Yes, an appraiser can use ChatGPT to draft market-conditions and neighbourhood text, provided every fact comes from data you have verified and you remain the author of every conclusion. Before signing, check each number against your source, remove anything unsupported, keep the prompt and output in the workfile, and confirm that no client or subject details were shared.
The reason is Advisory Opinion 41, adopted by the Appraisal Standards Board on 23 April 2026. It creates no new standards, but it says technology use does not shift responsibility: the appraiser decides whether a tool's output is accurate, relevant and appropriate for the assignment. ChatGPT is good at turning a table of figures into readable sentences. It is poor at knowing which figures matter in your market.
Which sections to hand over, and which to keep
Think of each part of a report as either a writing problem or a judgement problem. A writing problem has the facts already; the job is to organise them into clear sentences. A judgement problem has the facts but the job is deciding what they mean for value. ChatGPT can help with the first kind and should stay out of the second.
| Report text | Hand to ChatGPT? | What you supply | What you keep |
|---|---|---|---|
| Neighbourhood boundaries and land use description | Yes, as a first draft | Your own boundary notes, map observations, zoning facts you have verified | The boundaries themselves and the final wording |
| Housing supply and demand narrative | Yes | Active, pending and closed counts for stated periods, taken from the MLS | Choice of period, area and property type |
| Marketing time and price trend sentences | Yes | Median days on market, sale-to-list ratios, median prices, computed by you | Whether the trend is real or noise at this sample size |
| Concessions and financing remarks | Yes, if you have the counts | The share of sales with concessions and the typical amount, from your own comparable review | Any link between concessions and adjustments |
| Effect of market conditions on value | No | Nothing | All of it |
| Adjustment support, reconciliation, value conclusion | No | Nothing | All of it |
| Comparable selection and comparable facts | No | Nothing | All of it. A model that has not seen the sale cannot supply it |
The bottom three rows matter most. If ChatGPT proposes a time adjustment, a percentage, or a statement that a comparable "sold for" a certain price, that is invented content. The tutorial on comps prep and draft report checks covers the data side. Here the task is only the narrative around your verified figures.
The four-step test from Advisory Opinion 41
The International Society of Appraisers, quoting AO-41, sets out its four-step framework: evaluate the tool, evaluate the data, perform the analysis independently, and evaluate the credibility of the tool's output. It translates into habits you can use on every report:
- Evaluate the tool. Know what ChatGPT does: it predicts fluent text and can state wrong things with the same confidence as right ones. Note the model or plan used in the workfile.
- Evaluate the data. Confirm the MLS export covers the period, the area and the property types you say it does, and that it excludes sales that are not arm's length.
- Do the analysis yourself. Calculate months of supply, median prices and ratios in a spreadsheet. Give the tool results to describe, not raw numbers to compute.
- Check the output. Compare every figure in the draft with the spreadsheet, and read every claim about cause.
The Complete Excel Course with AI Integration
From formulas and pivot tables to Power Query, DAX, VBA and AI with Copilot, ChatGPT and Claude — hands-on from the first lesson.
By Sagnik Bhattacharya and Virginia Thorn, Primetime Emmy Award winner
Buy nowA worked example: months of supply and a price trend
Say a residential appraiser has an area of about 900 detached homes and an MLS export for the last six months. This is an illustration. All figures are invented for teaching and do not describe any real market. In the spreadsheet she computes two three-month periods.
| Measure | Prior three months | Latest three months |
|---|---|---|
| Closed sales | 39 | 51 |
| Closed sales per month | 13 | 17 |
| Active listings at period end | 41 | 38 |
| Months of supply (active divided by monthly sales) | 3.2 | 2.2 |
| Median sale price | $405,000 | $412,000 |
| Median days on market | 31 | 24 |
| Median sale-to-list ratio | 98.4% | 99.1% |
Her prompt gives the table, the period definitions and the rules. It contains no address, borrower or lender.
You are helping an appraiser draft the market-conditions paragraph of a residential report.
Use ONLY the figures below. Do not add any number, cause or forecast that is not in the table.
Area: detached homes, about 900 in the area. Periods: prior 3 months vs latest 3 months.
[table pasted]
Write 120 to 150 words in plain, neutral language. Describe the direction of each measure
and name the period for every figure. Do not describe causes, do not use words like "hot",
"strong" or "booming", and do not comment on value. End with one sentence saying the
sample is small (51 sales in the latest period).A realistic first output looks like this. It is illustrative, and the marked points are what she would fix.
Over the latest three months, 51 sales closed against 39 in the prior three months, an
increase of 31 per cent. Active listings fell from 41 to 38, so months of supply moved from
3.2 to 2.2. The median sale price rose from $405,000 to $412,000, up 1.7 per cent year over year [1].
Median days on market fell from 31 to 24 and the median sale-to-list ratio rose from 98.4 per
cent to 99.1 per cent, which indicates strong buyer demand and a seller's market [2].
These figures are drawn from a small sample of 51 sales and should be read with caution.The corrections:
- [1] "year over year" is wrong. The comparison is between consecutive three-month periods, not twelve months apart. The percentage itself is right: 7,000 divided by 405,000 is 1.7 per cent. Replace with "compared with the prior three months".
- [2] "Strong buyer demand and a seller's market" is a conclusion the data cannot carry. The ratio may hint at more competitive buying, but "seller's market" depends on thresholds that differ by market and property type, so it is her call. She deletes it or replaces it with the measured fact: "months of supply shortened, and homes sold faster and nearer to list price".
- The 31 per cent is correct (12 divided by 39), but she checks it anyway. One recomputed figure per draft is the habit that catches the day the tool is wrong.
The edited paragraph is hers, in her voice, with her figures. ChatGPT saved ten minutes of phrasing, not the thinking.
Neighbourhood text: describe the place, not the people
A neighbourhood section tends to attract adjectives, and adjectives are where risks hide. A model asked to "describe the neighbourhood" will often reach for words such as "up and coming", "desirable" or "family-friendly", and may add details about who lives there. Keep the section to observable physical, economic and land-use facts, and write the prompt so the model has nothing else to draw on.
Draft a neighbourhood description of 100 words from ONLY these notes. Do not add anything.
Do not describe the residents, their characteristics or the "feel" of the area.
Notes: boundaries are the rail line (north), the ring road (east), the river (south), Park Street (west).
Mostly detached homes built 1960 to 1985, lots 6,000 to 9,000 sq ft. One neighbourhood shopping strip
on the east edge. Elementary school inside the boundaries. No vacant lots over an acre.
Industrial use beyond the rail line. Public transport on Park Street.Read the answer for any word you did not supply. If the output says "tight-knit community" or "popular with young families", delete it. Statements about people can touch fair-lending and fair-housing concerns, and they are not support for a value opinion in any case. The tutorial on AI prompts for appraisers: narratives, neighbourhoods and reconciliation offers more prompt shapes for this section.
Concessions and marketing time: a filled-in second example
Suppose she reviewed the 51 sales and counted concessions: 9 had seller concessions, typically 2 to 3 per cent of price, and the rest none. She gives the model exactly that.
Write two plain sentences about seller concessions. Facts: 9 of 51 closed sales in the latest
three months included seller concessions, typically 2 to 3 per cent of the sale price. In the prior
three months, 11 of 39 did. No other information. Do not say whether this affects value.A good output states that concessions appeared in 9 of 51 sales, about 18 per cent, down from 11 of 39, about 28 per cent. Check those percentages: 9 divided by 51 is 17.6 per cent, and 11 divided by 39 is 28.2 per cent. Say "about 18" and "about 28". If the model rounds "2 to 3 per cent" into "around 2.5 per cent", change it back to your range. Whether concessions require an adjustment to a comparable is for you to analyse using paired data or other support, not for a tool to assert.
A thin market: when the sample is too small to describe
The worked example had 51 sales in three months. Many appraisers work in places where that number is five. Take a rural area with 6 closed sales in the latest three months and 4 in the prior three. Two sales a month means each active listing is worth half a month of supply, and with 8 listings on the market one sale more or less in the quarter moves the figure from 4.0 months to 3.4 or 4.8. A median of six prices means little. A prompt that simply says "describe the trend" invites the model to write confident sentences about noise.
The better approach is to change the data, not the prompt. Widen the period to twelve months, widen the area, or both, then tell the model what you did and why. A prompt for this case might say: "Twelve months, 23 closed sales, area widened to include two adjoining neighbourhoods because the subject's own area had too few sales. Do not describe a trend. State the counts and medians and say that the sample is small." The output should be flat and factual. That is the right output. If the draft hints at a trend, strike it. A reader of the report is better served by an honest "limited data" than by a smooth paragraph that says nothing true.
A falling market needs the same discipline
Drafting is easiest when the numbers are moving upwards and the words come easily. A softening market is where tool output tends to go wrong, because the model leans towards balanced, reassuring phrasing. Say the next period shows 28 closed sales against 40, active listings up from 30 to 44 and median days on market up from 22 to 35. The months-of-supply figure goes from 2.3 (30 divided by 13.3 sales a month) to 4.7 (44 divided by 9.3). A cautious draft might say supply "normalised" or the market "cooled slightly", which understates months of supply doubling, a 30 per cent drop in sales and a thirteen-day increase in marketing time. Compare each adjective with its number. If the number is double, the word should not be "slightly". Your prompt can help: "Do not use softening adjectives. State each change as a percentage or a count."
Before and after: from draft to signable
This pair shows the kind of edit that makes a draft yours. The first paragraph is what a tool might hand back for the supply and demand section. The second is what the appraiser leaves in the report. Both are illustrative.
Draft from the tool:
The local housing market remains solid, supported by healthy buyer demand and limited inventory.
Homes are selling quickly, and prices continue their upward trajectory, reflecting the area's desirability.
After the appraiser's edit:
In the latest three months, 51 sales closed, up from 39 in the prior three months, and active listings fell
from 41 to 38. Median days on market was 24 (prior period 31) and the median sale price was $412,000
(prior period $405,000). The sample is small, and these figures describe the area's detached homes only.Every phrase the appraiser removed was a conclusion without a number: "solid", "healthy", "limited", "quickly", "upward trajectory", "desirability". Every phrase left in has a figure and a period beside it. As a working rule, if a sentence in the market-conditions text has no number, a period or a source behind it, it should not be in the report.
A filled-in sign-off checklist
Print this and tick it on every report where AI helped with wording. The entries are an illustration for the area above.
- Source of every figure: MLS export dated, area and property type recorded. Done.
- Recomputed one figure by hand: months of supply, latest period (38 divided by 17, which is 2.2). Done.
- Periods named beside each figure: yes, "latest three months" and "prior three months". Done.
- Causes, forecasts and adjectives removed: deleted "strong buyer demand" and "seller's market". Done.
- No identifying details in the prompt: no address, borrower, lender or order number. Done.
- Prompt, output and edits saved to the workfile: yes, with the date. Done.
- Engagement terms checked for any AI restriction: none found for this client. Done.
Confidentiality and what to leave out of the prompt
USPAP's Ethics Rule treats confidential information as protected and limits disclosure to the client and a short list of permitted parties; read the exact list in the edition you work under. The drafting jobs above never need confidential information, so the simplest compliant habit is to keep it out. Leave out:
- The subject's address, parcel number, photographs and owner or borrower names.
- The lender, AMC or client name and the order number.
- Any figure that, combined with the area, identifies the subject, such as its exact sale history.
Market-level data from the MLS can raise a separate issue: the MLS's own licence terms may restrict what you can paste into third-party tools, so read those terms rather than assuming. Some clients and appraisal management companies also set rules on AI use in their engagement terms. If yours do, the engagement terms come first.
On privacy plans, OpenAI describes ChatGPT Business as not training on business content by default, at $20 per user per month billed annually or $25 billed monthly with a two-seat minimum. Consumer plans let you switch off training in settings. Neither choice replaces the habit of leaving identifying data out. For vendor questions on data handling, see questions to ask before buying AI that touches client data.
Time and cost, honestly counted
For an appraiser who writes thirty reports a month, the market-conditions and neighbourhood text might take ten to fifteen minutes each by hand once the data is in the spreadsheet. With a tested prompt, the typing falls to a few minutes, but the checking stays. If that frees five to eight minutes a report, thirty reports give back roughly two and a half to four hours a month, though only your own timings can confirm it. The cost is the plan: a paid individual plan or a business seat in the range quoted above. The real saving only appears once the prompts are stable, so spend the first week on two or three templates and test them on finished reports where you already know the right answer.
What goes into the workfile
AO-41 addresses workfile requirements, and the Record Keeping Rule already expects your file to support what the report says. A light routine costs five minutes a report. Save: the spreadsheet with the figures; the prompt you used with the date; the raw output; and a short note on what you changed and why. If the report states a figure, the file should show where it came from. The checklist in USPAP workfile checklist: what to keep and how AI can file it lists the wider contents.
Disclosure: what the report needs to say
AO-41, as quoted by the International Society of Appraisers, says USPAP contains no requirement to disclose tools or technologies simply because they were used in an assignment. What the report must do is avoid misleading intended users and give enough information for them to understand it properly. AO-41 says it may be necessary to disclose the steps you took if omitting them would mislead. If ChatGPT only helped phrase text from your calculations, most appraisers would see no change to the report's meaning. If a tool produced the analysis itself, such as a trend line you did not check, that is a different matter, and one to avoid. ISA has said it does not consider routine AI-specific disclosures necessary under current requirements, though it lists cases where one may be appropriate, such as AI contributing materially to the assignment results. That is one body's view, not a rule. Check the engagement terms and your state board's guidance, and use the Appraisal Institute's guidance on AI as a second reference.
Four ways this goes wrong
Quoting a statistic you never supplied. Asked to "describe the market", the model adds a regional trend or an interest-rate comment. Fix by instruction ("use only the figures below") and by reading for any number not in your table.
Period drift. A model told "last quarter" may describe "this year". State both periods in every prompt, and ask it to name the period beside each figure.
Quiet causal claims. "Prices rose due to low inventory" sounds sensible and is unproven. Ban cause words in the prompt, then search the output for "due to", "driven by" and "because".
Template creep. After ten reports, the paragraphs start to sound alike, and a stock sentence ends up in an area it does not fit. Re-read each draft against that area's data, not the last report's.
A routine you can run in fifteen minutes
- Export MLS data for the area and the two periods. Compute counts, medians, months of supply and ratios in a spreadsheet (3 to 5 minutes).
- Prepare the table and the notes with no identifying details (2 minutes).
- Paste the prompt for the section and read the output (2 minutes).
- Compare every number with the spreadsheet; delete causes, adjectives and value comments (4 minutes).
- Save the prompt, output and your edits to the workfile (2 minutes).
The saving is real but modest: the time you save goes mostly on typing, and none goes on analysis. If you find yourself trusting the draft more than the data, stop and recalculate one figure by hand.
Questions appraisers ask about AI-drafted report text
Do I have to disclose that ChatGPT drafted part of a report?
Not by default. As quoted by the International Society of Appraisers, AO-41 says USPAP has no requirement to disclose a tool simply because it was used; the duty arises only if leaving it out would mislead intended users. ISA itself says routine AI-specific disclosures are not necessary. Check your client's engagement terms too, because some clients restrict tools.
Can I paste the subject address and borrower name into ChatGPT?
Avoid it. The Ethics Rule's confidentiality duty covers client and assignment information, and the drafting job does not need it. Give the tool market data for the area, with no borrower, lender or street address. If you use a business plan that does not train on your content by default, that helps with data handling, but it does not replace your duty.
Which ChatGPT plan should an appraiser use for this?
A paid business plan is the safer starting point: OpenAI states that ChatGPT Business does not train on business content by default, and it costs $20 per user per month billed annually or $25 monthly, with a two-seat minimum. A solo appraiser can use an individual paid plan instead, with Improve the model for everyone switched off in Data controls. Whichever you use, never paste client identifiers.
Can ChatGPT calculate absorption rate and months of supply for me?
It can, but check it. Months of supply is active listings divided by monthly closed sales, and the model can mix up the period or the count. Do the sums in your spreadsheet from MLS exports and give ChatGPT the results to describe. Then recompute one figure by hand before the text goes into the report.
Further reads
- AI-Ready Appraisal Software: What UAD 3.6 Means for Your Practice — What the newer form means for appraisal software and AI tools.
- Can AI Help Draft a Reconsideration of Value Response? — Drafting a response to a lender's reconsideration request.
- UAD 3.6 Checklist for Solo Appraisers: Templates, Photos and AI Tools — Get templates and photos ready for the newer form.
- AI for Commercial Appraisers: Lease Abstracts and Income Prep — The commercial side: lease abstracts and income analysis.
- Can ChatGPT Read PDFs, Spreadsheets and Photos? What Breaks — What breaks when you give ChatGPT tables and PDFs.
- AI for Canadian Appraisers: CUSPAP, Privacy and Practical Uses — How the same questions look under a different standards body.
- AI for Real Estate Appraisals: Report Narratives and the USPAP Checks — Draft narratives and adjustment notes with AI while keeping USPAP responsibility, confidentiality and the workfile where they belong: with you.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: the Appraisal Standards Board's Advisory Opinion 41 (adopted 23 April 2026) as quoted in the International Society of Appraisers' posts on AO-41 and on AI disclosure; BVResources' report of the adoption; OpenAI's ChatGPT Business pricing page and its Data Controls FAQ; checked 4 November 2026.