How to Run a First-Pass CRE Underwrite With AI (Without the Made-Up Numbers)

Blackstone and Brookfield are pouring money into custom AI. Most of the coverage filed that under bad news for proptech. Read it the other way and it tells you exactly where your own money should go.

The giants aren’t trying to out-train Anthropic. Nobody is, and nobody should. They’re treating the model as a rented commodity and spending on the part that compounds: how it plugs into their deal process, their data, the way they price an asset. That’s the lesson for everyone who isn’t Blackstone. You won’t own a better model than the shop down the street, and you don’t have to. You can own the process. Rent the brain. Own the workflow.

So let’s own one workflow, start to finish: the first-pass underwrite.

The most thankless hour in acquisitions

A fresh OM, a rent roll, a T-12, and twenty to thirty minutes of pure data entry before anyone has an opinion worth having. At ten or fifteen deals a week, that’s most of a working day lost to typing instead of thinking. Here’s how you hand off the typing and keep the thinking.

Step 1: Ingestion

Feed it the rent roll and the T-12 and tell it to pull the line items into a clean structure. Nothing else. No interpreting yet. This is the part going commodity fastest, so treat it as plumbing, not edge.

Step 2: Rebuild the honest NOI

Have it rebuild in-place income the way you would if you had the hour: drop the model and employee units, strip the concessions, reset the tax line to your basis instead of the seller’s, normalize the expenses. Now the headline number and the real one sit side by side, and the gap is the conversation.

Step 3: A first-pass return on your assumptions

Run a first-pass return off your assumptions, not its. Cap rate, cash-on-cash, and a flag everywhere the two NOIs disagree. The model proposes; your assumptions decide.

Step 4: The rule that makes it trustworthy

Here’s the part people skip, the one that turns a handy tool into a real liability. A model will fill an empty cell with a confident, plausible number every single time you let it. In an underwrite, a made-up rent is worse than a blank one, because it looks finished. So you hand it one hard rule: use only what’s in these documents, and where a number is missing, write a CONFIRM tag that names exactly what to go check. Never a guess.

What you get back

A first-pass underwrite in minutes, every figure either sourced or flagged for diligence, and the call, pursue or pass, still entirely yours. The machine typed and added. You decided. That’s the trade you want, and it’s exactly what an AI underwriting copilot should do for an acquisitions team: speed without the made-up numbers.

The ingestion grunt work is going commodity, so your edge moves up to judgment and to whose process the model runs inside. For more on that, read this week’s AutomateRE or compare the field in the best AI tools for CRE underwriting. Want to try the building blocks first? Start with the free Claude skills for real estate pack.