AI Appraisal Software in 2026: What It Does, and What It Shouldn't

Every software pitch in 2026 says "AI." If you appraise houses for a living, you have probably seen AI appraisal software promised as the thing that will write your report, pick your comps and hand you a value. That pitch should make you nervous. You sign the report. Your license is on it. When the AMC calls at 9pm with a revision request, or a state board opens a complaint, the model is not on the phone. You are. This guide is for appraisers who want the useful part of automation without giving away the part that makes you an appraiser.

Key takeaways

• Good AI for appraisers does clerical work and catches errors. It should not form the value opinion or write your analysis.
• Automated valuation models (AVMs) are a lender's tool, now regulated under an interagency quality control rule effective October 1, 2025. The rule does not govern an appraiser's development of an appraisal.
• UAD 3.6 brings 728 GSE compliance rules. Automated checks against those rules are where software saves you the most time right now.
• Your own archive of past reports is the most defensible data set you have for adjustments. Software can turn it into numbers you can cite.
• Ask any vendor one question: can I see where every number came from?

What "AI" means in appraisal software today

"AI" is a marketing word that covers several very different things. It helps to split them apart, because they carry very different risks for your license.

The first three are clerical. They take typing and checking off your plate. The last two try to do the thinking. That is the line this whole article draws.

Machine learning appraisal, in plain terms

When people search "machine learning appraisal," they usually mean a model trained on many past sales that predicts a price for a new property. That is an AVM. It finds patterns in data: square footage, location, age, recent sales nearby. It is good at the typical house in a deep market. It struggles where appraisers earn their fee: the rural property with comps three miles out, the house with a failing roof the listing photos hide, the street that backs up to a highway, the market that turned last month.

Machine learning can also be used for smaller, safer jobs: summarizing your own grid adjustments, spotting a field that does not agree with another field, or noticing that a GLA you typed does not match the sketch. Same family of techniques, very different stakes.

Where AVMs fit, and the rule that now governs them

An automated valuation model is, in the words of the federal rule, any computerized model used by mortgage originators and secondary market issuers to determine the value of a consumer's principal dwelling. Lenders use them for home equity lines, portfolio monitoring, and to decide whether a loan even needs an appraisal.

In 2024, six agencies (the OCC, the Federal Reserve, the FDIC, the NCUA, the CFPB and the FHFA) issued a final rule on AVM quality control. It took effect October 1, 2025. Covered institutions must have policies and controls designed to:

1. ensure a high level of confidence in the estimates AVMs produce,
2. protect against the manipulation of data,
3. seek to avoid conflicts of interest,
4. require random sample testing and reviews, and
5. comply with applicable nondiscrimination laws.

That fifth factor matters. A model can carry patterns forward from the sales it learned from, and lenders that rely on AVMs now need controls that address fair lending.

Note: The CFPB's small entity compliance guide for the rule says it does not cover "the development of an appraisal by a certified or licensed appraiser." Your appraisal work is governed by USPAP and your state, not by this rule. But if a lender hands you an AVM figure, it is their tool, not your analysis.

For a longer look at what AVMs can and can't do, and at the job outlook, read Will AI replace appraisers?.

What AI appraisal software should do

Here is the honest list. These are the tasks where automation saves time without moving the judgment away from you.

1. Fill the facts you would otherwise retype

You know the routine: type the subject address into the order, then the report, then the map tool, then the flood tool. Look up the parcel ID. Copy the tax-roll year built, lot size and assessed value. None of that needs a license.

Good software fills the subject's address, parcel ID, coordinates and public-record facts from parcel and tax-roll data where it's available, and shows you where each value came from. It starts every new report from your firm profile so your license, E&O and contact details are already there. It lets you clone a past report on the same subdivision so the neighborhood description and site data don't start from zero.

Watch out: Public record is wrong more often than people think. Bedroom counts, GLA and year built on the tax roll are often out of date. Auto-fill is a starting point you verify during the inspection, never the final answer.

2. Check the report against the rules while you work

This is where UAD 3.6 changes the math. The GSEs publish Appendix H-1, the UAD Compliance Rules for the redesigned URAR: 728 rules, 592 Fatal and 136 Warning. A Fatal finding stops the submission at UCDP. Your client finds out, and then you find out, usually at the worst time.

A few real examples from Appendix H-1 give the flavor:

• UAD1461 (Fatal): "The 'Net Adjustment Total' does not equal the sum of the individual adjustments."
• UAD1557 (Fatal): "The 'Inspection Date' must be on or before the 'Effective Date of Appraisal'."
• UAD1624 (Fatal): at least one defect, damage or deficiency must be provided when the value is "Subject to Repair."
• UAD1279 (Fatal): the mandatory room photos (kitchen, full bathroom, half bathroom) must be included.
• UAD1767 (Fatal): commentary in certain fields exceeds the 540-character limit.

These are exactly the clerical errors software is good at catching. A rules engine never gets tired at 11pm. It never forgets that the half bath needs a photo. And it doesn't have opinions about your value.

Carat's AI-powered readiness checks run hundreds of the H-1 rules locally as you fill the report, and each finding links straight to the field. The goal is simple: catch it before the AMC does. We wrote a whole post on how AI catches AMC revision requests before you deliver, including what rules can't catch.

3. Turn your archive into support for adjustments

Most appraisers have years of reports sitting in an old program. Inside those reports are thousands of comps you already verified and adjustments you already made and defended. That archive is a data set. Fannie Mae's Selling Guide says adjustment rates can be determined "through statistical analysis, modeling, paired sales, or other commonly accepted methods", and that adjustments must reflect the market's reaction, not a rule of thumb.

Software can read that archive and show you your own median adjustment rates by town and price tier. It doesn't tell you what to use. It shows you what you have used, and how consistent you have been. That's a useful check against your own drift, and a starting point for paired sales and grouped analysis. The full method is in data-driven adjustments from your own past reports.

Carat does this with a bulk legacy import: a zip of ClickFORMS files or any MISMO 2.6 GSE XML export (most desktop programs can produce one). Anything unclear goes to a review queue rather than silently into a report. If you are moving programs, the guide to switching appraisal software without losing your comps covers the export side.

4. Do the arithmetic and the layout

Running totals in the grid. GLA calculated from the sketch under ANSI Z765. Flood, aerial, plat and location maps generated from the coordinates. Photos placed in the right UAD category. MISMO 3.6 XML validated against the schema on every build. None of this is glamorous, and calling it "AI" would be a stretch. It is automation, and it gives you back an hour on every report.

What AI appraisal software shouldn't do

It shouldn't write your analysis

Commentary is where you explain your thinking: why this comp and not that one, why the condition adjustment is what it is, why the market is stable rather than increasing. Under USPAP, that reasoning is yours and the report has to reflect it. Generated text can read well and still say something you didn't conclude. If a reviewer asks you to explain a sentence, "the software wrote it" is not an answer.

There is a practical problem too. UAD 3.6 sets character limits on commentary fields (140, 210, 540 and 2,500 characters appear in the H-1 rules). Short fields reward clear, specific sentences from someone who saw the house. Generic filler doesn't fit, and reviewers notice it.

It shouldn't hand you a value

An AVM number on your screen while you work is an anchor. It's human nature to drift toward it. If a tool shows you a "suggested value" before you have done the analysis, ask why you need it. Your value has to come from your comps, your adjustments and your reconciliation.

It shouldn't pull data you can't trace

If a field fills itself, you should be able to see the source. If an adjustment rate is suggested, you should be able to see which sales it came from. "Nothing guessed" is a good standard: data that's unclear should come to you as a question, not go into the report as an answer.

It shouldn't sign or submit for you

The appraiser signs. The lender or AMC submits to UCDP (and to EAD for FHA). Software can package the UCDP-ready zip, but the act of delivering a signed opinion of value is yours. Be wary of anything that blurs that line.

Desktop, hybrid and the push to automate the inspection

The biggest structural change from technology isn't inside your software. It's in how lenders scope assignments. Fannie Mae allows desktop appraisals on certain one-unit, principal-residence purchases with an LTV of 90% or less and a DU Approve/Eligible recommendation. It allows hybrid appraisals where a trained, vetted third party collects interior and exterior data and the appraiser relies on it. It also offers value acceptance + property data, where a data collector visits the home, no appraisal is obtained, and Fannie Mae accepts the lender's value estimate. Freddie Mac has parallel options, including ACE+ PDR and the hybrid Form 70H.

The redesigned URAR handles this with one report that changes shape by scope. In the UAD 3.6 schema, the interior and exterior inspection method for the appraiser can be Physical, Virtual or No Inspection, and the H-1 rules check that your certification matches the inspection you reported.

These scopes put more weight on the data you receive and on your ability to question it. We cover the details, the forms and the risks in hybrid, desktop and bifurcated appraisals.

A day with and without the clerical automation

Picture a normal Tuesday: two inspections, one report to finish, one revision to answer. Without automation, the report starts with retyping the subject from the order, looking up the parcel on the county site, copying the tax-roll facts, measuring the sketch and typing GLA into three places, adding up the grid by hand, and then reading the whole thing one more time hoping you catch the missing photo. With it, the subject and parcel facts are already there for you to verify, GLA comes from the sketch, the grid totals itself, and the readiness panel tells you the half bath has no photo before you package anything.

What didn't change: you still drove to the house, you still picked the comps, you still decided the condition rating, you still made and explained every adjustment, and you still reconciled to a value you're willing to sign. That's the whole idea. The time you save goes to the parts of the job that a client is actually paying for, or back to your evening.

How to evaluate AI tools for appraisers

Use these questions in any demo. They work whether you are looking at Carat or anyone else.

For a broader comparison of programs, including how to weigh price, see the honest buyer's guide to appraisal software for UAD 3.6.

Appraisal automation and USPAP

USPAP doesn't ban software. It holds you responsible for what you sign. A few general points worth keeping in mind:

• Competency. If you use a tool to support an analysis, you should understand what it did well enough to explain it.
• Credible results. An adjustment or trend that came from software is still your adjustment. The workfile should show the support.
• Record keeping. Your workfile has to be retained for at least five years after preparation, or at least two years after final disposition of a judicial proceeding in which you testified, whichever is later. If a tool held part of your support, make sure you can get it back out. The Appraisal Foundation publishes the current edition.

This isn't legal advice. It's a reminder that the easiest software to defend is the software that shows its work.

Where this leaves the appraiser

The appraisal profession has absorbed technology before: the move from typewritten 1004s to forms software, from film to digital photos, from faxed reports to UCDP. Each change took clerical work away and left the judgment. AI should be the same. Let it fill the parcel data, add up the grid, check the 728 rules and remind you what you paid for a garage bay in that town last year. Keep the inspection, the analysis, the reconciliation and the signature.

The UAD 3.6 mandate arrives November 2, 2026. If you are rebuilding your workflow anyway, build it so the machine does the typing and you do the appraising.

A short checklist to take with you:

1. Turn on rules checking and fix every Fatal finding before you deliver.
2. Treat auto-filled public record as a draft you verify on site.
3. Pull your own adjustment history and compare it with what you're about to use.
4. Write your own commentary, short and specific, inside the UAD 3.6 limits.
5. Keep your workfile complete and exportable.

Carat is built on that split: AI-powered readiness checks, adjustment rates learned from your own reports, and smart auto-fill from parcel data, for $99 a month per appraiser, unlimited reports. You sign. Always your report. If that fits how you work, you can start an account here.