AI for Appraisers in 2026: Orders, URAR Data and the Workfile

AI can already fill much of a URAR from photos, MLS data and public records, suggest comps, draft narrative and screen a report before delivery. It can't inspect a home, reconcile sales or sign the certification. Under USPAP's Advisory Opinion 41, adopted in April 2026, checking every output stays your job.

The timing makes this harder to sort out. Freddie Mac's UAD FAQ makes UAD 3.6 mandatory for new reports submitted on or after November 2, 2026, so the AI pitch now tends to arrive bundled with a software decision. This page covers US real property appraisers: which tools fit which practice, what each one does at each stage of an assignment, and the lines USPAP, Fannie Mae and Regulation Z draw.

Fee, commercial or assessor: the tools don't overlap much

Search results for "AI for appraisers" mix three different buyers, and a product built for one rarely helps another. Sort yourself first.

C3 AI also lists mortgage and portfolio configurations, but it quotes deployments of about three months for residential and nine for commercial, which makes it an enterprise project and not something a two-person shop buys on a Tuesday.

One assignment, from bid to delivered report

Most of the appraiser's admin sits at the two ends of an assignment, and most appraisal AI sits in the middle. O*NET's task list shows the research load in between: public records, county land values, legal descriptions, zoning, nearby sales.

Autofill under UAD 3.6: reading the pitch

UAD 3.6 replaces the 1004, 1073, 1025, 2055, 1004C and 1004D with the redesigned URAR and two companion reports for updates and completions. Freddie Mac's FAQ names no required software and sends you to your vendor for functionality and testing, so AI isn't a condition of compliance.

Lender readiness adds a wrinkle. Cotality's summary of the Fannie Mae and Freddie Mac announcement of September 30 says lenders holding a GSE exception can keep using UAD 2.6, with UAD 3.6 required for all initial submissions from May 20, 2027. You may be asked for either format for a few months, and a new tool should handle what your clients still order.

Aivre sells itself as AI-first report software. Its site lists annual plans of $1,500, $2,500 and $3,500 at list price and says the platform is free to use until January 1, 2027, which gives you a no-cost way to test it. It also advertises 82% faster reporting and calls itself GSE verified and USPAP compliant. The page gives no method behind the percentage, so treat that figure as marketing. And a software vendor's claim of USPAP compliance doesn't transfer to your report, for reasons covered below.

Condition and quality ratings are where autofill can mislead you

Restb.ai's post says its computer vision sorts photos by area, scores each against the UAD condition and quality scales and combines them into overall ratings, and it leaves the final rating with the appraiser. Take that seriously. A photo of a clean kitchen says nothing about the soft floor under it, and a suggested rating already sitting in a field is easy to accept with the report due tomorrow. Rate the property from your notes first, then compare, and look hard at any disagreement.

Comp suggestions and the source of each fact

Ranking is the useful part; the choice is yours. Fannie Mae's list of unacceptable practices includes passing over the comps most similar in location and physical characteristics, making adjustments that don't reflect market reaction, and using comp data from parties with a financial interest in the sale without verifying it with a disinterested source. An auto-populated grid can do any of those quietly. Treat a ranking as a list of sales to investigate, and if you can't say why the tool preferred a sale, give the ranking no weight.

Narrative drafts and the words Fannie Mae bars

Ask a chatbot to describe a neighborhood and read the result for listing-brochure language, because that register can break a Fannie Mae rule. The Selling Guide names terms such as "pride of ownership," "good neighborhood," "crime" and "desirable" location as unsupported opinion, and its neighborhood section expects factual, specific description. Keep the drafting tool away from characterizing the area. Give it facts you've verified (boundaries, price range, marketing time) and search every draft for the barred terms before it goes in the report. Valcre's guide raises the related risk of fabricated narrative and answers it with validated source data.

What Advisory Opinion 41 expects when a tool touches an assignment

AO-41 adds no new USPAP requirements; it applies the existing Ethics, Competency, Scope of Work and Record Keeping rules to AVMs, regression software, machine learning and generative AI, on the premise that a tool can't comply with USPAP and only the appraiser can. The Appraisal Foundation's USPAP page records the April 23, 2026 adoption and links to the text. The points below follow Dennis Badger & Associates' summary, so read the opinion itself before you change your procedures.

• Competency. You don't need to reverse engineer an algorithm, but you do need to know a tool's capabilities and limits well enough to spot unreliable output. NAR pressed for that reading during the comment period, and a second draft followed.
• Confidentiality. Before confidential information or assignment results go into a tool, especially an AI or cloud system, take reasonable steps to avoid improper disclosure. On a personal ChatGPT account, OpenAI's data controls page says chats can feed model training until you switch that off under Settings > Data controls, so strip addresses and client names or use a business plan.
• Workfile. If you rely on tool output, the workfile needs enough to support your conclusions: possibly the output, related data, prompts and your analysis of how you used it. That's a reason to save prompts on the day you use them.
• Disclosure. USPAP doesn't require listing every tool. Disclosure is needed when leaving it out would mislead, such as when generative AI materially contributed to an analysis. Spell check and formatting generally don't count.
• Client restrictions. A client can prohibit or require a tool, but not to the point where credible results become impossible. If that happens, change the scope, decline or withdraw.

When a value dispute lands in your inbox

Revision requests are a good job for AI until they touch the number. Regulation Z's valuation independence rule lets a lender ask you to consider additional comparable properties, give more detail on your conclusion or correct errors. It bars pressure toward a target value and bars you from materially misrepresenting value.

So use a drafting tool to pull the specific requests out of a long reconsideration email, set the suggested sales beside your grid, and write up why each was or wasn't used, once you've made that call. Never prompt a tool to find sales that support a figure. That prompt is a record of the wrong question, and if you rely on the output it may belong in your workfile.

Your lender clients now answer for AI too

Lender AI governance rules took effect in 2026: Freddie Mac's Bulletin 2025-16 on March 3 and Fannie Mae's Lender Letter LL-2026-04 on August 6. Cooley's summary says Fannie's version reaches vendors and subcontractors, and Fannie's own listing now points to an October 2026 supplement that replaced the letter, so read the supplement. Neither summary says whether an independent appraiser counts as a third party, so ask each lender, and keep a written list of your tools and the data each receives in case a panel asks. The AI page for mortgage brokers covers how those rules look from the loan side.

Where Merlin helps an appraisal office, and where it doesn't

Merlin, the Mac assistant we build, handles the correspondence around an assignment, and nothing inside the report. It doesn't fill report forms, connect to MLS, pull comps, reach a value, run orders or do invoicing. Phone calls and text messages are outside it too, and orders in AMC portals stay there. Gmail and Google Calendar are required at setup, which rules Merlin out for an office that lives in Outlook.

• The coordinator nobody answered. Merlin's People panel shows a sender still waiting 2 to 30 days on your reply, such as a loan officer asking about an ETA or a homeowner about access. A voicemail from the listing agent won't appear, because that list comes from Gmail alone.
• Access and delay emails. Tell Merlin to write the listing agent for permits, roof age and upgrade dates ahead of the inspection, or to let a client know the report moves to Thursday. You read it first, change what you like, and nothing leaves until you say yes.
• Revision alerts while you're in the field. A routine can fire on mail from a client's domain, or with "revision" in the subject, and post a summary to chat. Routines run unattended, so give that one no permission to send anything.
• Calls you take on the Mac. Merlin can record them and, on paid plans, transcribe with speaker labels. A call on your cell in a driveway isn't captured. Recording law differs by state, so tell the other person first.

Connecting Gmail gives Merlin access to whatever sits in it, including engagement letters and borrower details. Weigh that under the same confidentiality duty as any other tool before you connect. For a cleaner inbox either way, try Gmail labels by client or AMC, and this AI note taker roundup covers other ways to capture calls.

A trial you can run on a slow Friday

Test on finished work, where you already know the right answer, and keep live assignment data out until you've read the vendor's terms.

1. Re-run one closed assignment. Put a delivered report's property through an autofill trial with client and borrower names removed, after reading the vendor's terms on addresses, then count the fields and condition ratings it got wrong against your workfile.
2. Search a narrative draft for the barred terms. Keep Fannie Mae's list as a saved search and run it on every AI draft.
3. Write a one-page tool policy. List each tool, what data it receives, its training and retention settings, and when you'd disclose its use. Check it against the text of AO-41.
4. Add a prompts folder to your workfile template. Save the prompt, output and your note on how you used it, each time you rely on one.