How I scope an AI project: the discovery-week template

Every engagement starts with a paid discovery week. Five working days, fixed fee, before the larger build is committed. The point is to know whether to commit at all, and on what terms.

Here is the shape, which has held across the last several engagements.

Day 1: the conversation that should have happened first. Two hours with the operator who knows the problem best (founder, COO, technical lead, lead clinician, whoever it is). I do not show up with a slide deck. I show up with a notebook and a list of questions. The deliverable is a written summary of the actual problem, in their words, by the end of the day.

Day 2: the data tour. I want to see the actual data, in the actual systems, in the state it is actually in. Not a clean export. The dashboard nobody uses, the Airtable that has the real numbers, the spreadsheet on someone's desktop. The point is to know whether the data exists, where it lives, and what shape it takes.

Day 3: the constraint pass. Compliance, vendor lock-in, hiring constraints, runway, internal politics that affect what can ship. Half of all AI engagements get re-scoped here, because constraints surface that change the answer. Better now than in week six.

Day 4: the architecture sketch. I write the smallest system that solves the problem. Not the most ambitious. Not the prettiest. The one I would actually ship in 8 weeks if I started Monday. Stack, integrations, data flow, evaluation strategy, deployment target.

Day 5: the writeup. Three documents:

The scoped engagement. Fixed scope, fixed price, milestones, an honest timeline, and the assumptions that would force a re-quote.
The risk register. The five things that could go wrong, ranked by likelihood, with mitigation notes.
The "we should not do this" report. If the answer at the end of the week is no, this is the document that says why and points you to a better path.

At the end of the week you have everything you need to decide. The discovery fee is credited toward the build engagement if we proceed. If we do not, the documents are still useful and you keep them.

This is the most underrated week of any AI project. It is also the week that decides whether the next ten weeks ship.