A large AI decision is a bet about a future nobody has seen, so every pilot result has to be read through a belief about that future. When the people in the room hold different beliefs and nobody has said them out loud, the same evidence means different things to each of them, and the committee freezes. Lay the beliefs down early, ask which ones the room rejects, agree what would prove the survivors wrong, and set a date to look.
The size of the problem
Between 40 and 60 percent of qualified deals ended in no decision at all. The buyer agreed that change was needed and still could not commit, out of fear of making a mistake they would be blamed for.
On The JOLT Effect, by Matthew Dixon and Ted McKenna. Page 80 in the reader
How a good deal stalls
In the book's example a strong Challenger seller brings a real insight to an insurer, runs a clean MEDDIC, and wins the pilot.
Then the deal reaches the committee. Legal says each new AI model takes months to approve, and asks what happens when a better one appears next quarter. The chief financial officer asks why the company should commit for three years to something that will change in one. Nobody had raised either question earlier, because nobody had asked what the room believed about the future. The deal does not go to a competitor. It goes to no decision, the outcome the JOLT research found in 40 to 60 percent of qualified deals.
The disagreement lives in the premises
Two careful executives can reason soundly from different starting points and arrive at opposite vendors. A chief financial officer starts from cost; a chief technology officer starts from never being locked in again.
They can argue for a month and neither will move, because they are arguing about conclusions. The disagreement does not live there. It lives in the premises, and nobody has said the premises out loud.
Evidence does not decide on its own
Picture two insurers that run the same two-week trial and get the same accuracy. One signs, because its committee agreed that its definitions will outlive any AI model, and the trial showed the definitions work. The other stalls, because its chief financial officer believes the market will settle in two years and sees no reason to commit before then. The evidence was identical. The beliefs decided.
Most of the stalled AI deals I have seen did not stall for lack of evidence. They stalled because nobody had agreed, beforehand, what evidence would be enough. That is the job the thesis does, and it is why it comes first.
What to do about it
In AI much indecision comes from a thesis nobody agreed on: the buyer is afraid of being wrong about the future. So the recommendation should come from the agreed thesis, and the thesis also limits the exploration, because it narrows the field to the products that fit it. The risk can be taken off the table without a reference: name what would prove the thesis wrong, and set an early checkpoint where you will both look.
The customer nods at everything. A buyer who agrees with all of your theses is not agreeing. They are being polite, or they have not engaged. Ask the question that breaks the politeness: "If you had to cross one of these out, which would it be?"
How you would know the book is wrong
It is wrong if teams that run thesis sessions, the working meetings where a team agrees its theses, and keep a ledger of how each thesis held up, see no change in no-decision rates or cycle length against their own baseline.
The framework's own falsifier, from the Introduction. Page 27 in the reader
Write the bet and its stop conditions down with the thesis memo builder.