AI for Real Estate Appraisals: Report Narratives and the USPAP Checks

Yes, as a drafting aid for text you check, rewrite and sign. USPAP does not prohibit tools, but it keeps the appraiser responsible for competency, credible analysis, confidentiality and a report that is not misleading. Use AI for wording and structure, keep client data out of public tools, and never let it supply a figure, a comparable or an opinion.

Where this goes wrong is quiet. A model writes a fluent paragraph saying "the subject's neighbourhood has seen steady appreciation" when your data shows two flat quarters, and the sentence survives because it reads well. The checks below are the ones that stop that from reaching a signed report, and they are mostly about discipline in the workflow rather than about the software.

Follow me on Instagram@sagnikteaches

What USPAP puts on the appraiser, and what that means for tools

I will describe these provisions in plain words as I understand them, and you should confirm the current wording in the edition you work under (the Appraisal Foundation's USPAP page still lists the 2024 edition, in effect since 1 January 2024, as current in November 2026). Treat this as a map, not the rulebook.

Connect on LinkedInSagnik Bhattacharya

On technology specifically, the Appraisal Standards Board adopted Advisory Opinion 41, Use of Technology in an Appraisal or Appraisal Review Assignment, on 23 April 2026. The Foundation describes it as guidance on appraisers' USPAP responsibilities when using tools such as automated valuation models, regression and statistical software, and generative AI, and says it will be folded into a future edition of the Guidance and Reference Manual. I could not open the opinion's own text, so read it from the Foundation before relying on any summary, including this one. The consistent message in published commentary is that the tool does not take on the responsibility: the appraiser remains accountable for the analysis and conclusions, and relying on outputs or terms you do not understand can raise a competency problem. Your licensing board and your lender or client may add rules of their own.

Subscribe on YouTube@sagnikteaches

The three-lane rule for report content

Sort every part of the report into one of three lanes before you open a tool.

1. Green: wording only. Text where the facts are already established and you want clearer, tighter or more consistent prose: the neighbourhood description, the highest-and-best-use narrative from your own conclusions, the scope-of-work boilerplate, a summary of the market-conditions data you have already analysed.
2. Amber: draft with data you supply. Adjustment explanations and reconciliation paragraphs, where the tool phrases the rationale for numbers you derived. Allowed only with a figure check (below).
3. Red: never. Selecting comparables, deriving adjustments, forming the value opinion, determining the extraordinary assumptions or hypothetical conditions, and signing the certification.

That lane system is deliberately blunt. It stops a convenient shortcut from becoming a habit that you cannot explain to a licensing board or a lender's review appraiser.

Keeping client data out of the tool

The Ethics Rule's confidentiality duty is the first practical constraint. An appraisal assignment includes the client's name, the property's address, the borrower and often the intended use. None of that is needed to write a good paragraph about, say, a suburban street of 1960s ranch homes.

• Generalise. Replace the address with "the subject", names with "the client", the street with "the neighbourhood". Put real details back after you have the draft.
• Choose the account. A business plan with written no-training and confidentiality terms is a safer place than a free consumer account. ChatGPT Business lists at $25 per user per month ($20 on annual billing) and Claude Team is similar; both need two seats minimum, so a solo appraiser either pays for two or uses an individual plan with training switched off. Prices change, so check the vendor pages.
• Never upload the whole report or the engagement letter. Paste the paragraph you need help with.
• Remove photographs with identifying details (house numbers, faces, licence plates) if you ever use an image tool.
• Check lender and client terms. Some engagement documents already restrict third-party tools. Read yours.

Drafting narratives: a prompt and what comes back

This example uses the neighbourhood section, the classic green-lane task. The facts come from your own data and observations.

Rewrite the paragraph below as a clear, neutral neighbourhood description for an appraisal report.
Use only the facts I give. Do not add statistics, trends, amenities or opinions that are not in the facts.
Keep it under 130 words, in plain professional English, past tense where describing observations.
If a fact is missing that a reader would expect (for example, a boundary), list it after the paragraph as "Missing".

Facts: boundaries are a major road to the north, a rail line to the east, a river to the south, and an arterial to the west.
Housing is mostly single-family ranch and split-level homes built between 1958 and 1972 on lots of about 7,500 square feet.
A primary school and a small shopping centre are within about a mile. Employment is mainly in the city centre, a 20-minute drive.
Several homes have additions. No adverse conditions observed other than rail noise on the eastern edge.

An illustrative output:

The subject neighbourhood is bounded by a major road to the north, a rail line to the east, a river to the south and an arterial road to the west. Housing is predominantly single-family ranch and split-level homes built between 1958 and 1972 on lots of about 7,500 square feet, and several properties have been extended. A primary school and a small shopping centre lie within about a mile. Employment is concentrated in the city centre, around a 20-minute drive away. Apart from rail noise along the eastern edge, no adverse conditions were observed. Missing: street names for the boundaries; the date of observation.

What you would fix: the output is good, and the "Missing" line is useful, but check it for anything that sneaked in. A model that was not told "no trends" might have written "a stable, well-established area", which is an opinion about market behaviour. Here it did not, which is why the prompt says what not to add. Also match the spelling your client's forms use ("neighborhood" or "neighbourhood"), and add the boundary names and the observation date yourself. Finally, check that "within about a mile" matches your notes; models sometimes tidy "0.8 miles" into "a mile" and "1.3 miles" into the same thing.

Adjustment notes: the amber lane in detail

The adjustment itself is yours. A paired-sales analysis, a regression, a cost-to-cure or a market extraction produces a number. The narrative explains it. AI can phrase that explanation consistently across six comparables, which saves time and reduces copy-paste errors. The condition is that every figure in the explanation is one you gave it and then checked.

Write one sentence for each adjustment below, explaining it neutrally for an appraisal report.
Use only the numbers and reasons I give. Do not state any number that is not in the list. Do not say the adjustment is "supported" or "market-derived" unless I say how.

Comparable 2: gross living area, subject 1,640 sq ft vs comparable 1,780 sq ft, adjustment -$7,000 at $50 per sq ft, derived from paired sales in this neighbourhood (my analysis).
Comparable 2: condition, comparable is in inferior condition to the subject (my condition ratings), adjustment +$9,000, derived from renovation cost estimates from a local contractor (my notes).
Comparable 4: garage, comparable has a two-car garage, subject has one-car, adjustment -$4,500, derived from my paired sales data.

Notice that the signs follow one convention: the comparable is adjusted towards the subject. Comparable 2 is 140 square feet larger, so it is adjusted down (140 times $50 is $7,000, entered as -$7,000); it is also inferior in condition, so it is adjusted up. Comparable 4 has the better garage, so it is adjusted down. A model does not know your convention and will happily word a wrong sign into a smooth sentence, so you verify the sign and the arithmetic against your grid before the prompt and again after the output.

What a drafting aid does well is the explanation structure: "Comparable 2 is larger than the subject by 140 square feet; an adjustment of $50 per square foot, derived from paired sales in the neighbourhood, was applied." Keep the wording that way: fact, figure, derivation. Delete any phrase such as "reflecting strong buyer demand" unless you have the evidence in the workfile.

A figure check for every amber paragraph

1. Underline every number, date, area and percentage in the draft.
2. Find each one in your grid, notes or data source and tick it. Anything without a source is deleted or sourced.
3. Read the draft once for direction words ("larger", "inferior", "superior", "above", "below") and confirm each matches the data.
4. Read it once for adjectives and trend words ("strong", "rising", "typical"). Keep only those you can support.
5. Re-add identifying details and spelling conventions.

This takes five to ten minutes for a full grid and is the difference between using a tool and being used by it. For more prompt wording in this area, the prompts tutorial for appraisers has further examples, and the comps and draft checks tutorial covers preparation.

Reconciliation and highest and best use: keep these yours

The reconciliation paragraph states how you weighed the approaches and why. It is the heart of the opinion, so an appraiser should write it from their own reasoning. A tool can help tidy a paragraph you wrote, but if you ask it to "write a reconciliation", it invents a rationale, perhaps one that fits the numbers but not what you actually did. In a review, "why did you weight the sales comparison approach most heavily?" needs your answer, not a plausible one.

Highest and best use is similar. The four tests (legally permissible, physically possible, financially feasible, maximally productive) need your analysis. A model can remind you of the structure, and it can polish your wording. It should not choose the use.

Using AI as a reviewer of your draft, not its author

One of the safest uses is to flip the roles: you write the paragraph, and the tool reads it as a sceptical reviewer would. This keeps the opinion yours, and it often catches the unsupported adjective you stopped seeing.

You are a review appraiser checking a draft narrative. Do not rewrite it.
List every statement that (1) states a trend, (2) states a condition or quality, or (3) gives a number,
and say what supporting data a reviewer would expect to find in the workfile for each.
Then list any wording that could mislead a reader who only read this paragraph.

[paragraph, identifiers removed]

An illustrative result for a market-conditions paragraph lists four items: "marketing times have shortened" (needs the days-on-market data by quarter), "buyers are active" (opinion; needs data or removal), "median sale price $412,000" (needs the source and date range), and a sentence that mixes two time periods without saying so. You act on each: attach the table, delete the opinion, add the source, split the sentence. The tool has told you what to check; you do the checking.

Scope of work and assumptions: a sentence you cannot delegate

The scope of work, extraordinary assumptions and hypothetical conditions are the parts of a report that most directly define what you did. They are also the parts where a smooth but inaccurate sentence does the most damage. An illustrative contrast:

The tool cannot know that the hatch was locked or that two sales were not verified with a party. Left alone, it writes the statement that is most common in reports, which is not necessarily the one that is true of this one. Write these lines from your notes, and keep them specific.

Disclosure and significant assistance: questions to settle

Two questions come up, and the answers depend on current standards, your licensing board, the client's engagement terms and what the tool actually did. I cannot answer them for you, but I can list how to approach them.

1. How should tool use be described? The certification's significant-assistance statement names individuals, so a tool's part is better set out where the report describes the work performed. Polishing prose you wrote and checked is very different from a tool producing an adjustment you adopted. Write down, for yourself, what each tool did on each report. If the answer is "wording only", say so in your internal notes.
2. Does the client require disclosure or forbid tools? Many lenders and management companies have their own rules. Read the engagement terms, and ask if unsure. If a client forbids third-party tools, that decides it.
3. What does your insurer expect? Ask your errors-and-omissions carrier whether it has a position on AI use, in writing.

Whatever you decide, apply it consistently, write it in your office procedures, and re-check when the standards or advisory material are updated.

A worked example: a single-family assignment, start to finish

An illustration with invented figures. A one-appraiser practice completes an interior inspection of a 1,640 square foot split-level, with six comparables, and a report due in four working days.

1. Day 1 (inspection and data, no AI). Measurements, photos and notes. The appraiser derives adjustments from paired sales and builds the grid in the appraisal software. No data has gone into any outside tool.
2. Day 2, morning (neighbourhood, 20 minutes). The appraiser pastes the generic facts, gets the draft above, fixes the boundary names, and notes the prompt and the output in the workfile folder.
3. Day 2, afternoon (adjustment notes, 40 minutes). The appraiser checks the grid, finds a sign error in the grid itself while preparing the prompt, corrects it, then gets consistent sentences for 14 adjustments. The figure check finds one phrase, "reflecting increased demand", that has no support; it is deleted.
4. Day 3 (reconciliation, own words). Written by the appraiser without AI. Read aloud once.
5. Day 4 (review and sign). Full read of the report against the workfile, the certification and the scope-of-work statement. The prompt log is saved.

The saving is perhaps 60 to 90 minutes on a typical assignment, almost all in the green and amber lanes. The appraiser might save more on a commercial report with long narratives, but the same lanes apply. If you would like to compare approaches for different property types, the market-conditions tutorial goes further into that section.

The workfile: what to keep when a tool helped

The Record Keeping Rule asks you to keep a workfile that supports the report. One licensing regulator's notice quotes the period as at least five years after preparation, or at least two years after final disposition of any court proceeding in which you testified about the assignment, whichever ends later. When a tool helped, a short folder inside the workfile does the job:

• The prompt as sent (with identifiers removed) and the date.
• The raw output.
• Your edited text, or a note of the sections where you changed the output.
• The figure-check tick-list.
• A one-line note on which account or tool was used.

If a reviewer asks how a sentence came about, you can show the facts you supplied, the draft and your edits. The workfile checklist covers the rest of the contents. Check that period against the edition you work under rather than relying on a figure in a tutorial.

Things that go wrong, and how they show up

Before you sign: a short list

AI-ASSISTED REPORT CHECK
[ ] No client-identifying or property-identifying data entered in any tool
[ ] Every figure, date and area traced to the grid, notes or data source
[ ] Directions of adjustments match the grid; arithmetic re-done by me
[ ] No trend, demand or condition statement I cannot support from the workfile
[ ] Reconciliation, highest and best use and the opinion written by me
[ ] Scope of work, assumptions and conditions are mine and accurately stated
[ ] Prompt, output and edits saved in the workfile
[ ] I have considered whether any tool use needs disclosure under current guidance
Signed: ______   Date: ______

If you want a second opinion on the workflow for your practice, that is what my AI implementation consultation covers, alongside the professional-body guidance in the tutorial on what appraisers should check first.

Appraisers and AI: common questions

Does USPAP forbid using AI in an appraisal?

USPAP does not ban tools. It makes the appraiser responsible for the analysis, the opinions and the report, including competency, credible results and not being misleading. The Appraisal Standards Board adopted Advisory Opinion 41 on using technology, generative AI included, in April 2026, so read its text. Whatever a tool drafts, you own every statement that you sign.

Can I paste a client's report details into a public chatbot?

Not without thinking hard. The Ethics Rule treats the assignment and its results as confidential, with limited exceptions. A public tool may store or learn from what you enter. Use a business plan with written no-training terms, strip identifiers, or keep client data out and use AI only on generic text.

Do I have to disclose that I used AI?

It depends on what the tool did and on the current standards, lender rules and your regulator. The certification's significant-assistance statement names people, so the real question is how your report describes the work a tool did. Read Advisory Opinion 41, your licensing board's position and the client's engagement terms, and ask your own errors-and-omissions insurer.

What should go in the workfile?

Keep what supports the report: data, sources, notes and the final version. If AI helped, keep the prompt, the draft output and your edits for the same retention period as the rest of the workfile, so you can show how a sentence was reached. Check the Record Keeping Rule for the current period.

Further reads

• UAD 3.6 Checklist for Solo Appraisers: Templates, Photos and AI Tools — Templates, photos and tools for the new data standard.
• AI for Commercial Appraisers: Lease Abstracts and Income Prep — Lease abstracts and income prep for commercial work.
• Can AI Help Draft a Reconsideration of Value Response? — Responding to a reconsideration request with AI help.
• AI-Ready Appraisal Software: What UAD 3.6 Means for Your Practice — Software readiness for the new data standard.
• 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 (2024 edition current; Advisory Opinion 41 adopted 23 April 2026), appraiser licensing regulators' pages on significant appraisal assistance and the Record Keeping Rule, published commentary on AO-41, and the ChatGPT Business and Claude pricing pages, all checked 4 November 2026. Read the current USPAP text yourself.