Most AI underwriting mistakes trace back to one decision the model made without you: it guessed. Hand an offering memorandum to Claude in Excel with no other context and it will build a clean five-year proforma in a few minutes. It will also fill in every assumption you never gave it, and on a real deal some of those assumptions run in the seller’s favor.
In a recent community workshop I underwrote the same eight-unit Los Angeles apartment deal twice, once cold and once with the firm’s buy box loaded in. The distance between the two outputs is the entire lesson in AI underwriting: the reliability problem is a context problem. Give the model the criteria you carry in your head before it touches your model, and the guessing mostly stops.
What the model invents when you stay quiet
The cold run looked professional and was wrong in specific places. It defaulted the exit cap rate to 5.75%, generous for an older building in a softer submarket where 6.25% is closer to honest. It set disposition costs at 2%, when 5% is the working number. It put insurance at roughly $1,000 a unit, which on an old LA building is not even close; $2,000 a unit is the reality. It budgeted nothing for interior renovations. And it accepted an 86.3% loan-to-value lifted straight from the broker’s pitch. At the asking price the model returned a 27.1% levered IRR. Re-run at a realistic 65% LTV, the same deal came back around 17%. All of that came from the same source: the model had no context to work from. As I told the room, the more assumptions the AI makes, the more hallucinations and problems you inherit.
The buy box is the context it is missing
A buy box is your acquisition criteria written down: the return thresholds you require, the asset criteria you buy (unit count, vintage, unit mix), your target submarkets, the regulatory flags that kill a deal, and the standard assumptions you apply to everything. It is the stuff a good analyst already knows and a fresh model has no way to know. Assembling it honestly takes a few hours (if not more), and that investment is the whole move. We effectively took what lives in our heads and put it on paper so the AI underwrites the way we do.
Build it once, as a skill
The durable version of this is a skill, which is just a reusable set of instructions the AI loads every time. Ours has three parts: an instruction file that defines the steps, an assets folder holding the actual underwriting template, and a reference folder holding the buy box plus the standard template inputs. The model reads all three on every run, so it stops reinventing your format and your assumptions deal after deal.
From offering memorandum to go or no-go
With the skill in place, attaching an OM triggers a fixed sequence. The model extracts the deal facts from the memorandum, screens them against your criteria, populates your exact underwriting model, and returns a verdict: go or no-go, negotiate or walk. The cold proforma earlier had no renovation budget and a fantasy LTV because nothing told it otherwise. The buy box version starts from your numbers, so the output lands far closer to something you would actually send.
What reliable context unlocks
Once the model works from your criteria, you can ask it to do things a fixed Excel template was never built for. In the same session we added a scenarios tab with base, upside, downside, and stress cases on the existing model, then ran a target-price solver: given a required 18% levered IRR, what purchase price gets you there. That kind of analysis usually means an Excel consultant or hours of manual rework. With the context in place, it becomes a prompt, and the buy box is what makes those answers trustworthy instead of confident guesses.
Review every number anyway
A buy box raises the floor; it does not remove the human. Even with full context, an AI-built model is a first draft. One model in the workshop carried a cash-flow error you could not catch from the summary alone, the kind of thing that looks fine until you trace it. Sending a model like that out without a thorough review is a disaster waiting to happen. Treat the output as a fast, useful draft and nothing more. Two honest limits apply. AI will not conjure clean comps in non-disclosure states, so pair it with comp-focused data tools rather than trusting whatever it scrapes. And if the figures are sensitive, you can have the model build the skeleton and structure without handing it your most confidential inputs.
One setting and one habit worth copying
Two small things from the workshop. In Claude’s settings you can opt out of having your data used for training, which matters once your buy box and models live inside it. And iterate on your skills with voice dictation rather than typing; describing the change out loud is faster and tends to produce a sharper instruction file.
The two runs of that eight-unit deal used the same model, the same offering memorandum, and the same software. The only variable was whether the AI had the firm’s criteria in front of it. The few hours it takes to write the buy box down is what stood between a 27.1% IRR built on assumptions nobody would sign and a screen the firm could trust.
We break down workflows like this inside the AI for CRE Collective most weeks, and walk through them in the newsletter. If you want the buy box structure to adapt for your own deals, that is where to find it.
