AI in Commercial Real Estate Appraisal: The Workfile, Not the Signature

Most of the hours in a commercial appraisal are document hours: reading leases and rent rolls, normalizing operating statements, assembling comparable records from filings and data services, and typing narrative that restates evidence the appraiser has already weighed. AI systems are now genuinely good at that layer. They can abstract a lease file into a grid with a reference beside each material field, pull the reported facts of a comparable sale into a standard format, and produce first drafts of the market, neighborhood, and improvements sections from approved source material.

What the systems do not produce is value. A language model can propose an adjustment and defend it fluently, but fluency is not evidence, and a model has no license, no inspection notes, and no accountability to a state board. The workable mental model is narrow and useful: AI compresses workfile production. The appraisal itself remains a professional opinion that a person develops and stands behind.

This is not an open regulatory question. For federally related transactions, the Interagency Appraisal and Evaluation Guidelines require an appraisal performed by a state-licensed or certified appraiser in conformance with USPAP, and they are explicit that an automated valuation model is not an appraisal. An AVM can support an evaluation where an evaluation is permitted, but it does not satisfy a requirement for an appraisal, no matter how much data sits behind it.

USPAP points the same direction from inside the profession: the signing appraiser is responsible for the analyses, opinions, and conclusions in the report, whatever tools helped produce the drafts. So the operating question for a valuation shop is not whether AI is allowed near the workfile. It is whether the appraiser who signs can trace every material number and sentence in the report back to evidence they have personally weighed.

The inputs are the file the assignment already has: rent roll, operating statements, leases, the subject facts, the comparable records, and the firm’s report template. The AI produces three things, each cited to its source: an abstraction of the subject documents, a normalized comparable grid in which every reported fact carries a reference to its record, and first drafts of the descriptive sections. The income approach arithmetic runs in a deterministic model, a spreadsheet or equivalent, never inside the language model.

The reviewer is the appraiser, and the approval gates are the parts of the assignment that were always the job: highest and best use, adjustment selection and support, capitalization rate reasoning, reconciliation, and the final opinion of value. Conflicts get surfaced instead of smoothed: when two records disagree on square footage or a sale term, the system flags the conflict for the appraiser rather than silently choosing one.

Inspection, market participant interviews, the reasoning behind every adjustment, reconciliation across approaches, and the signature. Those are not production tasks that better tooling will eventually absorb; they are the substance of the license. An appraiser who lets AI take the typing while keeping those five is faster without being different. An appraiser who lets a model reason about value has handed the license to a system that cannot hold it.

The honest cautions run both ways. A model-generated comparable “fact” that cannot be traced to a record does not enter the grid, because a fluent narrative can carry an unsupported conclusion past a tired reviewer, and citation discipline is the defense. And keep a record of what the tools did in the assignment: clients, reviewers, and regulators are beginning to ask, and the shop that can answer from its workfile will be glad it kept the trail.