Built AI-first · Chapter 1 of 9
AI does not fix a process. It amplifies one. Point it at a function that was already sound and you get a faster sound function. Point it at the function that has been quietly broken for three years — the one everybody works around — and you get that, faster, with better grammar.
This is the chapter nobody wants, because almost every AI programme currently running is a retrofit, and it is producing real-feeling wins. Here is why those wins stall at about week six, and what to do instead.
The claim
The structure is the thing that decides who owns a piece of work, what they read before they act, what they are not allowed to do, and what happens when they fail. Almost no organisation has written any of that down. It lives in habit, in one long-serving person, and in the shape of an inbox.
A retrofit leaves all of that exactly where it was and inserts a very capable assistant into the middle of it. The assistant is genuinely good. The structure is unchanged. And a structure that cannot say who owns a thing does not get better when the thing arrives twice as fast.
The question is not whether AI made the task faster. It did. The question is whether the business can absorb a faster version of a task nobody owns.
The mechanism
Not "fails to help" — makes worse. These are the three we have watched happen, including to us.
Mechanism one
Every process has an error rate. Throughput scales; the error rate does not improve on its own.
If intake is a shared mailbox nobody owns, drafting replies faster produces more wrong replies, sooner. It is worse than the old state for a reason people miss: the old tell was that a rushed reply looked rushed. A confident, well-formed, wrong reply defeats the only quality check most small organisations actually have — the reader's instinct that something was dashed off.
Mechanism two
In a retrofit the AI sits inside one person's window, and what it read never leaves it.
Work enters the business with no record of the source it relied on or the reasoning it applied. Questions you could previously answer by asking a colleague — why did we quote that, where did that figure come from, who decided we do it this way — stop having answers. For a business that may be asked those questions by an auditor, a board, or a public-records request, this is not a productivity trade. It is a loss of the thing being traded for.
Mechanism three
Drafts, summaries and suggestions are not output. They are decisions waiting for the owner.
A retrofit multiplies artefacts, and every artefact needs somebody to accept, reject or fix it. That somebody is the owner, the chief, the principal — the person who was the constraint before any of this started. This is the mechanism that kills programmes, and it is cruel about it: for roughly six weeks it feels like enormous progress, because volume is up and nothing has broken yet.
The exit
Everyone looking for the way out reaches for a better model, a better tool, a better prompt. The way out is upstream of all three. It is the question you ask at the start.
● Code red · the retrofitter asks
Every honest answer to this question is a task: write this email, summarise this meeting, clean up this spreadsheet, draft this policy. Tasks are real and the help is real. But a task has no owner, no boundary, no cadence and no record — so nothing you build on top of it accumulates. Six months in you have a faster version of the same day and a chat history nobody can audit.
● Code green · the AI-first builder asks
Every honest answer to this question is a job: who owns intake, what they must read before acting, the decision rules they apply, what they escalate, what they may never do, and what happens when they fail. A job can be held to a standard. A job survives a model change. A job is an asset with a name on it.
A task is something you hand to a tool. A job is something you can hold someone to.
That is the entire shift, and it is why our own re-engineering did not start with a tool at all. It started with twenty written job descriptions — a professional background to reason from, a specialism, the sources the role is allowed to trust, its decision rules, what it escalates, and what it may never do. The document is the agent. Swap the model underneath and the role survives, which is the only sensible definition of an asset in a field moving this fast.
The test
Run these against the last AI thing your organisation adopted. They take about ten minutes and they are uncomfortable in a useful way.
| Ask | What the answer tells you |
|---|---|
| What do we own now that we did not own before? | A retrofit produces a chat history and a subscription. A re-engineering produces a written job, a body of context, a routine with a cadence, and a log. Three of those four appreciate. |
| If the model we use were withdrawn tomorrow, what survives? | If the honest answer is "nothing", the thing you built was a habit, not a structure. The job descriptions, the context and the routines are the parts that survive — and they are the parts almost nobody writes down. |
| Who does the extra output land on? | If every improvement ends in the owner's approval queue, you have made the bottleneck busier and called it a win. Real progress moves work away from the constraint, not toward it in nicer packaging. |
| Can we answer "how do you know?" about anything it produced? | If not, the speed was bought with traceability. In a regulated or public-facing organisation that is not a trade you were entitled to make on a Tuesday afternoon without telling anyone. |
In your world
An assistant that writes incident narratives and grant sections quickly is a genuine relief — and it is a retrofit if nobody changed who owns the record, what it must cite, or what happens when the underlying export is stale. The output now looks authoritative whether or not the data was fresh.
Drafting council packets and records responses faster is the most requested help in a small city. It is also where a retrofit is most expensive: a well-formed response to a public-records request that cites the wrong version of a policy is a disclosure problem, not a typo.
Quoting is the obvious first target and often a real win. The trap is that the quote was never the constraint — the follow-up was. Doubling quote output into a business with no owned follow-up produces more stale quotes, not more signed work.
A practice that can generate proposals in an afternoon, staffed by people who are billable, has moved its bottleneck from writing to reviewing — and review is done by the principal. That is mechanism three, arriving on schedule and disguised as capacity.
What we will not claim
Where this sits
Knowing that a retrofit amplifies structure is only useful once you can see which part of your own structure is going to be amplified. That is chapter two, and it is the one to read with a pen.