Built AI-first · Chapter 9 of 9
Ninety days, one year, three years. The dates are the least important part of this page — what matters is the order, because each stage is the precondition for the next and the usual failure is starting three stages in.
Measure, then context, then skills, then routines, then validation, then autonomy. Reorder it and you will be building on something you have not established.
The shape
Each stage below ends with a gate, and a gate is a question about evidence rather than a list of completed tasks. "Did we write the context library" is a milestone and it can be satisfied by a document nobody uses. "Can somebody outside the build answer how do you know about last month's output" is a gate, and it cannot.
We run these as outcome gates in the spirit of PMBOK 7 — principle-led and tailored to the organisation rather than a fixed deliverable checklist. The tailoring is real: a fire district and a four-person contractor run the same sequence at very different sizes, and the gate question is identical in both.
If a gate will not close, the answer is never to proceed anyway with a note. It is that the stage is not finished, and the next stage would have been built on sand.
Stage one
Month one
Nothing is automated this month, and the temptation to skip it is the single most common way this goes wrong. You are establishing what is currently true, so that anything you change later can be shown to have changed something.
Gate: you can state the current condition of the chosen zone from a record rather than from memory — and somebody else could read that record and reach the same description.
Stage two
Month two
Deliberately narrow. One written method and one standing cadence, in the zone you chose, at the lowest two rungs of autonomy.
Gate: the routine has run its full cycle at least three times, the log exists, and a failure has been observed and handled — an untested routine that has never failed has not been tested.
Stage three
Month three
This is where a working automation becomes something an organisation can defend. It is also the stage most programmes skip, because the first two months produced visible wins and this one does not.
Gate: a person who did not build it can pick any output from the last month and get a real answer to "how do you know?" — with a source and a date.
Stage four
Months 4–12
The ninety-day loop repeats. It gets faster — the second zone borrows the library, the validator and the queue from the first — but it does not get fast, and an organisation that claims to have done all seven in a year has usually done one properly and declared the rest.
Gate for year two: the organisation ran a normal week without its owner — not as an exercise, as a week that actually happened — and the people involved would describe it as ordinary.
Stage five
Years 2–3
By this point the interesting property is not what the agents do. It is what the organisation owns.
The real test: not a dashboard. Whether the week the owner is unreachable is an ordinary week — the same question as the seventh tell in chapter two, asked three years later.
The ordering
| If you skip… | What breaks later |
|---|---|
| Measurement | Every later decision is made against an estimate, and you will not be able to tell whether anything improved. This is the one we got wrong ourselves. |
| Context | Skills are written against facts that live in somebody's head, so the method is only as portable as the person who wrote it. |
| Skills | Routines run an unwritten method, which means every run is a fresh improvisation and nothing improves in one place. |
| Routines | There is nothing to grant autonomy over. Authority is granted to a routine, not to an agent — so without routines, chapter eight cannot be applied at all. |
| Validation | You reach the autonomy conversation with no evidence, and the honest answer to "how do you know it is reliable" is that you do not. |
In your world
Instrument one thing first: what is overdue in the inspection cycle, by how long, and in which occupancy classes. Few districts can answer that at the population level on a Monday morning, though every record needed is already held. Then write the notice method, give it a cadence and an owner, and keep release with a person. Administrative throughout — nothing near an operational system.
Constraints written first, then the method, then the clock as a routine, with release staying with a person permanently. Narrow scope is what makes this defensible at a council meeting.
Measure how many quotes go quiet and for how long — almost nobody knows. Then write the follow-up method and give it a cadence and an owner. Drafted, never sent automatically.
High frequency, immediately visible to the client, and the first thing to slip in a busy week. Writing it down as a method and a routine is a fortnight of work and it survives the next bad month.
What we will not claim
The end of the series
Which three things in your organisation are currently held together by one person remembering them? Everything in these nine pages is an answer to that, and the honest version of the answer usually takes about ten minutes to say out loud.