What AI should never touch in your underwriting.

Every month another shop announces it is now AI powered, and every month another investor quietly asks us to audit a model no human in the building can explain. These are the same story at different stages. The technology is not the problem. The missing line between drafting and deciding is.

We run proprietary AI inside our own underwriting and we implement AI inside client shops, which is exactly why we keep a short list of things the machine is never allowed to touch. Not because it cannot produce an answer. Because it will, confidently, and confidence is not the same thing as accountability.

The machine drafts these

Document intake: rent rolls, trailing twelves and leases extracted into clean grids, reconciled against source totals before anything downstream trusts them. Screening: a hundred deals ranked against your stated criteria so the week is spent on the top three. First drafts: memos, abstracts and summaries produced in minutes and marked as drafts. Arithmetic checking: recalculating a model's outputs in a second engine, which is how our own audit practice works. Every deliverable is verified against an independent replica before release.

The machine never decides these

**The assumptions.** Exit cap, rent growth, vacancy, rate path. These are not facts to extract, they are positions to take, and a position needs an owner with something at stake. A model can show you the sensitivity grid. It cannot decide which cell you are willing to live in.

**The final number.** Whatever reaches an investor, a lender or a committee gets reconciled against the source by a person whose name is on it. One reconciliation, every time, no exceptions for being busy. This single habit separates shops that scale AI from shops that scale embarrassment.

**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 a person can be asked why in two years.

**Anything a blank cell should answer.** When a rent roll has a blank, the blank is information about the seller's records. Machines fill blanks helpfully. Helpful is the failure mode. Blanks stay blank until the source document answers.

The test for everything else

One question sorts every new AI use case: if this output is wrong, does the error surface before capital moves? If yes, automate it and enjoy the speed. If no, the machine drafts and a human decides. That is the whole framework, and it fits on an index card, which is the point. Frameworks that need a slide deck do not survive a busy Tuesday.

AI drafts. Humans decide. Deals close. The order matters.

For informational and educational purposes only. Not investment, legal or tax advice.