The House-Buying Machine That Ate Itself

01 — The BetZillow's Machine for Buying Houses

In the spring of 2021, Zillow did something no real estate company had ever tried at scale: it handed its algorithm the keys.

Zillow had spent fifteen years building the Zestimate — a machine learning model that estimated the value of nearly every home in the United States. By 2021, the company believed its model was accurate enough to do something more than estimate. It could buy. In February 2021, Zillow announced that the Zestimate would serve as a live cash offer on approximately 900,000 eligible homes across 23 markets. A homeowner could visit Zillow's site, see the model's estimate of their home's value, and accept it as an actual purchase offer — no negotiation, no agent, no waiting.

The pitch was elegant. Zillow would use algorithmic precision to buy homes at fair value, perform light renovations, and resell them for a small profit. At scale, even a thin margin on thousands of transactions would add up to something significant. The real estate market had been rising for a decade. The model had been trained on that rising market. It knew what homes were worth.

It did not know what they would be worth next quarter.

02 — The Implosion9,790 Homes, $569 Million, 2,000 Jobs

In the third quarter of 2021, Zillow bought more homes than it had in the previous eighteen months combined.

This was not an accident. As competition from rival iBuyers like Opendoor intensified, Zillow's team had adjusted the algorithm — a practice they called, internally, "offer calibration." When the model suggested a purchase price, operators would sometimes raise the actual offer by thousands of dollars above that figure to increase the odds of winning the deal. The algorithm said one number. Zillow bid higher.

The result was visible from the outside before Zillow acknowledged it publicly. Real estate analyst Mike DelPrete, working from public listing data at the University of Colorado Boulder, noticed it in October 2021. "Every key metric I've seen from Zillow over the past few months just doesn't make sense," he said. "It's like it's making decisions two to three months too late relative to the market."

Individual transactions confirmed the pattern:

On October 17, Zillow halted new home purchases, citing "operational capacity." It was holding 9,790 homes in inventory and had another 8,172 under contract — homes it was contractually obligated to still purchase.

On November 2, Zillow announced the shutdown of Zillow Offers entirely. The Q3 write-down was $304 million. Total projected write-downs: up to $569 million. On November 3, Zillow's stock fell 25%.

03 — The ModelWhat the Algorithm Got Wrong — and Why Nobody Stopped It

The technical failure had a name: concept drift.

The Zestimate model had been trained on years of housing market data, almost all of it reflecting a rising market. When the market began to cool in mid-2021 — partly because the pandemic-era suburban rush was fading as vaccines became available — the model did not detect the shift. It continued to predict prices as if the conditions that had prevailed through its entire training history were still in effect. The world changed. The model didn't know.

By the time the losses were visible to an outside analyst looking at Phoenix listing data, Zillow had accumulated nearly 18,000 homes at inflated prices — 9,790 in inventory plus 8,172 more it was still obligated to purchase.

04 — The CostWhat Was Lost

The financial accounting was large enough that the numbers lost their texture.

A $569 million write-down. A 25% workforce reduction. A stock that fell 25% on the announcement day and never recovered its pre-announcement high. A securities class action filed by investors who alleged the company had misled them about the Zestimate's accuracy and the viability of the Offers program.

Behind those numbers were the 2,000 employees who received severance packages — ten weeks of pay, outplacement services, extended healthcare — and were told their jobs no longer existed because the algorithm they had built and operated had systematically overpaid for nearly 10,000 homes.

In the markets where Zillow had been most active, the inflated purchase prices it had paid may have temporarily distorted local comparable sales data, affecting what ordinary buyers paid for nearby homes. The algorithm's consequences were not contained to Zillow's balance sheet.

05 — The LessonThe Machine You Trust Is the Machine You Don't Watch

The most common way to tell this story is as a failure of AI. The algorithm was wrong. But that framing misses the more specific failure underneath it.

Algorithms are always approximations. They're accurate under the conditions they were trained on and increasingly inaccurate as conditions drift. This is not a bug — it is the nature of any predictive model operating in a changing world. The risk is not that the model will be wrong. The risk is that no one will notice it becoming wrong until the cost is already enormous.

Zillow had every incentive to monitor its model's performance in real time. It was using that model to make purchase offers on actual homes with actual capital. Instead, it compounded the drift with manual overrides that pushed further in the wrong direction, and scaled its operations through August 2021 even as the warning signs accumulated.

Zillow named its valuation model the Zestimate. The diminutive suffix was meant to signal humility — an estimate, not a certainty. In 2021, the company forgot the distinction. The 2,000 employees who lost their jobs that November did not build a rogue AI. They built one that worked — right up until it didn't, and no one in the building was watching the right numbers.