►EXECUTIVE SUMMARY 80% ACCURACY IS A FEATURE FOR PICKING RESTAURANTS, NOT FOR BETTING $1M+ AT 90% LEVERAGE: LLMs get it right four times out of five. On a leveraged acquisition, that fifth time takes your house. You need 99.9%, and ChatGPT isn't within a mile of it. STACKED 80%s ARE STACKED 20% ERRORS — AND THAT MATH COMPOUNDS AGAINST YOU: Every layer of AI analysis multiplies the wrong numbers underneath it. By the time you've built a thesis on it, you're not buying a business, you're buying someone else's rounding errors. GARBAGE IN, GOLD OUT ISN'T A THING: Most small business financials are a mess. Feeding that into an algorithmic tool doesn't clean it up — it launders bad data into confident-looking output. That's worse than no analysis at all. THE FAILURE MODE NOBODY TALKS ABOUT: RECONCILIATION: A real data room has 2-3 sources of truth for every number that matters. When two datasets don't reconcile, the LLM quietly picks one — often the wrong one — or averages them. It won't flag it. That silent choice is what screws you. MAYBE IN A YEAR OR TWO. NOT TODAY: Use Claude and ChatGPT to organize data. Do not use them as your diligence. I want AI to do this job — it would save me enormous time. It can't yet.