Case study
Zillow Offers: the model that only knew how to win.
The system didn't crash. It stayed confident right up until the world changed.

Rujuta Singh
4 min read
Zillow spent years telling people what their homes were worth. Then it started buying them. That went well.
You've probably used the Zestimate, the number that tells you what your home is worth. Give or take. In April 2018, Zillow decided to put some money where its model was. Zillow Offers launched in Phoenix: buy your home, fix it up a little, sell it on. The prices came from Zillow's own models.
The logic made sense. If anyone had the data to value a house, it was Zillow. For a while, it worked. Prices were rising. Zillow expanded to 25 markets. The model was rolling.
Then the market turned
By late 2021, Zillow was paying more for homes than it could sell them for. In its own words, it was "unintentionally purchasing homes at higher prices" than it expected to sell them for.
Q3 2021: a $304 million write-down, with another $240 to $265 million expected right behind it. The home-buying arm lost $421.6 million in a single quarter. On 2 November 2021, the board voted to shut it down. About 2,000 people lost their jobs, a quarter of the company.
The CEO, Rich Barton, didn't blame a bug. He said the unpredictability of forecasting home prices "far exceeds what we anticipated."
The model wasn't broken. It learned in a market that went up.
It had never seen a turn, so it couldn't price one. And Zillow wasn't careless. It bet on the risk it could see, which was growth. The risk that sank it was the one the model couldn't see: a market that stops behaving like the past.
Most AI failures look like this one. The system doesn't crash. It stays confident right up until the world changes. What was true six months ago isn't true today, and your model doesn't know that unless someone tells it.
Before you hand a decision to a model, ask
- What did it learn from?
- What has it never seen?
- Who gets to say "we might be wrong" before the bill arrives?