: that is the context step, and it is welcome. What no person can do is hold two hundred systems' connections in their head, current, at once. Evidence from the systems plus judgement from people beats either alone, and this page is the first half done properly.
Who should be the named approver for agent work?
Someone accountable for the outcome in the estate, not someone accountable for the platform. In practice that is usually the person who would be asked to explain the change to a regulator. If nobody can be named in a sentence, that is the finding, and it is worth surfacing before an auditor surfaces it.
Is this not just a traceability matrix?
Yes. The map is built from the org as it is, debt included, turning cleanup into ordinary evidenced changes.
Does this slow delivery down by adding approval steps?
Some of them genuinely do, and their judgement matters: that is the context step, and it is welcome. What no person can do is hold two hundred systems' connections in their head, current, at once. Evidence from the systems plus judgement from people beats either alone, and this page is the first half done properly.
How does this work on a brownfield estate with ten years of history?
That is the harder and more common case. The map is built from what already exists, including the parts nobody documented. Long-lived estates usually see the biggest difference, because reconstruction cost per transition is highest where the decisions are oldest.
"Our bids are our secret sauce."
Yes. The map is built from the org as it is, debt included, turning cleanup into ordinary evidenced changes.
"Every RFP is different."
Some of them genuinely do, and their judgement matters: that is the context step, and it is welcome. What no person can do is hold two hundred systems' connections in their head, current, at once. Evidence from the systems plus judgement from people beats either alone, and this page is the first half done properly.
"Will it read like AI wrote it?"
The opposite, and here is why: what makes bids read machine-made is generic filler, and filler exists because humans padding forty pages at midnight reach for it. Here, every sentence must attach to your record or it does not survive the checkpoint, so what remains is specific, in your voice, and signed by your people. The blandest bid your buyer will read this quarter was written by exhausted humans.
"Our MuleSoft estate is already a mess."
Yes. The map is built from the org as it is, debt included, turning cleanup into ordinary evidenced changes.
"We already have accelerators and templates."
Some of them genuinely do, and their judgement matters: that is the context step, and it is welcome. What no person can do is hold two hundred systems' connections in their head, current, at once. Evidence from the systems plus judgement from people beats either alone, and this page is the first half done properly.
"Will it fit our Anypoint setup and our release process?"
Hapie reads your Anypoint estate to build the map and the catalogue view, and everything ships through your own pipelines and approvals. Your setup stays your setup; Hapie supplies the method and the proof.
"We tried this before and it stalled."
Yes. The map is built from the org as it is, debt included, turning cleanup into ordinary evidenced changes.
"We cannot freeze the business for the duration."
Some of them genuinely do, and their judgement matters: that is the context step, and it is welcome. What no person can do is hold two hundred systems' connections in their head, current, at once. Evidence from the systems plus judgement from people beats either alone, and this page is the first half done properly.
"What if the new platform behaves differently?"
That is what the side-by-side run exists to catch. Every connection processes the same real inputs on both platforms until the outputs match, line by line, and the quiet edge cases, rounding, empty fields, odd dates, are exactly what it makes loud.
"AI tools already build apps from a prompt. Why is this different?"
Yes. The map is built from the org as it is, debt included, turning cleanup into ordinary evidenced changes.
"Our requirements always change mid-build."
Some of them genuinely do, and their judgement matters: that is the context step, and it is welcome. What no person can do is hold two hundred systems' connections in their head, current, at once. Evidence from the systems plus judgement from people beats either alone, and this page is the first half done properly.
"We do not have people free to write requirements for months."
You do not write them; you approve them. The agents draft requirements from the workshops and documents you already have, and your people answer a short open-questions list and sign the scope. The heavy documentation lifting is exactly what is automated.
"We are already building in Agentforce."
Yes. The map is built from the org as it is, debt included, turning cleanup into ordinary evidenced changes.
"Models change constantly. Won't the agent drift?"
Some of them genuinely do, and their judgement matters: that is the context step, and it is welcome. What no person can do is hold two hundred systems' connections in their head, current, at once. Evidence from the systems plus judgement from people beats either alone, and this page is the first half done properly.
"Our legal and compliance teams will never sign this."
They will not sign vibes, and they are right not to. What they can sign is what this process manufactures: written limits with named owners, exam results against agreed behaviours, complete action logs, and a rollback path. Several sections of this site, Security, and Documents you share with us, are written to be sent to them first. Bring them to the demo; they are the easiest audience we have.
"Our systems are a mess. No tool can read them."
Yes. The map is built from the org as it is, debt included, turning cleanup into ordinary evidenced changes.
"We cannot let an AI change systems that are running.
Some of them genuinely do, and their judgement matters: that is the context step, and it is welcome. What no person can do is hold two hundred systems' connections in their head, current, at once. Evidence from the systems plus judgement from people beats either alone, and this page is the first half done properly.
"The people who understood the old parts have left."
Much of what they knew is still in what they left behind: the code, the tickets, the history of changes. Hapie recovers that part. What cannot be recovered is confirmed once by your team, then kept on the map, so it is never lost again.
"We did an assessment last year."
Yes. The map is built from the org as it is, debt included, turning cleanup into ordinary evidenced changes.
"Our consultants already know our estate"
Some of them genuinely do, and their judgement matters: that is the context step, and it is welcome. What no person can do is hold two hundred systems' connections in their head, current, at once. Evidence from the systems plus judgement from people beats either alone, and this page is the first half done properly.
"We cannot open our systems to an outside tool."
It is read-only, it runs in your environment, and building the as-is changes nothing. The access conversation is with your security team on your terms, and the Documents and Security pages on this site are written to be sent to them first.
“Our org is a decade of debt. Start there?”
Yes. The map is built from the org as it is, debt included, turning cleanup into ordinary evidenced changes.
“We already have admins and a partner.”
Keep them. They become the approvers and standard-setters while the agents do the building.
“Will this fight our release process?”
No. It ships through your sandboxes, pipelines and release calendar. What it adds is the map and checkpoint.
USE CASE · AS-IS TO TO-BE ROADMA
Know where you stand. Decide where to go. In weeks, with evidence
Hapie reads your systems directly, in read-only mode, and builds the honest as-is: what exists, what connects to what, what is healthy, what is risky, and what is unclear. Then it lays out target options with trade-offs, and a roadmap in waves with estimates drawn from measured past deliveries. Every finding links to its evidence, and the roadmap converts straight into delivery when you say go.
THE FOUR THINGS EVERY ASSESSMENT OWNER RECOGNISES
Why assessments cost a fortune and change nothing
If you have ever commissioned one, you will recognise these.
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
01
the-lost-weekend
"Three senior architects, two weekends, one deadline. The people who bill the most, writing documents that usually lose."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
02
the-copy-paste-bid.
"Last year's response with the client's name changed. Except the one place it wasn't."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
03
the-promise-delivery-discovered-later.
"Somewhere on page forty, we committed to something. Delivery found out after we won."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
04
the-loss-with-no-lesson.
"We lost. Procurement said 'scoring'. Nobody knows which answer cost us, so the next bid repeats it."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
01
the-specialist-queue.
"Good MuleSoft people are scarce and expensive. Your whole programme moves at the speed of the three you have."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
02
the-api-that-already-existed.
"Nobody checks the catalogue under deadline pressure. Three teams have now built the same customer lookup, three ways."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
03
the-rules-that-live-in-a-wiki.
"The three-layer rules and naming standards are written down beautifully. On a deadline, they lose."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
04
the-handover-that-stops-at-go-live.
"The delivery partner rolled off. The documentation stops at go-live, and the people who knew why things were built that way went with them."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
01
the-quote-that-starts-with-discovery.
"The vendor's plan begins with a nine-month discovery phase, before anything moves, because nobody can say what is actually on the old platform."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
02
the-list-nobody-has.
"Ask three people how many connections run on the old platform and you get three numbers, hundreds apart."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
03
the-ones-that-might-be-dead.
"Some connections run every night for nobody. You keep paying for them, because switching one off is scarier than keeping it alive."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
04
the-last-attempt.
"The previous migration stalled partway. The people who ran it have moved on, and nobody is sure which connections were finished."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
01
lost-in-translation.
"What the business asked for and what got built are cousins, not twins. Nobody can say where it drifted, because nothing was written down with sources."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
02
the-prototype-that-shipped.
"The demo impressed everyone, so it quietly became the product. No error handling, no tests, no documentation. Three years later, everyone is afraid of it."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
03
the-creeping-scope.
"Every sprint adds a little. By month six the budget is spent and the core still is not finished."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
04
the-documentation-debt.
"The documentation was written in the last week, from memory, by whoever was left."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
01
pilot-purgatory.
"The demo wowed the board in March. It is September, and launch is still two weeks away, because nobody can say what would make it safe to say yes."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
02
the-improviser.
"In the demo it answered perfectly. Nobody knows what it says when a customer swears at it, lies to it, or asks it for a discount."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
03
the-3x-estimate.-copy
'What is our AI agent strategy?' Every vendor answers 'buy ours'. Nobody answers the real question: how would we trust one?
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
04
the-unowned-promise.
"If the agent promises a customer a refund, whose promise is it? Today, nobody has signed for that."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
01
the-out-of-date-chart.
"The chart showing how your systems connect was last correct in 2021. Three of the people on it have left the company."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
02
the-connection-nobody-dares-touch.
"One old connection moves payroll data. The person who built it left. Nobody touches it, so every project builds an expensive detour around it."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
03
the-3x-estimate.
"A two-week change is quoted at ten weeks, because eight of them go on working out what it might break."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
04
the-shelf.
"The improvements you wanted in 2024 are still a slide in a deck. Not for lack of value, but because changing running systems feels too risky and too expensive."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
01
the-deck-that-never-became-delivery.
"The last assessment cost a fortune, took five months, and lives in a drawer. Beautiful slides. Nothing shipped."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
02
the-interview-truth
"It was built from forty interviews. People told what they remembered, defended what they built, and nobody mentioned the system nobody owns."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
03
the-assessment-that-expired.
By the time decisions were made, the estate had moved on. The document described last year."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
04
the-roadmap-with-no-numbers-behind-it
"The plan said eighteen months and a big figure. Asked how it was calculated, the consultants said: experience."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
01
the-object-nobody-can-read.
Hundreds of fields added over a decade. Nobody knows which are used, so nobody dares remove them.
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
02
the-friday-release-that-broke-monday.
A new automation collided with an old one on the same records.
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
03
the-custom-code-nobody-remembers-why.
Custom code now does what a standard feature can do out of the box.
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
04
the-queue-behind-three-admins.
Every department’s asks funnel into one backlog behind the same few people.
flip card
THE AS-IS, READ FROM THE SYSTEMS THEMSELVES
An assessment built from interviews is a survey of opinions. Yours is built from the systems.
Hapie connects read-only to what you run: your platforms, your code, your project tracker, your documentation, your deployment records and traffic. From that it builds the as-is: every system and connection, what each does in plain words, what depends on it, where the risk concentrates, and, honestly, what is unclear. Unclear things become an open-questions list for your team, not a guess in a slide. Your people and ours add the context machines cannot know: history, politics, plans. But every finding in the final picture links to evidence, not to memory alone.
"The systems do not misremember, defend their choices, or leave the company."
THE TO-BE, AS OPTIONS WITH TRADE-OFFS
Not one recommendation. Options, priced in trade-offs, for you to choose.
From the as-is, Hapie lays out target options: typically a steady path (fix the riskiest things, keep the shape), a balanced path (consolidate and modernise in waves), and a bold path (replatform what matters most). Each option states, with its evidence: what changes, what it makes possible, what it risks, roughly what it takes, drawn from measured past deliveries, and what it deliberately leaves alone. The choice is yours; the evidence is the same for every option.
Genuinely custom. Built to the integration pattern, documented, and added to the map for the next team.
Genuinely custom. Built to the integration pattern, documented, and added to the map for the next team.
Genuinely custom. Built to the integration pattern, documented, and added to the map for the next team.
Genuinely custom. Built to the integration pattern, documented, and added to the map for the next team.
Genuinely custom. Built to the integration pattern, documented, and added to the map for the next team.
Genuinely custom. Built to the integration pattern, documented, and added to the map for the next team.
Genuinely custom. Built to the integration pattern, documented, and added to the map for the next team.
Genuinely custom. Built to the integration pattern, documented, and added to the map for the next team.
Genuinely custom. Built to the integration pattern, documented, and added to the map for the next team.
Genuinely custom. Built to the integration pattern, documented, and added to the map for the next team.
FROM KICKOFF TO AN APPROVED ROADMAP
From first access to a decided roadmap, with proof at every step.
One real change, walked end to end.
Read-only access, and the survey begins.
No workshops needed to start. Hapie reads the systems and builds the map while calendars are still being negotiated.
The honest as-is lands.
Bring one request from your backlog. Leave with it scoped.
Context is added by people.
Your team and ours add what machines cannot know: history, constraints, plans already in motion. Context is recorded next to evidence, never instead of it.
The options are laid out.
Steady, balanced, bold, each with trade-offs and estimates drawn from measured past deliveries. Your leaders choose with the dial in their hands.
The roadmap is approved.
Waves, dependencies, estimates with their working shown, and a named person signs it, the same checkpoint discipline as delivery.
It converts into delivery on day one.
The approved roadmap becomes the backlog, the map is already built, and wave one starts without a re-discovery phase.
THE AS-IS, READ FROM THE SYSTEMS THEMSELVES
An assessment built from interviews is a survey of opinions. Yours is built from the systems.
Hapie connects read-only to what you run: your platforms, your code, your project tracker, your documentation, your deployment records and traffic. From that it builds the as-is: every system and connection, what each does in plain words, what depends on it, where the risk concentrates, and, honestly, what is unclear. Unclear things become an open-questions list for your team, not a guess in a slide. Your people and ours add the context machines cannot know: history, politics, plans. But every finding in the final picture links to evidence, not to memory alone.
"The systems do not misremember, defend their choices, or leave the company."
Start
Bring one request from your backlog. Leave with it scoped.
Fifteen minutes on an org like yours: see the map, make the standard-before-custom call, and see the collision check in action.
Read-only to start · Ships through your own release process · Reply within two working days
The three doubts, answered straight
What Salesforce leaders ask.
Is the 95% failure statistic real?
The study is real, the number is loose. MIT's Project NANDA published The GenAI Divide in July 2025 based on 52 interviews and 153 survey responses. Its traceable finding is that 60% of organisations evaluated these tools, 20% piloted and 5% reached production. The "95% get zero return" framing has been publicly questioned by academics for not being derivable from the document.
If agents only succeed 35% of the time on multi-turn tasks, why use them?
Some of them genuinely do, and their judgement matters: that is the context step, and it is welcome. What no person can do is hold two hundred systems' connections in their head, current, at once. Evidence from the systems plus judgement from people beats either alone, and this page is the first half done properly.
How long does it take to get a first agent into production properly?
Someone accountable for the outcome in the estate, not someone accountable for the platform. In practice that is usually the person who would be asked to explain the change to a regulator. If nobody can be named in a sentence, that is the finding, and it is worth surfacing before an auditor surfaces it.
Does the EU AI Act apply to us if we only use agents internally?
Yes. The map is built from the org as it is, debt included, turning cleanup into ordinary evidenced changes.
Is Salesforce Shield enough on its own?
Some of them genuinely do, and their judgement matters: that is the context step, and it is welcome. What no person can do is hold two hundred systems' connections in their head, current, at once. Evidence from the systems plus judgement from people beats either alone, and this page is the first half done properly.
Who should be the named approver for agent work?
Someone accountable for the outcome in the estate, not someone accountable for the platform. In practice that is usually the person who would be asked to explain the change to a regulator. If nobody can be named in a sentence, that is the finding, and it is worth surfacing before an auditor surfaces it.
Is this not just a traceability matrix?
Yes. The map is built from the org as it is, debt included, turning cleanup into ordinary evidenced changes.
Does this slow delivery down by adding approval steps?
Some of them genuinely do, and their judgement matters: that is the context step, and it is welcome. What no person can do is hold two hundred systems' connections in their head, current, at once. Evidence from the systems plus judgement from people beats either alone, and this page is the first half done properly.
How does this work on a brownfield estate with ten years of history?
That is the harder and more common case. The map is built from what already exists, including the parts nobody documented. Long-lived estates usually see the biggest difference, because reconstruction cost per transition is highest where the decisions are oldest.
"Our bids are our secret sauce."
Yes. The map is built from the org as it is, debt included, turning cleanup into ordinary evidenced changes.
"Every RFP is different."
Some of them genuinely do, and their judgement matters: that is the context step, and it is welcome. What no person can do is hold two hundred systems' connections in their head, current, at once. Evidence from the systems plus judgement from people beats either alone, and this page is the first half done properly.
"Will it read like AI wrote it?"
The opposite, and here is why: what makes bids read machine-made is generic filler, and filler exists because humans padding forty pages at midnight reach for it. Here, every sentence must attach to your record or it does not survive the checkpoint, so what remains is specific, in your voice, and signed by your people. The blandest bid your buyer will read this quarter was written by exhausted humans.
"Our MuleSoft estate is already a mess."
Yes. The map is built from the org as it is, debt included, turning cleanup into ordinary evidenced changes.
"We already have accelerators and templates."
Some of them genuinely do, and their judgement matters: that is the context step, and it is welcome. What no person can do is hold two hundred systems' connections in their head, current, at once. Evidence from the systems plus judgement from people beats either alone, and this page is the first half done properly.
"Will it fit our Anypoint setup and our release process?"
Hapie reads your Anypoint estate to build the map and the catalogue view, and everything ships through your own pipelines and approvals. Your setup stays your setup; Hapie supplies the method and the proof.
"We tried this before and it stalled."
Yes. The map is built from the org as it is, debt included, turning cleanup into ordinary evidenced changes.
"We cannot freeze the business for the duration."
Some of them genuinely do, and their judgement matters: that is the context step, and it is welcome. What no person can do is hold two hundred systems' connections in their head, current, at once. Evidence from the systems plus judgement from people beats either alone, and this page is the first half done properly.
"What if the new platform behaves differently?"
That is what the side-by-side run exists to catch. Every connection processes the same real inputs on both platforms until the outputs match, line by line, and the quiet edge cases, rounding, empty fields, odd dates, are exactly what it makes loud.
"AI tools already build apps from a prompt. Why is this different?"
Yes. The map is built from the org as it is, debt included, turning cleanup into ordinary evidenced changes.
"Our requirements always change mid-build."
Some of them genuinely do, and their judgement matters: that is the context step, and it is welcome. What no person can do is hold two hundred systems' connections in their head, current, at once. Evidence from the systems plus judgement from people beats either alone, and this page is the first half done properly.
"We do not have people free to write requirements for months."
You do not write them; you approve them. The agents draft requirements from the workshops and documents you already have, and your people answer a short open-questions list and sign the scope. The heavy documentation lifting is exactly what is automated.
"We are already building in Agentforce."
Yes. The map is built from the org as it is, debt included, turning cleanup into ordinary evidenced changes.
"Models change constantly. Won't the agent drift?"
Some of them genuinely do, and their judgement matters: that is the context step, and it is welcome. What no person can do is hold two hundred systems' connections in their head, current, at once. Evidence from the systems plus judgement from people beats either alone, and this page is the first half done properly.
"Our legal and compliance teams will never sign this."
They will not sign vibes, and they are right not to. What they can sign is what this process manufactures: written limits with named owners, exam results against agreed behaviours, complete action logs, and a rollback path. Several sections of this site, Security, and Documents you share with us, are written to be sent to them first. Bring them to the demo; they are the easiest audience we have.
"Our systems are a mess. No tool can read them."
Yes. The map is built from the org as it is, debt included, turning cleanup into ordinary evidenced changes.
"We cannot let an AI change systems that are running.
Some of them genuinely do, and their judgement matters: that is the context step, and it is welcome. What no person can do is hold two hundred systems' connections in their head, current, at once. Evidence from the systems plus judgement from people beats either alone, and this page is the first half done properly.
"The people who understood the old parts have left."
Much of what they knew is still in what they left behind: the code, the tickets, the history of changes. Hapie recovers that part. What cannot be recovered is confirmed once by your team, then kept on the map, so it is never lost again.
"We did an assessment last year."
Yes. The map is built from the org as it is, debt included, turning cleanup into ordinary evidenced changes.
"Our consultants already know our estate"
Some of them genuinely do, and their judgement matters: that is the context step, and it is welcome. What no person can do is hold two hundred systems' connections in their head, current, at once. Evidence from the systems plus judgement from people beats either alone, and this page is the first half done properly.
"We cannot open our systems to an outside tool."
It is read-only, it runs in your environment, and building the as-is changes nothing. The access conversation is with your security team on your terms, and the Documents and Security pages on this site are written to be sent to them first.
“Our org is a decade of debt. Start there?”
Yes. The map is built from the org as it is, debt included, turning cleanup into ordinary evidenced changes.
“We already have admins and a partner.”
Keep them. They become the approvers and standard-setters while the agents do the building.
“Will this fight our release process?”
No. It ships through your sandboxes, pipelines and release calendar. What it adds is the map and checkpoint.
Use case · Salesforce delivery
Your Salesforce backlog, delivered in weeks. Your org, cleaner than you found it.
Hapie maps your org: what exists, what is actually used, and what depends on what. Then it delivers your backlog with a standard-first rule, checks every change against everything else running in the org, and a named person approves before anything goes live.
The four things every Salesforce team recognises
Why the backlog grows and the org slows.
If you run or deliver Salesforce work, you will recognise these.
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
01
the-lost-weekend
"Three senior architects, two weekends, one deadline. The people who bill the most, writing documents that usually lose."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
02
the-copy-paste-bid.
"Last year's response with the client's name changed. Except the one place it wasn't."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
03
the-promise-delivery-discovered-later.
"Somewhere on page forty, we committed to something. Delivery found out after we won."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
04
the-loss-with-no-lesson.
"We lost. Procurement said 'scoring'. Nobody knows which answer cost us, so the next bid repeats it."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
01
the-specialist-queue.
"Good MuleSoft people are scarce and expensive. Your whole programme moves at the speed of the three you have."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
02
the-api-that-already-existed.
"Nobody checks the catalogue under deadline pressure. Three teams have now built the same customer lookup, three ways."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
03
the-rules-that-live-in-a-wiki.
"The three-layer rules and naming standards are written down beautifully. On a deadline, they lose."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
04
the-handover-that-stops-at-go-live.
"The delivery partner rolled off. The documentation stops at go-live, and the people who knew why things were built that way went with them."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
01
the-quote-that-starts-with-discovery.
"The vendor's plan begins with a nine-month discovery phase, before anything moves, because nobody can say what is actually on the old platform."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
02
the-list-nobody-has.
"Ask three people how many connections run on the old platform and you get three numbers, hundreds apart."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
03
the-ones-that-might-be-dead.
"Some connections run every night for nobody. You keep paying for them, because switching one off is scarier than keeping it alive."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
04
the-last-attempt.
"The previous migration stalled partway. The people who ran it have moved on, and nobody is sure which connections were finished."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
01
lost-in-translation.
"What the business asked for and what got built are cousins, not twins. Nobody can say where it drifted, because nothing was written down with sources."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
02
the-prototype-that-shipped.
"The demo impressed everyone, so it quietly became the product. No error handling, no tests, no documentation. Three years later, everyone is afraid of it."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
03
the-creeping-scope.
"Every sprint adds a little. By month six the budget is spent and the core still is not finished."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
04
the-documentation-debt.
"The documentation was written in the last week, from memory, by whoever was left."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
01
pilot-purgatory.
"The demo wowed the board in March. It is September, and launch is still two weeks away, because nobody can say what would make it safe to say yes."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
02
the-improviser.
"In the demo it answered perfectly. Nobody knows what it says when a customer swears at it, lies to it, or asks it for a discount."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
03
the-3x-estimate.-copy
'What is our AI agent strategy?' Every vendor answers 'buy ours'. Nobody answers the real question: how would we trust one?
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
04
the-unowned-promise.
"If the agent promises a customer a refund, whose promise is it? Today, nobody has signed for that."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
01
the-out-of-date-chart.
"The chart showing how your systems connect was last correct in 2021. Three of the people on it have left the company."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
02
the-connection-nobody-dares-touch.
"One old connection moves payroll data. The person who built it left. Nobody touches it, so every project builds an expensive detour around it."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
03
the-3x-estimate.
"A two-week change is quoted at ten weeks, because eight of them go on working out what it might break."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
04
the-shelf.
"The improvements you wanted in 2024 are still a slide in a deck. Not for lack of value, but because changing running systems feels too risky and too expensive."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
01
the-deck-that-never-became-delivery.
"The last assessment cost a fortune, took five months, and lives in a drawer. Beautiful slides. Nothing shipped."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
02
the-interview-truth
"It was built from forty interviews. People told what they remembered, defended what they built, and nobody mentioned the system nobody owns."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
03
the-assessment-that-expired.
By the time decisions were made, the estate had moved on. The document described last year."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
04
the-roadmap-with-no-numbers-behind-it
"The plan said eighteen months and a big figure. Asked how it was calculated, the consultants said: experience."
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
01
the-object-nobody-can-read.
Hundreds of fields added over a decade. Nobody knows which are used, so nobody dares remove them.
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
02
the-friday-release-that-broke-monday.
A new automation collided with an old one on the same records.
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
03
the-custom-code-nobody-remembers-why.
Custom code now does what a standard feature can do out of the box.
flip card
The problem arrived in our own projects
NexGen Architects delivers software for enterprises. When AI entered the toolchain, output speed multiplied, and so did the doubt. Code appeared in hours that would once have taken weeks, but who had checked it? Which requirement did it trace to? What would we tell a client’s auditor? The industry’s answer was, in effect, “trust us”. That is not an answer a serious delivery firm can give.
STEP
04
the-queue-behind-three-admins.
Every department’s asks funnel into one backlog behind the same few people.
flip card
The map of your org
First, Hapie asks the org, not the admin who left.
Hapie reads your org directly, in read-only mode: objects and fields, automations and code, integrations, and what is actually used. From that it builds the map: what exists, what depends on what, what is duplicated, and what only looks alive.
“Your org remembers everything, including the things your team no longer does.”
Standard before custom
The cheapest customisation is the one you never build.
Custom code is how orgs get powerful, and also how they get slow, fragile and expensive to change. So the playbook's first rule is standard before custom: every request is checked against what Salesforce can do out of the box and what your org already has, before anything new is built. When custom is genuinely needed, it is built to the patterns, documented, and added to the map, so it never becomes the 2019 mystery. The org's future maintainers are treated as customers of every change.
Genuinely custom. Built to the integration pattern, documented, and added to the map for the next team.
Genuinely custom. Built to the integration pattern, documented, and added to the map for the next team.
Genuinely custom. Built to the integration pattern, documented, and added to the map for the next team.
Genuinely custom. Built to the integration pattern, documented, and added to the map for the next team.
Genuinely custom. Built to the integration pattern, documented, and added to the map for the next team.
Genuinely custom. Built to the integration pattern, documented, and added to the map for the next team.
Genuinely custom. Built to the integration pattern, documented, and added to the map for the next team.
Genuinely custom. Built to the integration pattern, documented, and added to the map for the next team.
Genuinely custom. Built to the integration pattern, documented, and added to the map for the next team.
Genuinely custom. Built to the integration pattern, documented, and added to the map for the next team.
One change, delivered into a live org
From request to released, with proof at every step.
One real change, walked end to end.
A request arrives
The request meets the map and its existing dependencies.
Standard before custom
Bring one request from your backlog. Leave with it scoped.
Collision check
The change is tested against everything else that fires on the records.
Check the whole org for collisions.
The change is tested against everything else touching the same records—existing routing, escalation rules, integrations, and automations—so there are no Friday surprises.
The checkpoint proves it’s ready.
Tests pass. Evidence is captured. Rules are verified. A named owner signs off on the release.
Release through your process. Update the map.
The change is deployed using your organisation’s release process, with a complete record kept. The map updates automatically to reflect the organisation as it now exists.
The map of your org
First, Hapie asks the org, not the admin who left.
Hapie reads your org directly, in read-only mode: objects and fields, automations and code, integrations, and what is actually used. From that it builds the map: what exists, what depends on what, what is duplicated, and what only looks alive.
“Your org remembers everything, including the things your team no longer does.”
Start
Bring one request from your backlog. Leave with it scoped.
Fifteen minutes on an org like yours: see the map, make the standard-before-custom call, and see the collision check in action.
Read-only to start · Ships through your own release process · Reply within two working days
The three doubts, answered straight
What Salesforce leaders ask.
Is the 95% failure statistic real?
The study is real, the number is loose. MIT's Project NANDA published The GenAI Divide in July 2025 based on 52 interviews and 153 survey responses. Its traceable finding is that 60% of organisations evaluated these tools, 20% piloted and 5% reached production. The "95% get zero return" framing has been publicly questioned by academics for not being derivable from the document.
If agents only succeed 35% of the time on multi-turn tasks, why use them?
Some of them genuinely do, and their judgement matters: that is the context step, and it is welcome. What no person can do is hold two hundred systems' connections in their head, current, at once. Evidence from the systems plus judgement from people beats either alone, and this page is the first half done properly.
How long does it take to get a first agent into production properly?
Someone accountable for the outcome in the estate, not someone accountable for the platform. In practice that is usually the person who would be asked to explain the change to a regulator. If nobody can be named in a sentence, that is the finding, and it is worth surfacing before an auditor surfaces it.
Does the EU AI Act apply to us if we only use agents internally?
Yes. The map is built from the org as it is, debt included, turning cleanup into ordinary evidenced changes.
Is Salesforce Shield enough on its own?
Some of them genuinely do, and their judgement matters: that is the context step, and it is welcome. What no person can do is hold two hundred systems' connections in their head, current, at once. Evidence from the systems plus judgement from people beats either alone, and this page is the first half done properly.
Who should be the named approver for agent work?
Someone accountable for the outcome in the estate, not someone accountable for the platform. In practice that is usually the person who would be asked to explain the change to a regulator. If nobody can be named in a sentence, that is the finding, and it is worth surfacing before an auditor surfaces it.
Is this not just a traceability matrix?
Yes. The map is built from the org as it is, debt included, turning cleanup into ordinary evidenced changes.
Does this slow delivery down by adding approval steps?
Some of them genuinely do, and their judgement matters: that is the context step, and it is welcome. What no person can do is hold two hundred systems' connections in their head, current, at once. Evidence from the systems plus judgement from people beats either alone, and this page is the first half done properly.
How does this work on a brownfield estate with ten years of history?
That is the harder and more common case. The map is built from what already exists, including the parts nobody documented. Long-lived estates usually see the biggest difference, because reconstruction cost per transition is highest where the decisions are oldest.
"Our bids are our secret sauce."
Yes. The map is built from the org as it is, debt included, turning cleanup into ordinary evidenced changes.
"Every RFP is different."
Some of them genuinely do, and their judgement matters: that is the context step, and it is welcome. What no person can do is hold two hundred systems' connections in their head, current, at once. Evidence from the systems plus judgement from people beats either alone, and this page is the first half done properly.
"Will it read like AI wrote it?"
The opposite, and here is why: what makes bids read machine-made is generic filler, and filler exists because humans padding forty pages at midnight reach for it. Here, every sentence must attach to your record or it does not survive the checkpoint, so what remains is specific, in your voice, and signed by your people. The blandest bid your buyer will read this quarter was written by exhausted humans.
"Our MuleSoft estate is already a mess."
Yes. The map is built from the org as it is, debt included, turning cleanup into ordinary evidenced changes.
"We already have accelerators and templates."
Some of them genuinely do, and their judgement matters: that is the context step, and it is welcome. What no person can do is hold two hundred systems' connections in their head, current, at once. Evidence from the systems plus judgement from people beats either alone, and this page is the first half done properly.
"Will it fit our Anypoint setup and our release process?"
Hapie reads your Anypoint estate to build the map and the catalogue view, and everything ships through your own pipelines and approvals. Your setup stays your setup; Hapie supplies the method and the proof.
"We tried this before and it stalled."
Yes. The map is built from the org as it is, debt included, turning cleanup into ordinary evidenced changes.
"We cannot freeze the business for the duration."
Some of them genuinely do, and their judgement matters: that is the context step, and it is welcome. What no person can do is hold two hundred systems' connections in their head, current, at once. Evidence from the systems plus judgement from people beats either alone, and this page is the first half done properly.
"What if the new platform behaves differently?"
That is what the side-by-side run exists to catch. Every connection processes the same real inputs on both platforms until the outputs match, line by line, and the quiet edge cases, rounding, empty fields, odd dates, are exactly what it makes loud.
"AI tools already build apps from a prompt. Why is this different?"
Yes. The map is built from the org as it is, debt included, turning cleanup into ordinary evidenced changes.
"Our requirements always change mid-build."
Some of them genuinely do, and their judgement matters: that is the context step, and it is welcome. What no person can do is hold two hundred systems' connections in their head, current, at once. Evidence from the systems plus judgement from people beats either alone, and this page is the first half done properly.
"We do not have people free to write requirements for months."
You do not write them; you approve them. The agents draft requirements from the workshops and documents you already have, and your people answer a short open-questions list and sign the scope. The heavy documentation lifting is exactly what is automated.
"We are already building in Agentforce."
Yes. The map is built from the org as it is, debt included, turning cleanup into ordinary evidenced changes.
"Models change constantly. Won't the agent drift?"
Some of them genuinely do, and their judgement matters: that is the context step, and it is welcome. What no person can do is hold two hundred systems' connections in their head, current, at once. Evidence from the systems plus judgement from people beats either alone, and this page is the first half done properly.
"Our legal and compliance teams will never sign this."
They will not sign vibes, and they are right not to. What they can sign is what this process manufactures: written limits with named owners, exam results against agreed behaviours, complete action logs, and a rollback path. Several sections of this site, Security, and Documents you share with us, are written to be sent to them first. Bring them to the demo; they are the easiest audience we have.
"Our systems are a mess. No tool can read them."
Yes. The map is built from the org as it is, debt included, turning cleanup into ordinary evidenced changes.
"We cannot let an AI change systems that are running.
Some of them genuinely do, and their judgement matters: that is the context step, and it is welcome. What no person can do is hold two hundred systems' connections in their head, current, at once. Evidence from the systems plus judgement from people beats either alone, and this page is the first half done properly.
"The people who understood the old parts have left."
Much of what they knew is still in what they left behind: the code, the tickets, the history of changes. Hapie recovers that part. What cannot be recovered is confirmed once by your team, then kept on the map, so it is never lost again.
"We did an assessment last year."
Yes. The map is built from the org as it is, debt included, turning cleanup into ordinary evidenced changes.
"Our consultants already know our estate"
Some of them genuinely do, and their judgement matters: that is the context step, and it is welcome. What no person can do is hold two hundred systems' connections in their head, current, at once. Evidence from the systems plus judgement from people beats either alone, and this page is the first half done properly.
"We cannot open our systems to an outside tool."
It is read-only, it runs in your environment, and building the as-is changes nothing. The access conversation is with your security team on your terms, and the Documents and Security pages on this site are written to be sent to them first.
“Our org is a decade of debt. Start there?”
Yes. The map is built from the org as it is, debt included, turning cleanup into ordinary evidenced changes.
“We already have admins and a partner.”
Keep them. They become the approvers and standard-setters while the agents do the building.
“Will this fight our release process?”
No. It ships through your sandboxes, pipelines and release calendar. What it adds is the map and checkpoint.
