Using AI for Appraisal Work: Comps Prep and Draft Report Checks

An appraiser can use AI to prepare comparable-sales data, such as cleaning exports, flagging gaps and outliers and building a tidy grid, and to check a draft for contradictions, unsupported claims and leftover template text. The tool never selects comparables, makes adjustments, rates condition or proposes a value; those stay with the appraiser, who owns the USPAP duty.

That line is easy to state and easy to blur. The blur happens when a tool offers a "suggested value" or a ranked comparables list and the output looks authoritative. So the line needs drawing in specific places: which prep jobs are safe, which draft checks catch real errors, which prompts to use and what a trustworthy output looks like.

Follow me on Instagram@sagnikteaches

The title says appraisal work generally, and the examples lean residential because that is where most solo practices work. The same principles carry to commercial files; for lease-heavy work see the lease abstracts tutorial.

Connect on LinkedInSagnik Bhattacharya

The line between prep and judgement

A simple test: if two competent appraisers could produce the same output from the same inputs, it is prep. If the output depends on their professional opinion, it is judgement. Sorting a spreadsheet by date is prep. Deciding that a sale from eleven months ago is still relevant in a changing market is judgement.

Subscribe on YouTube@sagnikteaches

The duty to your client and to USPAP does not change because a tool touched the file. Whatever appears in the report is your statement. Keep a one-line note in the workfile of how AI assisted (see the workfile tutorial in this series).

The profession has also put this in writing. On 23 April 2026 the Appraisal Standards Board adopted Advisory Opinion 41, Use of Technology in an Appraisal or Appraisal Review Assignment, which the Appraisal Foundation describes as guidance on appraisers' USPAP responsibilities when using automated valuation models, regression and statistical software and generative AI. The Foundation says it will be incorporated into a future USPAP edition. Read it before you settle your own routine, because the prompts below are only as safe as the responsibilities you keep around them.

Protecting the data before you paste anything

Comps data usually comes from licensed sources with their own terms, and your assignment carries client confidentiality. Before the first prompt:

1. Read the data source's terms on use with third-party tools. If they restrict it, stop there and do the prep in a spreadsheet.
2. Replace the subject's address and any owner details with "Subject". Comparable addresses can usually be shortened to a reference such as "Comp 1" for the AI step, then put back in your own file.
3. Choose a tool whose terms keep your inputs out of training and set the option accordingly. What to check in an AI tool's privacy policy lists the questions.
4. Keep a record of which tool and settings you used on the file.

Comps prep, step by step

Step 1: Set your screening rules in writing

Before any AI step, write the rules that you would apply by hand: maximum months since sale, distance, size band, style, any exclusions such as distressed sales. The AI applies your rules; it does not invent them. Writing them down also makes them a workfile exhibit.

Step 2: Clean the export

Raw exports contain mixed date formats, size fields with units, blank bedroom counts, concessions tucked into remarks. A prompt:

Below is a table of six sales exported from a data source, with identifiers replaced by Comp 1 to Comp 6. Return a clean table with columns: Comp, sale date (YYYY-MM-DD), sale price (number), gross living area (number, square feet), bedrooms, bathrooms, concessions mentioned in remarks (quote or "none"), and a notes column listing any field you could not read or that looks inconsistent. Do not estimate missing values. Do not add comments on value or adjustments.
[paste table]

The instruction not to estimate matters. Left to itself, a tool will happily fill a blank bedroom count with a plausible number. You want blanks left blank and flagged.

Step 3: Flag, never filter silently

Ask the tool to flag sales that break your rules and say which rule, rather than removing them. You then decide. A sale removed silently is a selection decision made by software, with no trace in your file.

Step 4: Check the arithmetic

Language models are unreliable at arithmetic across many rows. Either ask the tool to produce a spreadsheet formula and run it yourself, or compute in a spreadsheet and use AI only for the narrative. For a how-to on this safe pattern, the ChatGPT and Excel guide covers asking for formulas rather than answers.

Main worked example: six sales, one subject

Say a solo appraiser is working on a three-bedroom house of about 1,800 square feet. The data export has six candidate sales. The rules written in step 1: sold within 9 months, within 1.5 miles, living area within 20 per cent of the subject (1,440 to 2,160 square feet).

The price-per-square-foot column comes from a spreadsheet formula, not the chatbot. The AI step's value is the flags: it read 600 characters of remarks on Comp 4 and spotted the concession, which is easy to miss in a hurry. What you do next is yours. You might keep Comp 4 and investigate the concession's size, drop Comp 5 on the date rule or keep it with a written reason because the neighbourhood was stable, and exclude Comp 3 as too large. Writing down those reasons is the professional work. Nothing in the table concludes a value, and the unit price column is a descriptive aid, not an indication of one.

The time saving in this example, illustratively, is perhaps ten to fifteen minutes over doing the clean-up by hand, plus a lower chance of missing the remark. The bigger benefit is consistency across files, because the same rules and prompt get applied each time.

Draft report checks: five that pay for themselves

The second job is reading your draft for slips. Give the tool the text and the data table, with identifiers removed, and ask narrow questions. Broad questions such as "is this report good?" return flattery.

Check 1: contradictions between text and data

Compare the narrative with the data table. List every contradiction in bedroom count, bathroom count, living area, dates, condition ratings, effective age and sale prices. Quote both sides each time. If there are none in a category, write "none found". Do not suggest edits.

An illustrative output: "Narrative: 'four-bedroom home'. Table: 3 bedrooms. Narrative: 'sold in March'. Table: sale date 2026-04-12." Both would embarrass you in a lender review.

Check 2: unsupported claims

List every sentence that states a fact about the market, neighbourhood or property that is not backed by a figure, source or observation in the text provided. Quote the sentence. Do not judge whether it is true.

This finds phrases like "demand is strong" with nothing behind them. You then add your data or delete the claim.

Check 3: leftover template text

A plain search for square brackets, "TBD", "[area]" and another property's street name catches most leftovers. AI adds little here; a text search is faster and exact. Use it first, then use the AI check for subtler leftovers, such as a sentence about a garage in a file with no garage.

Check 4: reasoning gaps

Read the reconciliation section. List any step where a conclusion is stated without a reason, such as why one comparable is given more weight. Quote each place. Do not supply reasons.

It will not know the true reason, which is the point. It shows you where a reviewer will ask "why?", and you supply the answer from your own thinking.

Check 5: tone and clarity

Ask for sentences over 35 words or with more than one hedge, and tidy them yourself. This one is optional and the lowest value; skip it when time is short.

Realistic failures and how they show up

• The invented contradiction. The tool claims the narrative says 1,900 square feet when it says 1,800. It misread a nearby number. Verify every flag against the draft before editing.
• The filled blank. A cleaned table shows 3 bedrooms for Comp 2, which was blank in the export. You told it not to estimate; it did anyway. Compare the cleaned table against the original for every row, every time.
• The confident adjustment. You ask for a quick note on a feature and the reply says "a typical adjustment would be around 5,000". That is a judgement dressed as data. Ignore it, and rephrase prompts so they never invite one.
• The polished paragraph. A rewritten sentence reads better but now claims "rising demand". Compare rewritten text with the original, sentence by sentence.

These failures share a cause: trusting the output because it is fluent. The protection is procedural, which is why a short review step matters; human review for AI work without slowing down describes a light version.

Where UAD 3.6 enters, and where it should not

UAD 3.6 became mandatory for new reports submitted to the GSEs' portal on 2 November 2026. Lenders that requested a one-time exception, announced on 30 September 2026, may keep submitting new UAD 2.6 reports through 19 May 2027, so you may see both versions for a few months. That affects your prep in one practical way: the new report is data-driven, so a clean comps table with consistent field names saves time when you key or import. The GSEs' UAD compliance rules, which you run through your appraisal software provider's API, test the data for completeness, validity and reasonableness but return no Collateral Underwriter messages. Use them. Beyond that, UAD-specific steps are covered in the UAD 3.6 checklist; the rest of this page stays on prep and checks that apply to any report version.

A draft check walked through on a short extract

Here is an illustrative paragraph from a draft, with the matching data row, so you can see what the checks catch.

Draft: "The subject is a four-bedroom, two-bath home of 1,800 square feet in average condition (C3), updated in 2019. The market has remained strong with rising prices. Comparable 2 sold in March for 391,000 and required an upward adjustment for its smaller garage."

Data: Bedrooms 3. Baths 2. GLA 1,800. Condition C4. Update year 2016. Comp 2 sale date 2026-04-12, price 391,000. Garage: Comp 2 larger than subject.

Checks 1 and 2 should return something like this (illustrative):

Contradictions:
- Bedrooms: text 4, data 3.
- Condition: text C3, data C4.
- Update year: text 2019, data 2016.
- Comp 2 sale month: text March, data April.
- Comp 2 garage: text says smaller than subject, data says larger.
Unsupported claims:
- "The market has remained strong with rising prices." No figure or source given.

Five hard contradictions in one paragraph is exaggerated for the sake of the example, but the pattern is real: slips cluster where text was edited after the data changed. Every flag points you to a place to look; none tells you which side is right. The garage line is the interesting one, because it may mean the adjustment direction in your grid is also wrong, and only you can open the grid and check.

A second case: a condominium with sparse data

Say the subject is a one-bedroom condominium in a building with few recent sales. The export has four sales, three in the same building and one nearby. The prep step looks different here.

• Add a column for floor level and view, copied from remarks, flagged as "from remarks, unverified" so you remember to confirm.
• Ask the tool to group the sales by building, and to list the monthly fees where present. Fees vary and affect buyer behaviour, so a blank fee is a reason to phone the listing agent, not to guess.
• Ask for any remark about special assessments or pending litigation, quoted word for word.

The tool reads the remarks faster than you do, but only you can contact the agents, confirm the facts and decide how much weight a different building's sale deserves. With sparse data, the temptation to let the tool "suggest" a conclusion grows. Resist it; say in the report that data was limited and explain your approach in your own words.

A small prompt library, kept in one file

Reuse beats improvisation. Keep prompts in a single document, with the date of the last change, and note which ones gave problems.

Review the library every quarter. When the tool or its model changes, the same prompt can behave differently, so rerun each prompt on a past file where you know the right answer and see whether the output still matches.

When a tool offers a value anyway

Some platforms bundle automated valuation features, or an assistant will volunteer a figure when a prompt is loose. The rule is the same: do not use it, and do not let it anchor you. If you have seen a number before forming your own view, record in your notes that you did, and make sure your analysis rests on your own data and reasoning. A conclusion you cannot trace to your comparables and your adjustments is not yours to sign.

The same applies to a client or lender who asks whether an AI estimate matches yours. You can answer that your opinion is developed from your analysis and is not compared against automated outputs, or follow your engagement terms. Do not adjust your conclusion to match a number from a tool.

Measuring whether it is worth it

Track three things for ten files: minutes spent on prep and checking, number of revision requests citing factual slips, and number of AI flags that turned out to be real. An illustrative log after ten files: 22 minutes added per file, one revision request for a slip (down from three over the previous ten), and 31 flags of which 19 were real. A 61 per cent hit rate means the checks are worth the time even with false alarms. If the hit rate falls below roughly a third, tighten the prompts or drop that check. The figures are yours to gather; there is no industry benchmark I can cite.

A one-file routine you can copy

1. Write screening rules and save them to the workfile.
2. Strip identifiers; run the clean-up prompt; compare the output with the original export.
3. Compute any numbers in a spreadsheet. Keep the formulas.
4. Choose comparables yourself. Write one sentence of reasoning for each kept and each excluded sale.
5. Draft the narrative from your own notes.
6. Run checks 1, 2 and 4 on the draft; search for leftovers yourself.
7. Fix confirmed issues. Ignore false alarms. Note the unresolved ones.
8. Save a one-line note of AI use and settings to the workfile.

With a practised hand this adds about 20 minutes to a file and removes the slips that cause revision requests. Revisions cost far more than 20 minutes, particularly when a lender asks for them after your next order is already waiting. For writing paragraphs from your own notes, the series has AI prompts for appraisers.

What to do on a very busy day

When five orders land together, the temptation is to skip the checks and trust the draft. Keep a minimum version instead: run only the contradiction check and the bracket search on each file. That takes about five minutes per report and catches the slips most likely to trigger a revision request. Leave the reasoning-gap check for the files where the market was unusual or the data thin, because those are the ones a reviewer will question. Record which files had the short version so you can look again if a query comes back.

Explaining your AI use to a client or reviewer

A short, honest statement is enough. For example (illustrative): "I use software, including AI assistants, to tidy data exports and to check my draft for inconsistencies. All comparable selection, adjustments, ratings and conclusions are my own." Check your engagement terms and the lender's guidance first; some clients restrict tools or ask for disclosure. If you are unsure, ask them in writing before the assignment starts.

If you would like to set this up around your own tools and files, that is what my AI implementation consultation covers. The core habit is simple enough to start tomorrow: write your rules down, flag rather than filter, and read every output as a suspect until you have checked it against your own data.

Questions about AI and appraisal judgement

Can AI choose my comparable sales?

It can sort and flag candidates by rules you set, such as distance, date and size. Choosing the comparables, and defending that choice, remains your professional judgement. If a tool presents a ranked list, treat it as a screening aid and write your own reasons for each selection.

Is it a USPAP problem to use AI at all?

USPAP duties attach to the appraiser whatever tools are used. You remain responsible for the data, analysis and conclusions. Read Advisory Opinion 41 on the use of technology, adopted by the Appraisal Standards Board in April 2026, check your licensing board's position, and keep a note of how AI assisted each file.

Should I tell clients I use AI?

Check your engagement terms and the lender's requirements. Some clients ask about AI use or restrict it. A short line in your engagement letter describing administrative and checking uses is a reasonable default, and it avoids surprises.

What should never go into a general AI tool?

Anything you have not checked the tool's data terms for: borrower names, loan numbers, full addresses with owner details and lender-confidential material. Strip identifiers first, or use a tool with business-grade data terms and training switched off.

Further reads

• AI for Real Estate Appraisals: Report Narratives and the USPAP Checks — Report narratives and the USPAP checks around them.
• ChatGPT for Appraisers: Drafting Market-Conditions Sections — Drafting market-conditions sections with care.
• Appraisal Institute on AI: What Appraisers Should Check First — What the profession's institute says to check first.
• Can AI Help Draft a Reconsideration of Value Response? — Handling a reconsideration request after delivery.
• How to Clean Up Customer Records Before You Add AI — The data-cleaning habits behind good comp prep.
• AI-Ready Appraisal Software: What UAD 3.6 Means for Your Practice — How report platforms add AI features.
• AI for Used Car Sales: Price Each Car Against Local Listings — How to price each used car: comparable adverts into a sheet, mileage and condition adjustments, a floor, range and advertised price, and a guide cross-check.
• Best AI for Real Estate Wholesalers: Underwrite a Deal Before You Call — Use AI to screen a wholesale deal before the seller call: comps, ARV and repair sanity checks, a maximum offer sum, and the checks you still do by hand.
• AI for Canadian Appraisers: CUSPAP, Privacy and Practical Uses — CUSPAP confidentiality, competence and work-file rules read against AI, plus privacy points and practical appraisal uses that keep client data out.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: the Appraisal Foundation's USPAP page (Advisory Opinion 41, adopted 23 April 2026); Freddie Mac's 30 September 2026 UAD temporary policy exception notice and the joint GSE UAD 3.6 FAQs for the mandatory date and compliance rules, checked 4 November 2026. Worked figures are illustrative.