Can AI Underwrite Real Estate Deals?.

Can AI underwrite real estate deals? It can draft one and it cannot decide one, and the difference between those two verbs is the whole answer. AI will extract a rent roll, rank a hundred deals and produce a first pass memo faster than any analyst alive. What it will not do is own the exit cap, sign the final number or say the no, because each of those is a position with a person's name attached and a machine has no name to attach. So the honest answer is yes to the drafting and no to the deciding, and a shop that keeps that line clean gets the speed without the exposure.

Underwriting is a decision, not a document

Most of the confusion in this question comes from treating underwriting as the spreadsheet. It is not. The spreadsheet is the artifact. The underwriting is the set of judgments the spreadsheet is arguing for: what rent actually gets collected, how fast it grows, what a buyer five years out will pay for that income, how much leverage is prudent against it. Software can compute every one of those lines. It cannot be accountable for a single one. Confusing the computation with the judgment is exactly how a shop ends up holding a model no one in the building can defend, which is the situation investors are in when they call us to audit one.

Keep that distinction in hand and the rest of the answer falls out cleanly, because you can sort any AI capability into one of two piles: the drafting it does well, and the deciding it must never do.

What AI does well: the drafting half

Four workflows carry almost all of the real value, and we run each of them inside our own shop before we install them in anyone else's.

Document intake first, because it is the highest value and the lowest risk. Getting a broker's rent roll, trailing twelve and leases out of a PDF and into a clean grid, reconciled against the source totals before anything downstream trusts it, saves a day of retyping per deal and pays for itself in week one. The reason it is safe is built into it: if the extraction is wrong the grid fails to tie to the T-12, and the error surfaces on a Tuesday instead of at a committee.

Screening second. Point a model at a hundred teasers, rank them against a box you wrote down in advance, market and size and vintage and return threshold, and spend the week on the top three instead of finding out which three. The discipline that makes this work is writing the box first. A machine ranking deals against your written criteria is a screener. A machine ranking deals against its own taste is a liability with a confident tone.

First drafts third. Memos, abstracts and deal summaries produced in minutes and clearly marked as drafts. A draft is a starting point for a person, never a substitute for one.

Arithmetic checking fourth. Reconcile a model's outputs in a second engine and you catch the broken formula the first engine hid. That is literally how our audit practice works, and every deliverable we ship is verified against an independent replica before release.

Notice what those four have in common. In every one, if the output is wrong the error surfaces before capital moves. That is the entire test for whether a task is safe to automate, and all four pass it.

What AI cannot do: the deciding half

The other pile fails the same test, and the failures are expensive.

The assumptions. Exit cap, rent growth, vacancy, the rate path. A machine can hand you a sensitivity grid with the exit cap running from six to seven. It cannot tell you which column you are willing to defend to an investor in year five, and that one column moves more money than almost anything else in the file. In our exit cap essay a hundred basis points of expansion erases three percent rent growth compounded over five years. The grid is arithmetic. The choice of which cell you will live in is a position, and a position needs an owner with something at stake.

The final number. Whatever reaches an investor, a lender or a committee gets reconciled against the source once, by a person whose name is on it. One reconciliation, every time, no exception for being busy. That single habit is most of what separates shops that scale AI from shops that scale their mistakes.

The no. Passing on a deal is the most expensive decision in the building measured by regret and the cheapest measured by cash. A ranking can suggest a pass. A person kills the deal, because only a person can be asked why two years later.

The blank cell. When a rent roll has a blank, the blank is information about the seller's records. Machines fill blanks helpfully, and helpful is the failure mode. The blank stays blank until the source document answers it.

We keep the full version of this list, the things the machine is never allowed to touch, in a companion piece written for operators. The short version lives here because the question this essay answers is the simpler one: can it, not how do you govern it.

So can it, or not

Put the two halves together and the grammar answers the question. AI can underwrite in the sense that a junior analyst drafts, quickly and tirelessly and usefully. It cannot underwrite in the sense that a principal signs, because signing is the part that carries the risk and a model carries none. A deal is underwritten when a person with capital at stake has chosen the assumptions, reconciled the number and accepted that the no was or was not the right call. Everything upstream of that is drafting, however good the draft.

That is not a hedge against the technology. It is the reason the technology is usable at all. The speed is real precisely because the accountability stays put.

Where a small shop should actually start

If you run a lean shop and you want the speed, do not start with the deciding. Start with the drafting, and start with the safest draft there is: document intake. Getting the rent roll and the T-12 into your model without a day of retyping is unglamorous, low risk and visible in week one, and it fails loudly when it fails, which is what you want from a first project. Screening is the natural second. The assumptions, the final number and the no stay exactly where they were, with you.

That sequence, draft first and decide last, is what we install when we implement AI inside a client shop, and it is the whole reason the speed never costs you a deal. It is also how we built our own models and the free screener on this site, every one of them verified against an independent replica before release.

Can AI underwrite real estate deals? It drafts them. You underwrite them. AI drafts. Humans decide. Deals close. The order is not decoration. Ask the machine to skip a step and you will learn, usually at the worst possible moment, why the order is the answer.

The model decides the deal. A person decides the model.