Say yes to the pilot, then ask two things before it starts: what would it have to show for the buyer to take the next step, and what would make them stop? Write both down in the buyer's words, inside a short thesis memo: the bet, why the room believes it, when it succeeds, when it fails, what the company keeps if it fails, and when everyone looks again.
Say yes, then ask one question in two halves
So when a buyer says, let's just pilot it, say yes. Then ask one question in two halves: what would the pilot have to show for you to take the next step, and what would make you stop? The answer to the first half is the buyer's thesis in disguise. The answer to the second is its falsifier. Write both down before the pilot starts, in the buyer's words, and the pilot stops being a demonstration and becomes the first test of the bet.
Why pilots stall
Here is the trouble with a pilot on its own. A pilot answers the question it is given, and most are given the wrong one.
In my experience the cause is rarely that the technology failed the test. It is that nobody agreed beforehand what the test was for. The pilot shows the tool can summarize a claim file. Nobody decided whether summarizing claim files was the bet the company was placing, so the result proves something nobody needed proved, and the decision stalls exactly where it was.
The thesis memo
What carries it is the thesis, written as a buying document. On a seller's slide, a thesis is part of the pitch. In the committee's file it becomes something else: a record of what bet the company is placing, why, and under what conditions the bet succeeds or fails. I call that document a thesis memo.
The bet. We will keep our definitions of a claim, a policy and a loss outside any one AI model, and buy only tools that can read them, so that every better AI model can use them without a rebuild.
Why we believe it. Three theses, each argued over in this room. Our definitions will outlive any tool we buy. Better AI models will keep arriving, so anything built around one AI model will be rebuilt. Approval in a regulated company moves more slowly than the AI models do.
It succeeds if. In the first two weeks, on the claims questions that already have a known right answer, the answers reach a level the claims team sets before the trial starts. By the end of the first quarter, a second approved AI model can run on the same definitions without rewriting them.
It fails if. The definitions have to be rewritten for each AI model, or the claims team cannot tell a right answer from a wrong one, or legal begins approving new AI models in weeks rather than months, which would undo the third thesis and change the plan.
What we keep if it fails. The definitions stay ours, in an open format. We lose two weeks, not three years.
When we look. Week two, the end of the first quarter, and ninety days before renewal, each time against this page.
What stays a checklist. Security controls, where the data is kept, audit and contract terms. These are stable, so they are scored the usual way.
The insurer’s memo in Chapter 9, invented like every story in the book. Page 171 in the reader
The memo is the plan for the pilot
Look again at Priya's memo. Its "it succeeds if" line is a pilot: two weeks, on questions with a known right answer, against a level the claims team sets before the trial starts. Its "it fails if" line is the stop condition. Its "what we keep if it fails" line is the contract safeguard, written in the buyer's terms. Its "when we look" line is the staged commitment. Every tool the careful buyer asked for is in the memo. What the thesis added is the reason each one is there: the pilot tests the belief the deal rests on, not whatever the product happens to do well.
There is one honest warning. You cannot write this memo for the buyer and hand it over, or it becomes your pitch on their letterhead. Offer the structure, argue over the theses, draft lines when asked, and let the champion and the committee own the words.
Build the sale from the bottom
The memo is the written form of the stack: thesis, implication, decision, and only then a product.
Start at the bottom and the order of trust flips. The thesis costs the buyer nothing to consider, because it is about the world and would be true with any vendor. By the time a product appears, it is sitting on three layers the buyer has already accepted.
Agreement is not approved spending
So treat it as two problems, and work both. The first is intellectual: does the room agree on what is true? The thesis is built for that one. The second is organizational: who owns the budget, when it opens, what the paper process requires, who can block the deal, and whether the company has the people to implement it.