Using AI in commercial real estate underwriting: what it can and can't do

Where does AI help in underwriting?

In the hours before the model: getting messy inputs into a consistent shape, and in the hours after: explaining what the model says. The four places it pays for itself most reliably:

• Normalizing the T-12. Mapping the seller's chart of accounts to your own line items, separating one-time items, and flagging months that do not look like the others.
• Reading the rent roll. Turning a PDF or an awkward export into one row per suite, with lease dates, rent and recoveries, and listing where the rent roll and the leases disagree. The detail is in AI rent roll and lease abstraction.
• Pulling comps. Querying sources you are allowed to automate, such as county sales files, and screening out nominal deeds and portfolio sales before they reach the model.
• Narrating the output. Writing the plain-language paragraph under a sensitivity table, and drafting the investment memo from the finished model.

How does AI normalize a T-12 and a rent roll?

By applying a mapping you define, consistently, and telling you where the mapping did not fit. Every seller labels expenses differently. Give the model your standard line items and the rules: repairs versus capital items, what counts as a management fee, how to treat a one-time insurance claim. Ask it to output the mapped statement and a list of every line it was unsure about.

What to check before using the result:

• Mapped totals equal the seller's totals for revenue, expenses and NOI.
• Every excluded or reclassified line is listed with its amount and reason.
• Annualized figures say which months they were annualized from.
• Rental income on the T-12 is within a stated tolerance of the rent roll's annual rent, and the gap is explained.

A starting instruction:

Map every line of t12.xlsx to the line items in chart.md, using the rules
in rules.md. Write the mapped statement to a new tab called Inputs, with a
source column giving the original line name and month. Do not change any
existing tab. List every line you reclassified, excluded or were unsure
about, with its amount and reason. Confirm that mapped revenue, expenses
and NOI equal the seller's totals, using formulas in the sheet.

What must AI not decide in underwriting?

Anything that is a judgment about the future or about value. Specifically:

• The cap rate. Going-in and exit cap rates are the most sensitive assumptions in the model, and they rest on a read of the market, the asset and the buyer pool. A model can summarize the evidence. It should not pick the number.
• The assumptions. Rent growth, vacancy, downtime, leasing costs, capital reserves and hold period are underwriting decisions with a name attached.
• The value conclusion and the bid. What the property is worth to you, and what you will offer, belong to the person who will defend it at investment committee.
• Pass or proceed. An AI summary that says a deal "looks attractive" is a sentence, not an underwriting.

A useful test: if the number changes the answer and a reasonable analyst could pick a different one, a person picks it.

What is model risk when AI touches underwriting?

It is the chance that a wrong number enters the model and looks right. AI introduces a few specific ways that happens:

• Invented figures. A model asked for a value it cannot find may produce a plausible one. Require "not stated" instead of a guess, and require a source for every number.
• Silent formula changes. An assistant editing a spreadsheet can overwrite a formula with a hard-coded value. Protect formula cells, and diff the workbook before and after.
• Arithmetic in prose. Numbers computed inside a chat reply, rather than in the spreadsheet, are not reliable enough to underwrite on. Let Excel do the math.
• Stale inputs. A comp set or market figure without a date can be months old. Every external input should print its as-of date.
• Confident narrative. A well-written memo makes weak numbers read as strong. Review the numbers before the prose.

How do you keep an audit trail?

Make every input traceable to a file and a page, and every change traceable to a person. In practice:

• Keep source documents in the deal folder, unchanged, and name them consistently.
• Have AI-produced inputs land on their own tab, with a source column citing file and page, rather than typed straight into the model.
• Save the prompt or Skill that produced each input alongside it.
• Keep a change log in the workbook: date, who, what changed, why.
• Save a version of the model at each decision point: first look, LOI, investment committee.

A reusable Skill is the natural place to write the house rules once, so every analyst's T-12 mapping follows the same logic.

Why should Excel stay the system of record?

Because the model is what your team, your lenders and your investors already audit. A spreadsheet shows every formula and every input; a chat transcript does not. Use AI to prepare inputs and explain outputs, and let the workbook hold the math.

Claude Cowork, which Anthropic began folding into the main Claude app on September 16, 2026, starting with Pro and Max, fits this well: it works on the workbook in an approved folder on your computer, so the file stays in Excel instead of being uploaded, rebuilt in a chat and downloaded again.