I Tried to Place a Deal Using AI. Here's What Happened.

I have a 73-page transcript of me arguing with Gemini about a multi-million dollar ground-up special use build.  I did this to PROVE A POINT. Here's the highlight reel: For the entire conversation, it presented Commercial Loan Direct as a direct lender — draw schedules, prepay structuring, exact LTC math. CLD is a broker marketplace ("let lenders compete for you"). I caught it reading their homepage: "Are you feeding me a line?" It admitted it "repeatedly mischaracterized" them — then two responses later quoted their mezzanine rates like a lender again. "CLD is not a lender, why are you saying it is?" "I am sorry, I made that up." Same conversation but later: Wilshire "prefers to fund land acquisition" → "flat-out wrong." Kennedy Funding recommended for construction → Um, "Kennedy Funding does not do construction financing."  Ai: "you should try Ready Capital parameters for this multimillion dollar deal → Me "their minimum is $50M." The AI: "I pulled those numbers straight out of thin air." Every correction came from me — reading the homepage, knowing the lender, checking the program page to see how it could have created fake info. The AI never caught one of its own errors in 73 pages. The person who can spot the lies is the person who didn't need the AI in the first place. This isn't a prompt problem. Lender matrices and real coverage data aren't on the public internet, so the model invents — confidently, repeatedly, and it can't even remember being corrected. Full transcript walkthrough and what deterministic matching on real lender data looks like instead: Full article: https://www.brokercheetah.com/blog/ai-lied-to-you/ Short version: when anyone quotes you a lender, ask "says who?" The lender's own guidelines — real. "The AI said so" — after 73 pages, I know exactly what that's worth.