Zillow's algorithmic home-buying programme (Zillow Offers) purchased homes at prices its model predicted would appreciate, then resold them for profit. The model failed to track rapidly changing market conditions in 2021. Zillow found itself holding 9,800 homes worth $2.7 billion, purchased at above-market prices. The programme was shut down with a $528 million write-down.
Zillow Offers used a machine learning model (Zestimate) to predict future home values and automate purchase bids. In 2021, US housing markets experienced unusual volatility — rapid price appreciation followed by a rapid deceleration. Zillow's model had been trained on historical data that did not reflect this pace of change. The model continued to recommend purchase prices based on optimistic projections even as the market shifted. By October 2021, Zillow had purchased approximately 9,800 homes it could not resell at the prices it had paid. The company took a $528 million write-down and shut down Zillow Offers, laying off 25% of its workforce.
How the Production Safety Framework maps to this failure
A D4 failure that destroyed $500M+ in value. The model was operating in a live market with real financial commitments, but there was no monitoring system that compared predicted valuations against actual resale prices in near-real-time. A basic D4 implementation — tracking the delta between predicted and realised values on a rolling basis — would have detected systematic over-prediction weeks before the damage was done. D4 observability in financial AI deployments is not optional: it is the mechanism by which model drift is caught before it becomes catastrophic.
Specific PSF controls mapped to each failure point
Zillow shut down Zillow Offers, wrote down $528M, and laid off 25% of staff. The event is a reference case for the necessity of production monitoring in financially consequential AI deployments.
The AIDA exam tests PSF knowledge across all 8 domains. Free to take, immediately verifiable.