ENTERPRISE AI

A reference picture and a first roadmap

Draw where a request goes, who approves it, and how you will know. Then run one bounded pilot. Week ranges in a method are not a delivery quote.

Figure 1: A first roadmap: the decision and the risk, the evidence you can trust, then one bounded pilot.

1. A picture of the request, not of the vendor

A reference picture is a layer cake you can point at. Someone captures the intent and reviews the result. A gateway classifies the request and enforces routing, budget, and policy. An orchestrator plans, calls tools, and stops. Context assembly selects only what this user may see. Model services, managed or private, sit behind a version and a regression gate. Tools perform bounded actions. Knowledge is the documents, the lineage, and the evidence. Verification measures quality. Security and governance run across the life of the system. Observability and the bill sit where an operator can see them.

You do not build every layer for a pilot. You mark which layers already exist, and which one the pilot would be inventing in a hurry. The hurry layer is the risk.

2. The first roadmap is a sequence

Discovery comes first: the outcome, the data classification, the model paths you allow, the failures you will not accept, and who has to adopt the thing. Then the evidence and the access: where the truth lives, who may read it, and who stewards a correction. Then a bounded pilot on one workflow, with training and a measurement. Scale and the operating life come after, and they do not end.

The source method prints week ranges, and it says the calendar shrinks when phases overlap. Those ranges assume access, procurement, and a security review that cooperate. They are planning bands, not a Cloudadorn quote. Do not put them in a proposal as a promise. Overlap the work you are allowed to overlap. Do not overlap the discovery with the production write.

3. Name the role, or the work has no owner

A role dictionary is short. A sponsor accepts the risk and the spend. A business owner accepts the outcome. An architect keeps the picture honest. Security owns the threat model. A data steward owns the meaning and the correction. An evaluation lead owns the golden set. A service owner owns it after release. Someone owns the capacity and the bill. Someone owns adoption. If two names share a role, nobody has it.

The risks worth writing down before the pilot are also short. Untrusted content that changes behavior. Memory that gets poisoned. An eval set that does not look like production. Reviewers who rubber-stamp. A private deployment whose savings never show up. A model change that regresses. Ownership split across committees. Each one needs a mitigation and a name.

Layman rule: A roadmap without a named owner for the outcome is a schedule of meetings.

4. Principles you can test

Design the system, not only the prompt. Make the model replaceable, then prove it with a regression test. Treat retrieved content, memory, tools, and model output as untrusted until something checks them. Keep the source. Use structured state for facts and open questions. Combine semantic and exact search when the identifier matters. Keep high-impact math and state changes deterministic. Grant the least privilege before you grant more autonomy. Evaluate repeated performance, not one demo. Use the full bill. Expand autonomy only when the trace, the check, and the owner are real.

5. What to draw this week

One page. The user, the decision, the unacceptable failure, the layers you already have, the pilot workflow, and the names. Stop there. The scorecard is how you tell, later, whether the pilot became work: a weekly scorecard.