THE ARTICLE · 10 MIN
Software now exists that values homes, scans listings and runs the rental maths, and some of it is sold with “AI” in the name. This page is about that kind of tool in general. In it, an “AI property tool” means something that does some mix of three jobs: estimating what a home is worth, finding listings (including foreclosures and homes that are not for sale), and working out rent, costs and loan cover. For each job, it looks at what the published evidence says, and it ends with questions you can ask of any such tool.
Nothing here is financial, investment or legal advice. For a decision about a property or a loan, speak to a licensed professional where you live.
What an automated valuation is
A study published in 2024 by the US Department of Housing and Urban Development describes an automated valuation model, or AVM, as “a computer-driven mathematical formula that uses property characteristics, local market information, and price trends to arrive at an estimated value for a property”.
Zillow’s Zestimate is one example, and Zillow publishes how it works and how often it misses. It says the Zestimate “incorporates public records, MLS data and user-submitted home details” into its own valuation model, which it describes as neural-network based. It also says Zestimates “are available only for residential properties”: large multi-family buildings, vacant land and commercial properties are not eligible.
How accurate the estimates are
Zillow states: “The nationwide median error rate for the Zestimate is 1.79% for homes that are on the market and 7.20% for off-market homes.” It explains what a median means here: for listed homes, the estimate is within 1.79% of the final sale price “half of the time”. The other half of the time, it is further off.
The company gives a reason for the gap. Listed homes have more up-to-date information, “including listing details and recent market activity”. In its table for listed homes in large metro areas, last refreshed on 16 September 2026, the median error runs from 1.18% in Raleigh to 2.56% in New York.
Two things follow from Zillow’s own figures. A home that is not for sale falls in the off-market group, where the median miss is about four times larger. And a percentage turns into money quickly: on a $300,000 home, a 7.20% miss is $21,600. Neither figure tells you how far off the estimate is for any one home. Zillow explains how it measures them: for listed homes it uses the last Zestimate before the home goes under contract, which already draws on the listing, and for off-market homes it counts “only homes that eventually sell”.
Zillow also says the Zestimate leaves out foreclosure sales, because its analysis found that they “are generally made at substantial discounts compared to non-foreclosure sales”. The Zestimate is meant to estimate the price a home “would fetch if sold for its full value”. That is a different question from what a foreclosure will sell for.
An estimate is not an appraisal
Zillow is direct about this: “It is not an appraisal and can’t be used in place of an appraisal.” Asked whether a Zestimate can be used to get a loan, its answer is no, adding that “most lending professionals and institutions will only use professional appraisals when making loan-related decisions”. It suggests visiting the home, a professional appraisal, or a comparative market analysis from a real estate agent.
The US rule described next draws the same line. It does not apply to the use of AVMs in “the development of an appraisal by a certified or licensed appraiser”.
The US rule for lenders’ automated valuations
Six US agencies (the OCC, the Federal Reserve, the FDIC, the NCUA, the CFPB and the FHFA) published a final rule on 7 August 2024 setting quality-control standards for AVMs. It took effect on 1 October 2025, and the federal eCFR listed its sections as in force when we checked on 1 October 2026.
Lenders and secondary-market issuers that use an AVM to value a home securing a mortgage must have policies and controls so that the model meets standards designed to:
- “ensure a high level of confidence in the estimates produced by AVMs”;
- “protect against the manipulation of data”;
- “seek to avoid conflicts of interest”;
- “require random sample testing and reviews”;
- “comply with applicable nondiscrimination laws”.
The rule does not prescribe how. The agencies point to “the flexibility provided to institutions under the final rule to design policies, practices, procedures, and control systems”.
Its scope is narrow. It covers mortgages “secured by a consumer’s principal dwelling”, and a consumer can have only one principal dwelling at a time, so “a vacation or other second home would not be a principal dwelling”. It is a rule about the models lenders use, so by its own definitions it does not set a standard for a value that a website or an investment tool shows you.
Bias: what one study found
In the rule’s background section, the agencies wrote of “increasing concerns about the potential for AVMs to produce property estimates that reflect discriminatory bias, such as by replicating systemic inaccuracies and historical patterns of discrimination”.
A study in HUD’s journal Cityscape in 2024, by Linna Zhu and Michael Neal of the Urban Institute and Caitlin Young of Yale Law School, tested this in counties of the Atlanta and Memphis areas. The authors note the hope that AVMs would help with appraisal bias by “eliminating the appraiser’s input”. They added data on each property’s condition and used machine-learning methods, and still found “evidence that AVMs yield larger valuation errors in majority-Black neighborhoods”. Condition mattered too: in one of the study’s regressions, a home rated poor rather than good had an error larger by 4.35 percentage points.
The gap is in percentage terms. In the study’s figures for 2000 to 2018, errors measured in dollars were mostly smaller in majority-Black neighbourhoods, where average home values were lower. This is one study, whose main analysis uses 2018 sales in two metro areas. It shows that a gap in percentage errors remained after the authors accounted for property condition; it cannot say how large the gap is elsewhere, today, or for any particular tool.
Valuing today is not forecasting tomorrow
On 2 November 2021, Zillow’s board decided to wind down Zillow Offers, the company’s business buying and selling homes. Its filing gave the reasons as “home pricing unpredictability, capacity constraints and other operational challenges”, and recorded a $304.4 million write-down on homes bought “at higher prices than Zillow Group’s current estimates of future selling prices after selling costs”. It expected the wind-down to cut its workforce by about 25%.
Zillow’s shareholder letter that day said the business rested on “the need to forecast the price of homes accurately three to six months into the future”, and a few sentences later: “We have been unable to accurately forecast future home prices at different times in both directions by much more than we modeled as possible.” The chief executive, Rich Barton, said “the unpredictability in forecasting home prices far exceeds what we anticipated”.
The same letter names a pandemic and an unprecedented rise in prices as part of the story, so this is not proof that prices can never be forecast. It is a company with its own valuation model explaining why its forecasts were not good enough to run that business. A valuation estimates today’s price; a deal often depends on tomorrow’s.
What deal-finding tools do, and what is known about them
Of the three jobs, the third is mostly arithmetic. One calculation such a tool may run is the debt service coverage ratio, or DSCR. The OCC’s handbook for bank examiners says the DSCR is “calculated by dividing the NOI by the annual debt service requirements”, where net operating income is “annual gross income less operating expenses”.
The formula is simple, and so are its limits. The ratio is only as good as the rent, the expenses and the loan terms put into it. If the rent is itself an estimate, any error in it carries straight into the result.
Do the AI parts find better deals than a person would? We looked for an independent test, a study comparing what such tools picked with how those deals turned out, and found none. What we found were vendors’ own case studies and review lists. That does not show the tools fail. It means the claim has not been tested in public in anything we could find.
What US regulators say about money-making pitches
In October 2021 the US Federal Trade Commission sent a notice to more than 1,100 businesses that pitch money-making ventures. It summarised earlier cases finding it unlawful to make false or misleading claims about earnings, “for example, representations that participants will make a profit, or that represented profits are typical”. It also listed “falsely telling consumers they do not need experience to earn income or that they must act immediately to participate”. Penalties at the time were up to $43,792 per violation; after yearly inflation adjustments, the maximum in force when we checked in October 2026 was $53,088.
The FTC’s consumer advice puts it plainly: “If it promises guaranteed income, large returns, or a ‘proven system,’ it’s likely a scam.” On real estate seminar and investment coaching scams, it says such schemes “routinely use made-up ‘stories’ and statistics, claim a guaranteed return on your investment, or stress how cutting-edge their offer is”. On success stories: “They might not be true or typical.”
On 25 September 2024 the FTC announced five law enforcement actions under the name Operation AI Comply. They were about a tool for writing fake reviews, an “AI lawyer” service, and “multiple companies claiming that they could use AI to help consumers make money through online storefronts”, not about property. In one, the FTC alleged that a scheme “falsely claimed its ‘cutting edge’ AI-powered tools would help consumers quickly earn thousands of dollars a month in passive income”. These are the FTC’s allegations; the case it filed in court is for the court to decide. The FTC’s chair at the time, Lina M. Khan, said: “there is no AI exemption from the laws on the books.”
Questions to ask of any property tool
These are questions, not advice. The answers are yours to weigh, with a licensed professional where the money is large.
- Where does the data come from, and how fresh is it? One large property site, for example, names its inputs and says it refreshes its estimates “multiple times per week”.
- Is the number an estimate or an appraisal? An estimate shown on a screen is not an appraisal. A lender using a model to value a home for a mortgage on someone’s principal dwelling falls under the rule above.
- Does it publish its error rate, for listed and unlisted homes separately? At least one large site publishes both, and the gap between the two is the useful part.
- Are the income figures typical, and what are they based on? The FTC suggests asking questions such as “How would the business generate income? What would my specific expenses be? Can I afford it? When would I expect to turn a profit?”
- If it offers or arranges loans, who is licensed? In the US, companies and individuals that originate mortgages must be licensed or registered through the Nationwide Multistate Licensing System, and anyone can check status on NMLS Consumer Access. For investment professionals, FINRA’s BrokerCheck links to the SEC’s adviser database. These are US registers; other countries keep their own.
- Does the pitch lean on urgency, guarantees or “no experience needed”? Those are the claims the FTC has named.
Sources
- Zillow, “What is a Zestimate?”, read 1 October 2026 (table last refreshed 16 September 2026).
- OCC, Federal Reserve, FDIC, NCUA, CFPB and FHFA, “Quality Control Standards for Automated Valuation Models”, final rule, 89 Fed. Reg. 64538, 7 August 2024; eCFR, 12 CFR part 34, subpart I.
- L. Zhu, M. Neal and C. Young, “Racial Disparities in Automated Valuation Models: New Evidence Using Property Condition and Machine Learning”, Cityscape 26(1), 2024, HUD Office of Policy Development and Research.
- Zillow Group, Form 8-K, 2 November 2021, with its press release (Exhibit 99.1) and shareholder letter (Exhibit 99.3).
- OCC, Comptroller’s Handbook: Commercial Real Estate Lending, version 2.0.
- FTC, “FTC Puts Businesses on Notice that False Money-Making Claims Could Lead to Big Penalties”, 26 October 2021.
- FTC, “When a Business Offer or Coaching Program is a Scam”, consumer advice.
- FTC, “FTC Announces Crackdown on Deceptive AI Claims and Schemes”, 25 September 2024.
- Conference of State Bank Supervisors, “Protecting Homebuyers with NMLS and Consumer Access”.
- FINRA, “Check Registration: Sellers and Investments”.
- FTC, “FTC Publishes Inflation-Adjusted Civil Penalty Amounts for 2025”, 11 February 2025; eCFR, 16 CFR 1.98.
Checked October 2026. What we read: Zillow’s Zestimate page; the final rule in the Federal Register and its status in the eCFR; the HUD study’s abstract and full text; Zillow’s 8-K, press release and shareholder letter of 2 November 2021; the OCC handbook’s definitions; the four FTC pages and the eCFR penalty table; and the CSBS and FINRA pages. What we could not find: any independent test of whether AI deal-finding tools pick better deals, or published error rates for those tools’ own estimates. If you can show any of this wrong, with a source, we want to see it.
- real estate
- ai
- money
- valuations
- scams