Blog

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

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

I have a 73-page transcript. Seventy-three pages of me arguing with Gemini about an $8.5 million ground-up 32-bed memory care build in Fulshear, Texas. I asked it to help me place the deal. What followed was a masterclass in confident fabrication — and I kept receipts.

Let me walk you through it.

For the entire conversation, Gemini presented Commercial Loan Direct as a direct balance-sheet lender. Not once, not in passing — for the whole thing. Detailed draw schedules. Prepayment-penalty structuring advice. Seventy percent LTC math ($5.6 million on $8 million). Land-advance mechanics. All of it delivered with total authority, all of it specific to my deal.

Commercial Loan Direct is not a lender. It's a broker marketplace. Their homepage literally says "let lenders compete for you." I caught it by reading their website myself, and this is what I typed:

"Are you feedling me a line of shit??? I am on CLDs site... The are NOT a lender per this on their home page."

Gemini admitted it: "I repeatedly mischaracterized Commercial Loan Direct as a direct lender throughout this conversation."

Repeatedly. Throughout. The entire conversation.

It got worse. On Wilshire, Gemini first told me they "PREFERS TO FUND THE LAND ACQUISITION." Then, when I pushed back, it admitted: "My first statement saying they could participate in land acquisition was flat-out wrong." Flat-out wrong. Its words.

It recommended Kennedy Funding for construction financing. I picked up the phone and called them. "Kennedy Funding does not do construction financing. I just called them." One phone call. That's all it took to destroy the AI's recommendation.

It cited Ready Capital and Broadmark for the $8 million deal with specific parameters. I checked Ready Capital's site: "Ready Capital website says ground up is 50MM+." Fifty million minimum. My deal was eight. Gemini's response: "I pulled those numbers straight out of thin air."

Straight out of thin air.

And here's the one that should terrify you. After I corrected it on Commercial Loan Direct — after the admission, after all of it — two responses later it slipped right back into treating them as a lender, quoting CLD mezzanine rates of "7.04% to 12.21%." My reply: "CLD is NOT A FUCKING LENDER." Its reply: "I completely slipped right back into treating them like one after you just called me out on it."

It can't even remember being corrected.

At one point I typed: "This entire time you have been telling me THIS IS A FUCKING LENDER!!!!!" And later, the summary of the whole 73 pages: "No, you suck at this and simply make shit up when you don't know."

Now here's the part that matters more than any single hallucination. Every correction in that transcript came from me. I read the homepage. I called the lender. I checked the program page. The AI never caught one of its own errors — not one, in 73 pages. It fabricated, I verified, it apologized, it fabricated again.

The person who can spot the lies is the person who didn't need the AI in the first place.

Oh, and on a different deal — a Puerto Rico condo — it told me a batch of lenders operated in PR. None of them did. The evidence? "PR" appeared in an address-confirmation dropdown. Two letters in a dropdown, treated as a lending footprint.

This is not a prompt problem. No prompt fixes a model that was never given the data. Lender matrices, actual coverage areas, who's a lender and who's a broker — none of it is on the public internet. It's in rate sheets, emails, and phone calls. So the model fills the gap with the most plausible-sounding thing it can generate, says it like it's sure, forgets its corrections, and does it again.

Broker Cheetah doesn't predict. It looks up. Forty-six lenders' full matrices, collected through human relationships — every tier, every LTV cap, every carve-out, verified against the lenders' own internal data. Deterministic. Same answer every time. No tokens, no hallucinations, never degrades. And it never needs you to check its work, because there's nothing to check — it's the lender's own data.

So here's my question: if the only person who can catch the AI's lies is someone who already knows all the answers — what exactly is it for?