What Happened
Zillow Offers, launched in 2018, used the company’s Zestimate machine learning algorithm to automatically determine home values and make instant purchase offers to sellers. The program aimed to buy homes below market value, make minor improvements, and resell at a profit. Instead, the algorithm systematically overpaid for homes, purchasing thousands of properties at prices higher than their actual market value.
By late 2021, Zillow was sitting on approximately 7,000 homes purchased at inflated prices and was unable to sell them without significant losses. The total write-down was $881 million.
Timeline
Zillow Offers launched in 2018 and expanded aggressively through 2020-2021. The company purchased over 27,000 homes in 2021 alone, accelerating acquisitions even as the algorithm’s accuracy deteriorated. In October 2021, Zillow paused new purchases. On November 2, 2021, CEO Rich Barton announced the company was exiting the iBuying business entirely, writing down $881 million and laying off approximately 2,000 employees (25% of staff).
Impact
The Zillow iBuying failure became one of the most expensive AI prediction failures in corporate history. It demonstrated that even sophisticated machine learning models can fail catastrophically when market conditions shift in ways the training data did not capture. The loss wiped out years of profits and permanently altered Zillow’s business strategy.
The incident also raised questions about algorithmic decision-making in markets where AI predictions can become self-fulfilling or self-defeating. If an algorithm systematically overpays, it can distort the very market it is trying to predict.
Response
CEO Rich Barton acknowledged that the algorithm had been unable to accurately predict home prices with sufficient precision for the iBuying model to work. He stated that the unpredictability of the housing market made it too risky to continue. Zillow sold the remaining homes over subsequent months, taking additional losses on many transactions.
Lessons Learned
The Zillow disaster demonstrated several critical lessons about AI in high-stakes financial decisions. First, model accuracy requirements in financial applications are extreme — even small systematic errors compound into catastrophic losses at scale. Second, models trained on historical data may fail when market conditions change rapidly (as they did during the post-COVID housing boom). Third, organizations must have circuit breakers that halt automated decision-making when losses exceed thresholds, rather than continuing to execute a failing strategy at increasing scale.
The case also showed that domain expertise matters as much as algorithmic sophistication. Experienced real estate professionals could identify overpriced properties that the algorithm could not, suggesting that human-AI collaboration would have produced better outcomes than full automation.