In real estate, AI mostly shows up as price estimation: models that predict what a home is worth or will sell for. The clearest public test of that came on November 2, 2021, when Zillow announced it would wind down Zillow Offers, its home-buying program, after taking a USD 304 million inventory write-down and saying it had been unable to accurately forecast future home prices. Regulators have since written rules for the same kind of model.
Key takeaways
- Zillow's November 2, 2021 shareholder letter ended its home-buying program and named price forecasting as a central reason.
- Six U.S. agencies published a final rule on August 7, 2024 setting quality control standards for automated valuation models in mortgage lending, effective October 1, 2025.
- In Canada, the Office of the Superintendent of Financial Institutions (OSFI) issued Guideline E-23 on model risk on September 11, 2025, effective May 1, 2027, and it explicitly covers AI and machine learning.
- An estimate from a model is a prediction with an error range, not a measured sale price.
What Zillow Offers tested
Zillow Offers bought homes directly from sellers and resold them, which meant the company carried the risk if its price estimates were wrong. In its third-quarter 2021 shareholder letter, filed with the Securities and Exchange Commission, Zillow reported a USD 304 million write-down on inventory it owned at the end of the quarter. The reason given was that it had bought homes at prices above its estimates of what they would later sell for.
The letter also reports that Zillow held 9,790 homes at the end of the quarter, with 8,172 more under contract, and expected a workforce reduction of about 25 percent over the following quarters. Management called the program too risky and too volatile for its earnings, and stated that the company had been unable to forecast future home prices accurately, in either direction, beyond what its models allowed for. The letter does not blame a specific algorithm. It describes a business that took on market risk, in a period when prices moved quickly, using forecasts that were less reliable than assumed.
Why a price estimate is not a price
A valuation model takes data such as past sales, size, location, and condition, and returns a number. The number carries uncertainty that the display often hides. The gap between an estimate and a later sale grows when the market moves faster than the training data, when a home has features that are not in the data, or when few comparable sales exist. The Zillow episode is an example of a buyer who relied on the estimate with its own money at stake. Most people meet the same limit differently: an online estimate can differ from an appraisal, an agent's opinion, or the price a buyer actually pays.
The U.S. rule for valuation models in mortgages
Lenders also use automated valuation models (AVMs) to estimate the collateral value of a home. Six agencies jointly issued a final rule on quality control standards for these models: the Office of the Comptroller of the Currency, the Federal Reserve Board, the FDIC, the National Credit Union Administration, the Consumer Financial Protection Bureau, and the Federal Housing Finance Agency. The Federal Register published it on August 7, 2024, and it took effect on October 1, 2025. The CFPB's rule page describes it as applying to mortgage originators and secondary market issuers that use AVMs in certain credit decisions on a consumer's principal dwelling.
Covered institutions must adopt policies, procedures, and control systems designed to meet five quality control factors:
- A high level of confidence in the estimates
- Protection against manipulation of data
- Avoidance of conflicts of interest
- Random sample testing and reviews
- Compliance with applicable nondiscrimination laws
The rule covers lending uses. It does not set accuracy requirements for consumer-facing price estimates on real estate websites.
Canada: a model risk guideline for banks
No Canadian rule is written for property valuation alone. For federally regulated lenders, OSFI's Guideline E-23 on model risk management was issued on September 11, 2025, and takes effect on May 1, 2027. It applies to banks, insurers, and trust and loan companies in proportion to their size and risk, and its definition of a model explicitly includes AI and machine learning methods. Its principles include an enterprise-wide view of model risk, a risk-based approach, an inventory of models including vendor models, and independent review. A bank's valuation model falls under it only as one of many models the institution runs.
Where each rule or fact applies
| Use of a model | Who relies on it | What governs it, as of October 2026 |
|---|---|---|
| Public price estimate on a listing site | Buyers, sellers, agents | No dedicated federal AI rule; the estimate is informational |
| Company buying homes with its own money | The company and its investors | Securities disclosure; business risk (Zillow Offers case) |
| Collateral value in a U.S. mortgage | Lenders and secondary market issuers | Joint AVM quality control rule, effective October 1, 2025 |
| Models at a federally regulated Canadian bank | The bank | OSFI Guideline E-23, effective May 1, 2027 |
Questions to ask about a property estimate
- Does it show a range or only a single number?
- When was it last updated, and how many comparable sales sit behind it?
- Does it use the home's actual condition, or only public records?
- Who bears the cost if the estimate is wrong?
This article gives general information about regulation and technology. It is not legal, financial, or real estate advice, and it does not recommend buying or selling any property.
FAQ
Did an AI model cause Zillow's loss?
The shareholder letter attributes the wind-down to the company's inability to forecast home prices accurately and to the risk and volatility of the business. It does not name one algorithm as the cause.
Are online home value estimates regulated?
The 2024 rule covers models used by lenders and secondary market issuers in mortgage credit decisions. It does not set accuracy rules for estimates shown to consumers on listing websites.
Does the U.S. AVM rule address bias?
Yes. One of its five factors requires institutions to comply with applicable nondiscrimination laws in how they use these models.
Does Canada have a similar rule?
Not specific to valuation. OSFI's Guideline E-23 covers model risk, including AI and machine learning models, at federally regulated financial institutions from May 1, 2027.
Related reading
- Telling real AI from marketing
- AI and industry benefits
- Which industries hold up against AI
- AI in another industry: shipping
More in AI Tools and Industry.
Last verified: October 2026
Sources
- Zillow Group, Q3 2021 shareholder letter (Form 8-K exhibit 99.3), U.S. Securities and Exchange Commission filing, November 2, 2021
- Quality Control Standards for Automated Valuation Models, final rule, Consumer Financial Protection Bureau, 2024
- Quality Control Standards for Automated Valuation Models, Federal Register document 2024-16197 (89 FR 64538), August 7, 2024, summarized by the Federal Housing Finance Agency
- Guideline E-23 Model Risk Management (2027), Office of the Superintendent of Financial Institutions, September 11, 2025