Twelve mistakes AI makes in real estate models when nobody checks

A chat AI writes the answer that seems most likely. It doesn’t go back to the memo to look again, and it doesn’t trace its own formulas. When the memo doesn’t give a number, it doesn’t leave a gap. It fills in something that fits. So a wrong model and a right one look exactly the same, and asking the AI to double-check itself doesn’t help: it checks with the same eyes that made the mistake.

And a wrong model doesn’t break. It just gives a different answer, quietly, by enough to change a decision.

The debt payment is worked out from the $24 million purchase price instead of the $14.4 million loan. Or the sale is priced off this year’s income instead of next year’s. The number looks like a debt payment, so nobody clicks into it. The returns are wrong from that cell on.

Total operating expenses is a formula, so it looks trustworthy. But it adds rows 20 to 34 when the expenses run to row 35, and the last line, the replacement reserve, never makes it into the total. Net operating income is too high, and every return after it.

The seller’s net operating income is $1,250,000 and the model comes to $1,238,000. Instead of finding the gap, the AI adds “+12,000” to the end of the formula itself, and now they match. Then it tells you the model ties out to the memo. It doesn’t: the $12,000 gap is still there, now hidden by the AI that made it.

The cash flow to the investors is worked out from net operating income, but the formula never takes out the debt payments. Every month shows far more cash than the deal produces, and the levered returns look excellent.

Tenants repay $380,000 of operating costs. The model adds the $380,000 to income and still leaves those costs out of expenses, or takes the replacement reserve out twice: once as a cost and again below the income line.

A $45,000 rent concession is added to income instead of taken off. Income is now $90,000 too high, and the model still balances.

Month 14’s utilities are typed in as $35,000 where the formula belongs. It looks like every other month. But change the expense growth, or the start date, and that month doesn’t move with the rest.

You raise the vacancy from 5% to 8% to see what happens, and the returns don’t budge. The formula that should read the vacancy has its own 5% typed inside it, so the box you changed isn’t connected to anything.

A $1,800 elevator inspection, billed once a year, goes in as $1,800 a month. That’s $19,800 a year of cost that doesn’t exist, in a building with millions of costs, so nothing looks out of place.

The memo shows rents of $32 a foot “at market.” Tenants pay $26 today. The model starts at $32, so year one counts rent nobody is paying yet, and every return after it is too high.

The memo has no tax figure, so the model shows one anyway: a round number with no source, sitting in the inputs as if the seller gave it. Nothing tells you it was assumed.

Three suites appear only on the stacking plan, not on the rent roll. The model never sees them. Its totals still add up, because they add up what it read.

You can catch most of these yourself if you check the model against the documents and trace its formulas, not by asking the AI. The checks are in How to check an AI-built real estate model. Just remember it means going through thousands of formulas, on every deal.

Or let us do it. Our checks are run by our own proprietary software, built to catch mistakes like these before a model reaches you. Anything your documents don’t say is asked about, never guessed (how we check).