We built an AI that attacks business plans. It gave ours a 14 out of 100.

What an adversarial AI finds when it reads a plan for failure instead of applause — and the map it reads from.

Every AI you've used is built to keep you going. Ask it about a shaky idea and it helps you polish the shaky idea. We wanted the opposite: a reader that assumes the plan is wrong and goes looking for where. So we built one — a Panel of hostile specialists that attacks a business plan and names the single thing to fix first.

The first plan we ran it on was our own. It scored us 14 out of 100. The verdict, verbatim: "one person selling a prompt pattern as a moat, priced against a free substitute that the plan's own evidence documents." We left it on our own sample report, unedited, because a tool that only works on other people's plans isn't a tool.

The number isn't the part worth your time, though. The map behind it is.

To attack a plan the same way twice, you need a vocabulary for how plans die. We coded one: 53 distinct failure modes, grouped into 15 families, each owned by the specialist trained to catch it — the hostile investor, the operator, the regulator, the actuary who prices one-way doors.

Count the modes per family and a pattern falls out that surprised us: plans don't usually die of a bad idea. They die in execution and unit economics — the two families with the most distinct ways to go wrong. The idea is rarely the problem. The schedule that assumes nothing goes wrong, the single founder who is four jobs at once, the CAC payback longer than the runway tolerates — that is what kills plans.

A sample of what the Panel actually reads for:

• Market — a top-down TAM with no bottom-up proof; a "why now" that is asserted, never shown.
• Moat — weekend-clonable, with no compounding asset; a data moat claimed but never designed to collect anything.
• Competition — the "no competitors" claim; the real competitor that is a spreadsheet, DIY, or doing nothing.
• Reputation — no answer to "what's the headline when this goes wrong."
• Irreversibility — a one-way door, left unpriced.

The full 53, with definitions, are here: https://www.badmars.com/reports/how-business-plans-fail

Fifty-three problems is paralysis, not progress — so the Panel ranks, and names one the Frog: the problem that, until it moves, makes the other fifty-two unfixable. Ours was the one founders hate to hear — one person was the entire build, design, legal-liaison and PR function, four jobs at once. Everything else on the list routed through that single constraint.

The uncomfortable thesis under all of it: your plan is going to be tested either way. The only question is where. Here — cheap, private, reversible — or by the market, which charges more and never sends a report.

If you want to read the full failure map, it is here: https://www.badmars.com/reports/how-business-plans-fail. And if you want to watch it attack a plan you are unsure about, the first one is free, no card: https://www.badmars.com.

Fail here first, before reality charges more.