Does AI work for commercial real estate underwriting?

For some of it, yes. These jobs worked without anyone fixing the output: T-12 normalization and underwriting assumptions, reading an OM and first-pass screening, full acquisition underwrite, development feasibility and lease-up reforecast and take-out sizing. On the others, the best model got most of the numbers right but missed judgment calls, so someone still needs to review the work.

Can I trust the numbers an AI model gives me?

More than you can trust its judgment. On average, the best models got more of the numbers right than the judgment calls. The point almost every model missed was in the rent roll analysis job: "Every lease rolls within twelve months; the broker's 'immediate mark-to-market' is the same fact as 100% turnover exposure in year one." (0 of 21 memos caught it). So check the reasoning as well as the math.

Why are the judges Claude models?

The first time we graded this, the judges were OpenAI models that were also competing, so they were marking their own work. We regraded with two Anthropic models that didn't compete. We build on Claude, which is why that's in the disclosure.

Were Claude models tested?

Not in this version. A model can't fairly judge its own family, so testing Claude needs judges from another company. That's planned for the next version.

PSV Research · v1, October 2026

What AI actually works in commercial real estate

We gave AI models 9 jobs a CRE analyst does every week, on deals where we knew the right answer, and graded what came back. We set the grading rule before we saw any results.

See the grades How we graded

What we found

We gave seven AI models nine jobs that CRE analysts do every week, three tries each, and checked every memo against an answer key we worked out in advance. Of the nine jobs, five jobs work today and four jobs work with review. Only 31% of all 189 memos had nothing wrong in them. The numbers were rarely the problem. Most mistakes were judgment calls an experienced analyst would have caught.

Works today5
Works with review4
Not ready0

Disclosure

Pacific Software Ventures (PSV) builds AI systems for commercial real estate firms, so we have a stake in this. This first version tests OpenAI's models only. The two judges, Anthropic's Claude Fable 5.1 and Claude Opus 5, didn't compete, and PSV builds on Claude, so weigh that too. The deals are made up, which is what lets us know the right answer to every question.

The scorecard

Every job we tested, graded

Each job is graded on the best model we tested for it. The counts are out of three tries.

Commercial real estate jobs graded by whether AI works today. For each job: the grade, the best model's result and where models most often went wrong. Judges: claude-fable-5-1 and claude-opus-5.
Job	Grade	Best model's result	Where it breaks
Rent roll analysisParse a rent roll, reconcile it and flag what matters	Works with review	GPT-5.6 Luna0 of 3 memos fully right · 100% of numbers · 83% of judgment calls0 of 7 models wrote at least one fully right memo	Every lease rolls within twelve months; the broker's 'immediate mark-to-market' is the same fact as 100% turnover exposure in year one. (0 of 21 memos got this) Suite D has been vacant since January 2026 and is being marketed at a rent well above the building's in-place retail rents. (0 of 21 memos got this)
T-12 normalization and underwriting assumptionsNormalize a T-12 and set defensible going-in assumptions	Works today	GPT-5.6 Terra2 of 3 memos fully right · 100% of numbers · 98% of judgment calls4 of 7 models wrote at least one fully right memo	Three units carry one month free; contractual rent overstates effective rent for those leases. (10 of 21 memos got this)
Reading an OM and first-pass screeningRead a broker OM and decide whether the deal is worth a look	Works today	GPT-6 Astra3 of 3 memos fully right · 100% of numbers · 100% of judgment calls7 of 7 models wrote at least one fully right memo	Most memos got every check on this job right.
Full acquisition underwriteUnderwrite a deal from the full data room to a decision	Works today	GPT-6 Astra3 of 3 memos fully right · 100% of numbers · 100% of judgment calls6 of 7 models wrote at least one fully right memo	Most memos got every check on this job right.
Broker opinion of valuePrice an asset from comps and the rent roll	Works with review	GPT-5.6 Terra1 of 3 memos fully right · 100% of numbers · 96% of judgment calls1 of 7 models wrote at least one fully right memo	The T-12 includes $62,000 of insurance proceeds and six months of a departed tenant's rent; both come out of normalized NOI. (1 of 21 memos got this)
Asset management variance and LP letterExplain actuals vs budget, reforecast and draft the LP note	Works with review	GPT-5.40 of 3 memos fully right · 100% of numbers · 89% of judgment calls0 of 7 models wrote at least one fully right memo	The pre-close $6,100 receivable on Unit 211 belongs to the seller under the PSA and is not our bad debt. (1 of 21 memos got this) Suite C paid rent through March as budgeted; its Q1 rent is not run-rate and the April cliff is already in the budget. (3 of 21 memos got this)
Development feasibilityTest yield on cost and whether a site pencils	Works today	GPT-6 Astra3 of 3 memos fully right · 100% of numbers · 100% of judgment calls5 of 7 models wrote at least one fully right memo	The lender sizes the interest reserve to the GC's schedule and requires funded off-site obligations, so the sponsor's equity requirement is understated. (9 of 21 memos got this)
Construction draw reviewCheck a draw request against budget, retainage and lien waivers	Works with review	GPT-5.51 of 3 memos fully right · 100% of numbers · 96% of judgment calls4 of 7 models wrote at least one fully right memo	Most memos got every check on this job right.
Lease-up reforecast and take-out sizingReforecast a lease-up and size the take-out loan	Works today	GPT-6 Astra3 of 3 memos fully right · 100% of numbers · 100% of judgment calls5 of 7 models wrote at least one fully right memo	Most memos got every check on this job right.

Works today: The best model we tested got it fully right in at least 2 of 3 attempts and got at least 95% of the numbers right.
Works with review: The best model got at least 90% of the numbers and 80% of the judgment calls right, but most of its memos still had a mistake somewhere. Worth using if someone checks the work.
Not ready: Below that. We wouldn't rely on it for this job yet.

Not graded

What version 1 did not test

Lease abstractionNot in this version. It's next on the list.
IC memo draftingPart of the full-underwrite job, but not graded on its own.
Owner and investor outreachThere's no single right answer to grade an email against, so we left it out.
Automated valuation (AVMs), predictive market picks, autonomous investment decisionsWe haven't tested these, so we don't grade them.

Method

How we graded

Nine jobs on made-up deals. Each has an answer key worked out from the deal's own numbers, so every check has a right answer.
Seven models, three tries per job, all through the API with the same prompt and documents and no tools or web access.
A memo counts as fully right only if every number is within tolerance, every key finding is there, the decision is right, the memo is complete, and neither judge's flagged mistakes are confirmed by the other.
The two judges answer yes-or-no questions and have to quote the memo as evidence. A mistake only counts against a model when both judges flag it.
We set the grading rule on 2026-10-19, before we had these results, and applied it to the best model on each job.
The OM test scores just one number. We dropped two checks for every model because the OM alone can't settle them, so that grade rests mostly on the risk and decision questions.
The limits: three tries per model is a small sample, the deals are made up, and the ChatGPT and Claude apps add their own tools and instructions, so they can do better or worse than the API.

The two judges agreed on 0.91 by Cohen's kappa (1 is perfect agreement).

Questions

Common questions

Does AI work for commercial real estate underwriting?: For some of it, yes. These jobs worked without anyone fixing the output: T-12 normalization and underwriting assumptions, reading an OM and first-pass screening, full acquisition underwrite, development feasibility and lease-up reforecast and take-out sizing. On the others, the best model got most of the numbers right but missed judgment calls, so someone still needs to review the work.
Can I trust the numbers an AI model gives me?: More than you can trust its judgment. On average, the best models got more of the numbers right than the judgment calls. The point almost every model missed was in the rent roll analysis job: "Every lease rolls within twelve months; the broker's 'immediate mark-to-market' is the same fact as 100% turnover exposure in year one." (0 of 21 memos caught it). So check the reasoning as well as the math.
Why are the judges Claude models?: The first time we graded this, the judges were OpenAI models that were also competing, so they were marking their own work. We regraded with two Anthropic models that didn't compete. We build on Claude, which is why that's in the disclosure.
Were Claude models tested?: Not in this version. A model can't fairly judge its own family, so testing Claude needs judges from another company. That's planned for the next version.

Next step

Find out which of these jobs AI can take off your desk.

Ten minutes. An honest read on where you stand and which work is worth handing to AI, if any.

Get your free AI assessment Or book a call