How AI Catches AMC Revision Requests Before You Deliver

It's 9:14pm. The email subject line says "REVISION REQUEST - URGENT." The AMC wants the half-bath photo, the net adjustment on comp 4 doesn't match the line items, and the reviewer asks why the zoning section says "legal nonconforming" with no comment on rebuild. None of it changes your value. All of it costs you an evening. This is the AMC revision request most appraisers know best, and it's the one software is best at preventing. Here is how AI-powered checks catch those requests before you deliver, which ones they can't catch, and a pre-delivery routine that works with any software.

Key takeaways

• Most revision requests fall into three buckets: rule failures, internal inconsistencies, and judgment or client-specific asks.
• UAD 3.6 publishes the rules: Appendix H-1 has 728 compliance rules, 592 Fatal and 136 Warning. Checking against them before delivery prevents the first bucket.
• Consistency checks (does the commentary agree with the grid?) prevent a lot of the second.
• The third bucket, "add a comp," "support your adjustment," is judgment. Software can help you prepare; it can't answer for you.

The three kinds of revision request

The first bucket is pure clerical error. It's also the one that most often comes back from UCDP as a hard stop under UAD 3.6, because the GSEs publish the rules and the lender's portal runs them.

What the UAD 3.6 compliance rules actually check

Appendix H-1 is a spreadsheet of rules that the GSEs use to check the MISMO 3.6 XML. Each rule has a message ID, the message text the lender sees on the UCDP Submission Summary Report, and a severity. A Fatal stops the submission. A Warning gets through but is visible to the lender and often to the AMC reviewer.

Here are real rules from H-1 that line up with requests appraisers get every week:

Notice how many of these are things you know to do and simply missed at the end of a long day. That's the point. A person reviewing their own report at 11pm skims. A rules engine doesn't.

The full rule set, and the ones that bite hardest, are in the 728 UAD 3.6 compliance rules: Fatal vs. Warning.

Note: Some H-1 messages say "For assistance with this message, contact the appraisal software vendor." Those are technical XML problems, not appraisal problems. Good software should never let you produce them in the first place.

How AI-powered checks work in practice

"AI" here is mostly disciplined rules evaluation, done continuously instead of once at the end. A useful checker does four things.

1. Runs while you work. You find out the half-bath photo is missing while you're still in the photo section, not after you've packaged the zip.
2. Understands conditional rules. UAD 3.6 is dynamic. A pool, an easement or a "subject to" condition switches on more required fields. The checker has to know which fields your report now needs.
3. Links each finding to the field. "UAD1461" is not useful on its own. "Comp 4: net adjustment doesn't equal the line items," with a click to the grid, is.
4. Separates Fatal from Warning. Fix every Fatal. Read every Warning and decide.

Carat's AI-powered readiness checks evaluate hundreds of the H-1 rules locally as you fill the report, with each finding linked straight to the field. Not every rule runs locally yet, and we say so; the share is large and growing. The final word is still UCDP's, at the lender's submission.

Consistency: the second bucket

Rules check that the data is present and valid. Reviewers also read the report as a whole, and they're good at spotting places where it disagrees with itself. Before you deliver, check:

• Commentary vs. grid. If your GLA commentary says $50 per square foot, every GLA adjustment in the grid should use $50 (or you should say why not).
• Sketch vs. report. GLA, room counts and below-grade area should match the sketch exactly. A sketch tool that writes GLA into the report removes this one.
• Value vs. approaches. If your final value sits outside the range of every approach you developed, the reconciliation has to say why.
• Condition and quality. The subject's ratings should match the photos and the defects you reported. A C2 with a "Subject to Repair" condition invites a question.
• Addresses and parties. Client name and address, lender, borrower. A wrong ZIP on the client block is a revision.
• Exhibits. License, E&O, maps, sketch, all present and current. If your exhibits come from a firm profile, an expired E&O declaration page is a one-time fix instead of a per-report risk.

Tip: Keep a short "standing requirements" note for each AMC you work with. If one client always asks for smoke and CO detector photos, or a statement on whether a legal nonconforming property can be rebuilt, put it in the report the first time.

The third bucket: judgment requests

"Please add a comp within 0.5 miles sold in the last 90 days." In a rural market that comp may not exist. "Please support your adjustments." "Please consider these three sales the borrower found." No checker can answer these for you, and none should.

Software can help you get ready for them:

• Show the sales you considered and rejected. UAD 3.6 includes a section for Additional Properties Analyzed but Not Used. Filling it answers the "did you consider..." question before it's asked.
• Show your support. If you can say "my last 40 reports in this town and price tier used a median of $X for a garage bay," the "support your adjustment" request becomes a sentence, not an afternoon. See data-driven adjustments from your own past reports.
• Keep your history searchable. When a reviewer sends three sales, you want to know in a minute whether you've appraised in that subdivision before and what you found.

For the full list of the most common requests and how to answer each one politely and fast, see the AMC revision request survival guide.

A 10-minute pre-delivery routine

Use this with any software.

1. Run the compliance check. Fix every Fatal.
2. Read every Warning. Fix it, or decide it's correct and make sure the report explains why.
3. Compare GLA and room counts to the sketch.
4. Compare every rate in your commentary to the grid.
5. Check the reconciliation against the range of indicated values.
6. Check photos: the subject views your client requires, kitchen, every bath, each comp, and anything you called out as a defect.
7. Check each commentary field fits its character limit.
8. Check the client's standing requirements.
9. Check license and E&O dates.
10. Read the final value, the effective date and the signature block one last time. Then sign.

What AI shouldn't do here

It's tempting to want a tool that also "fixes" the report. Be careful. A checker that rewrites your commentary to satisfy a rule, or changes an adjustment to match a pattern, is changing your analysis. The right behavior is to point and explain, then let you decide. The broader case for that line is in our pillar on AI appraisal software in 2026.

Revision requests are part of why so many appraisers feel worn down; we wrote about that honestly in why appraisers are burned out. You won't get to zero. But the ones that come from a missing photo or a sum that doesn't add up are worth eliminating, and you can.

Carat runs those checks as you work, for one flat price with unlimited reports. If you'd like to see your next report come back clean, create an account.