The Algorithm That Ate Zillow
In 2021, Zillow scaled an ambitious home-buying program driven by algorithmic pricing. The model predicted housing values with extraordinary confidence. Dashboards updated in near real time. Charts were clean. Confidence intervals were tight. To the teams watching the screens, the system appeared to know what it was doing.
The program purchased over 27,000 homes.
By the end of the year, Zillow had written down more than half a billion dollars in losses. The program was shut down. Thousands of employees were laid off. The company exited the home-buying business entirely.
What happened was not a technology failure. The model was doing exactly what models do. It was optimizing for the variables it could see. The problem was everything it could not see.
When the Dashboard Shouts and the Human Whispers
Zillow’s pricing model was trained on historical data, square footage, zip codes, comparable sales, and market trends. Within those parameters, it performed well. What it could not measure was the texture of reality. The smell of mold behind a wall. The difference between a quiet cul-de-sac and a street recently rerouted for bus traffic. The fact that a specific neighborhood’s desirability had shifted for reasons no dataset had captured.
Veteran real estate professionals saw the problem early. Regional directors and experienced agents noticed that the model was paying too much for certain properties, that the prices did not match what they were seeing on the ground. But their feedback existed in a different language than the algorithm’s confidence. They spoke in anecdotes and gut feelings. The model spoke in charts and precision.
In most organizations, when the anecdote meets the algorithm, the algorithm wins. Not because it is right, but because it looks right. A dashboard with a clean number is easier to defend in a meeting than a field observation that begins with “something feels off.”
This is what happens when speed operates without the counterweight of judgment. Not a single dramatic failure, but a systematic overconfidence that compounds until the correction arrives all at once.
The Accelerator Trap
There is a cognitive profile I encounter frequently in high-growth organizations. Call it the accelerator instinct. These are leaders and teams who equate velocity with value. They believe, often correctly, that the first mover wins and the cautious competitor loses. They are wired for action, and they are rewarded for speed.
In stable environments, this instinct is an asset. In environments where the data is incomplete, the model is limited, and the stakes are high, it becomes a trap.
The accelerator accepts the output because questioning it feels like slowing down. Slowing down feels like losing. And so the organization moves faster and faster in a direction that was never properly interrogated, until the gap between the model’s reality and actual reality becomes too large to ignore.
Zillow did not lack smart people. It lacked a mechanism for those smart people to challenge the model when the model was wrong. The structural incentive was to trust the machine and move fast. The structural cost of pausing was measured in lost deals. The structural cost of not pausing was measured in half a billion dollars. But only one of those costs was visible in advance.
The Convergence That Did Not Happen
Imagine a different version of this story. One where the data team and the field team had a regular, structured process for surfacing disagreements between the model’s output and on-the-ground reality. Where a regional director’s observation that prices felt inflated in a specific market was not dismissed as anecdotal but treated as a signal worth investigating.
This is not a fantasy of slower organizations. It is a description of organizations that have learned to pair speed with judgment deliberately. Where the person who sees the pattern and the person who feels the friction are required to sit in the same room and reconcile their views before capital is deployed.
That reconciliation is uncomfortable. It takes time. It often produces ambiguity rather than clarity. But ambiguity, in a world of overconfident models, is frequently the most honest assessment available.
The Lesson for Every Leader
You do not need to be in the real estate business to learn from Zillow’s experience. Every organization that relies on algorithmic recommendations faces the same structural vulnerability. The model sees what it was trained to see. It misses what it was not. And the humans around it, under pressure to move fast, are incentivized to accept the output rather than challenge it.
The question worth asking is not “is our model good?” It is “what would we miss if our model is wrong, and do we have a process for catching it?”
If the only people in the room are those who trust the dashboard, the dashboard is running the company. And dashboards do not feel the ground shift beneath them.
Speed without judgment does not just fail. It fails at scale. And by the time the correction arrives, the cost has already been paid.
