A revenue chief has 400 closed deals, three competing explanations and a CRM she does not trust. This is the two-page brief an AI agent wrote for her board after reading every call. Each claim in it opens the buyer quotes and the deals behind it.
A fictional sample client. Lanternmere HR, its people, its competitors and all 400 deals are invented: synthetic deals with planted patterns, written for this project. The model calls that read them are real. Recovering planted patterns tests the method. It does not prove the agent would be as accurate on real deals.
How to read it: text marked like this is a claim. Click one to see its evidence.
The sample client’s starting point
Our win rate has gone from 28% to 19% in four quarters. In last week’s exec meeting I heard three explanations and no evidence.
Sales says we are too expensive. Product says we are missing features. Marketing says the broker channel sends us weak leads.
The CRM loss reason says “Price” on 61% of losses. I don’t believe it, because it is the first option in the dropdown.
We record every call. The answers are in the transcripts, and nobody has time to read 400 deals. I need a brief for the board meeting: what is causing the losses, by segment, with quotes I can read aloud, and what each team should do about it.
Method note
The deals were generated in one step with planted patterns and an answer key. The analysis ran in a separate step and never saw the key. Afterwards the brief was scored against it. These are all seven checks, with counts.
The answer key, the generation prompt and every writing brief are published in the repository, so a reader can see the test was fair.
Evidence appendix
Every deal the agent read, with what the CRM says and what the agent coded. Open one to read the call.