Ninety days is achievable. It is achievable specifically when the first six weeks are spent on the part everyone wants to skip.
Here is the sequence, and what happens when it is reordered.
Weeks 1–2: pick one decision and one entity
Not a use case. A decision — a specific choice the business makes repeatedly, with a known volume, a known current cost and a known owner.
And one entity that decision is about. Customer, supplier, asset, claim, patient. One.
The instinct is to scope broadly to justify the investment. Broad scope is why programmes take two years. A single decision with a single entity gets to production, and production is what unlocks the next one.
The test for a well-chosen decision: someone can tell you how many are made per week, how long each takes now, and what happens when one is wrong.
Weeks 2–6: the governed record
This is the part that gets compressed, and compressing it is the single best predictor of a programme that stalls in month five.
Enumerate the systems that hold the entity. Honestly — including the shared mailbox and the spreadsheet. Resolve it across them. Get survivorship rules written down and signed by an owner, per attribute. Stand up quarantine with a named owner and an expiry policy. Capture lineage from the first record.
Nothing here demos. All of it determines whether anything built on top can be defended.
Weeks 6–9: the decision path
Now build. Three evidence lanes: retrieval for grounding, deterministic checks for the non-negotiables, reasoning over both.
Write the eligibility policy before the model work, with the risk owner. Not a confidence threshold — an explicit test of when this decision may be made automatically.
Build the trace store in this window, not later. Retrofitting traces onto a running system is unpleasant and you will want them from the first shadow-mode decision.
Weeks 9–11: shadow mode
Run the whole path in production, on production volume, deciding nothing. Compare against what humans actually decided.
This is where you find the things no test caught: the 6% of cases with a data shape nobody mentioned, the policy exception that lives in one team's heads, the upstream system that changes its response format on Sundays.
Two weeks minimum. Resist shortening it. Shadow mode is the cheapest place you will ever find these.
Weeks 11–13: staged live
Start with the narrowest eligible slice — lowest value band, clearest cases. Widen weekly against measured agreement.
The blind sample starts here and never stops.
The two decisions that are hard to reverse
Where lineage is captured. If you do not record which specific records produced which specific decision, at write time, you cannot reconstruct it later. Sources move on. This is the decision most often deferred and least often recoverable.
How cost is attributed. Tag every request with its business context from the first day. Retrofitting attribution across a live estate is a project, and until you have it, every capacity and value conversation is opinion.
Everything else — model choice, chunking strategy, prompt design, even the vector store — is changeable in an afternoon or a sprint. These two are not.
What ninety days does not include
It does not include the second use case, which will be faster. It does not include change management, which runs in parallel and takes longer than the build. And it does not include the integration into the thirty-year-old system of record, which you should start in week one regardless of when you need it, because it will be the critical path.
- Delivery
- Programme
- Platform