Built AI-first · Chapter 3 of 9
Ask what the most valuable AI asset in a business is and you will hear "the model" or "the data". It is neither. It is the maintained body of writing that says what is true here, what we decided and why, what we will never do, and how we sound — the thing every agent reads before it acts.
Almost nobody builds it correctly, and the reasons are specific rather than mysterious. Five of them are below, along with the shape of an entry that actually works.
The claim
Every organisation using AI today is renting the intelligence and will be renting a different one within eighteen months. What does not change hands is the written answer to the questions a capable new hire would ask in their first fortnight: how do we price, what did we promise this client, which policy governs this, what did we decide about that and why, what must never leave the building.
Almost every organisation answers those questions from memory, one person at a time, forever. Writing them down once is the single highest-return piece of work in an AI programme, and it is unglamorous enough that it keeps getting deferred behind tool selection.
If the model you use were withdrawn tomorrow, what would you keep? Whatever you just named is your context library. For most businesses the honest answer is a subscription and some chat history.
What it contains
The reason "just upload our documents" does not produce a working library is that three of these four are not documents anywhere in the business. They have never been written down at all.
Kind one
What is true, with a date and a source.
Prices, terms, roster, service areas, thresholds, who owns what. The part people assume the library is. It is also the part that rots fastest, which is why an entry without an as-of date is worse than a missing one — it is confidently wrong at an unknown moment.
Kind two
What we chose, when, and the reasoning we chose it for.
The highest-value entries and the rarest. A document records what was written; a decision records what was settled — and settled decisions are what stop the same argument being had every quarter by people who were not in the room.
Kind three
What must never happen, and what we do not know.
Regulatory boundaries, contractual limits, claims we have retired, and — crucially — the list of things the organisation has not measured. A library that only contains conclusions produces confident answers in exactly the places where it should hesitate.
Kind four
How we sound, and what we would never say.
Not a brand deck. A short, blunt description of register, the words we do not use, and the two or three sentences that are actually ours. This is the difference between output you edit and output you send.
Why it goes wrong
In rough order of how often we meet them.
The shape
Five fields. It is deliberately boring, and the boredom is the point — an entry that takes four minutes to write is an entry that gets written.
| Field | Why it is there |
|---|---|
| The claim, in one sentence | If it needs a paragraph it is two entries. Single claims are what make a citation meaningful — "it relied on entry 14" has to mean something specific. |
| Source | Where this came from: a contract, a statute, a meeting, a measurement, or a judgement call by a named person. "A judgement call by the owner" is a perfectly good source, and far better than an unattributed assertion. |
| As of | The date the claim was last known true. Not the date the file was edited. This is the field that lets anything be aged, reviewed or retired. |
| Owner | One name. The person who will be asked whether it is still true, and whose job includes answering. |
| Status | Settled, provisional, or retired. Retired entries stay in the library with the reason — that is how you stop a claim you deliberately killed being rediscovered and reused eight months later. |
We keep our own settled decisions in exactly this form, dated, with the reasoning next to each one, under a heading that says they are settled and are to be quoted rather than re-argued. It reads as bureaucratic until the first time somebody reopens a question that cost a week to close, and you can answer it with a line and a date.
Decide once that an old receivable is not collectible, write it down, and no agent ever raises it again. That is what memory that compounds actually means.
In your world
Ten is the right size for a first pass. It is small enough to finish in an afternoon and large enough that the gaps become obvious.
Which occupancy classes sit on which re-check interval and why; what counts as a completed check; who may grant an extension and on what grounds; retention periods; which policy version is current. The dates are already in the records system. The judgements behind them are with the fire marshal, and they are decisions, not facts.
The records categories, the exemptions you actually rely on, the response clock, and who has authority to release. Constraints first here, because the cost of a confident wrong answer is highest in this zone.
Your rate structure, what is included, the jobs you no longer take and why, and the three customers you would not work for again. The refusals are decisions and they are almost never written down anywhere.
Scope, the commitments made verbally, the assumptions the estimate rested on, and the things you have already agreed not to do. This is the library that keeps delivery out of one person's head.
What we will not claim
Where this sits
A context library on its own is a reference book. It becomes an operating system when two other things sit beside it — the methods that use it and the schedule that runs them. That is the three-file architecture.