Will AI Kill Real Estate Valuation? AVMs, Zestimate, and the $880M Lesson Every Investor Should Know

AI-powered valuation tools promise instant property appraisals. But Zestimate's 7% off-market error rate and Zillow's $880M iBuying collapse reveal the limits. Here's what cross-border investors need to know in 2026.

Published on: April 23, 2026

Quick answer: AI is not killing real estate valuation, it is bifurcating it. Automated valuation models now cover more than 116 million U.S. homes and are genuinely useful for screening, portfolio monitoring, and loan origination at scale, but Zillow's Zestimate still carries a median error of roughly 1.8–2.4% on-market and 7.0–7.2% off-market, and Zillow lost more than $880 million when it bet its iBuying program on its own model. For unique, luxury, thin, and cross-border assets, most of what serious international investors buy, a human appraiser remains irreplaceable. The winning approach is hybrid: AI for speed and screening, human expertise for the deciding input.

Ask any proptech founder in 2026 and you'll hear the same pitch: artificial intelligence is about to do to real estate appraisal what Uber did to taxi medallions. Algorithms process millions of data points in seconds. They don't miss comparables. They don't get tired, biased, or paid by the deal. And crucially, they cost a fraction of a licensed appraiser.

The numbers sound convincing. Industry data suggests firms using machine learning in commercial real estate have lifted net operating income by up to 10%, and by mid-2025 more than 90% of commercial real estate companies had launched AI pilots. Residential AVMs, automated valuation models, now cover more than 116 million U.S. homes, generating values in milliseconds.

So the question is no longer whether AI can value property. It can. The question is whether it should be the primary input for a six- or seven-figure investment decision. And the answer, if you look past the marketing, is more interesting than either the boosters or the skeptics want to admit.

What an AVM actually is (and what it isn't)

An automated valuation model is a piece of software that estimates a property's value using a combination of public records, tax assessments, recent comparable sales, property characteristics, and, increasingly, machine learning models trained on millions of transactions.

Most modern AVMs stack two or more techniques. A hedonic model assigns values to individual property attributes (an extra bedroom is worth X, proximity to a top-ranked school is worth Y). A comparable-sales model operates the way an appraiser would: it finds similar recently sold properties nearby, adjusts for differences, and triangulates a value. More sophisticated systems layer in a repeat-sales index, tracking how specific properties have appreciated over time, and weight each method by a confidence score.

The output looks like a single clean number. Under the hood it's a probability distribution with a standard deviation, a confidence score, and a quiet assumption: that the data it trained on still describes the market today.

This is where things get interesting.

The Zestimate: useful, but don't confuse it with an appraisal

Zillow's Zestimate is the most publicly tracked AVM in the world, which makes it an unusually honest case study. The company publishes its own accuracy data, and the picture is more nuanced than the headline suggests.

As of 2025, Zillow's published nationwide median error rate is roughly 1.8–2.4% for on-market homes and 7.0–7.2% for off-market homes. Here's what that actually means in dollar terms:

• $500,000 home, on-market: typical miss of ~$9,000–$12,000
• $500,000 home, off-market: typical miss of ~$35,000–$36,000
• $1,000,000 home, off-market: typical miss of ~$70,000+

And that's the median, meaning half of all estimates are off by more than those figures. Zillow reports that around 99% of properties land within 20% of the sale price, which is impressive in aggregate but leaves enormous room for individual error on any one deal.

Accuracy also collapses at the edges. In rural Vermont, off-market Zestimates have historically shown median errors above 12%. One Virginia agent described a Blue Ridge Mountains listing that Zillow estimated at $650,000, and that sold for more than $1.1 million. The algorithm wasn't broken; it simply had nothing comparable to benchmark against.

The general rule: AVMs perform best where they have the most data (dense urban markets with frequent, similar transactions) and worst where buyers actually pay the biggest premiums (unique architecture, luxury properties, view lots, heritage homes, rural and exurban markets).

The $880 million lesson: when Zillow bet the company on its own algorithm

If you want to understand the limits of AI valuation, look at what happened when one of the most data-rich real estate companies on earth tried to use its own model to buy houses for cash.

Zillow launched Zillow Offers, its iBuying program, in 2018. The thesis was elegant: use the Zestimate to make instant cash offers, buy homes at a small discount, do light renovations, and resell. Scale the Zestimate into a balance sheet, not just a marketing tool.

It ended in November 2021. Zillow shut down the program, wrote down hundreds of millions of dollars in inventory losses, and cut roughly 25% of its workforce. Total losses on the iBuying program are widely estimated at more than $880 million.

The post-mortems are worth reading, because the failure wasn't where people expected. The algorithm didn't break mathematically. It broke because:

1. Concept drift. The housing market shifted in 2021. The model kept pricing homes as if the hot market of early 2021 would continue indefinitely. When the market cooled, Zillow was sitting on thousands of homes bought at peak assumptions.
2. Seller adverse selection. Sellers with properties the model overvalued were more likely to accept Zillow's offer. Sellers with properties the model undervalued simply listed with a broker. Zillow's portfolio concentrated, systematically, in the homes its algorithm got most wrong.
3. Unmodeled features. The AI couldn't see what human buyers actually respond to: architectural charm, neighborhood feel, noise from a nearby highway, a neighbor who doesn't mow their lawn, a recent kitchen remodel never recorded in public data.
4. Operational reality. Renovation costs, contractor availability, and resale timelines didn't fit inside the model's assumptions.

There's an academic afterword. Harvard Business School researchers found that after Zillow Offers shut down, home sellers began deviating more from Zestimate when setting list prices, suggesting consumer trust in the algorithm itself was damaged by the iBuyer failure, even though the Zestimate kept running.

The takeaway isn't that AI valuation is useless. It's that when a company with effectively unlimited data and engineering talent tried to bet on its own model at scale, the model lost.

Where AI actually wins

Despite all of the above, the direction of travel is clear: AI is eating more and more of the valuation workflow. The honest map of where it's winning looks like this:

Portfolio-level monitoring. Asset managers overseeing hundreds or thousands of properties cannot re-appraise each asset quarterly. An AVM can flag which ones have drifted meaningfully in value and deserve human attention. Errors wash out at portfolio scale; they bite at the single-asset level.

Initial screening and deal sourcing. For an investor looking at 500 listings, running them through an AVM to filter undervalued candidates is obviously better than eyeballing them. The model doesn't need to be right; it needs to be directionally useful.

Loan origination at scale. U.S. federal regulators issued quality-control standards for AVMs that took effect October 1, 2025, and lenders now widely use AVMs for certain refinances, home equity lines, and portfolio monitoring where a full appraisal would be overkill. At roughly $500 per appraisal on average, the cost savings are real.

Commercial real estate underwriting support. JLL and others report that AI-augmented workflows are producing roughly 23% faster transaction times and claim 18% more accurate valuations in aggregate. The framing matters: "augmented," not "replaced."

Mass appraisal for tax and insurance. Tax assessors don't need per-house precision; they need consistency across millions of parcels. AVMs are well-suited to this.

Where AI still loses, and where your money is

The failure modes of AI valuation map almost perfectly onto the profile of the properties cross-border investors actually care about.

Unique or luxury assets. Branded residences in Dubai, restored Ottoman mansions on the Bosphorus, historic Lisbon townhouses, ski chalets in the Dolomites, the inventory that drives premium pricing is the inventory AVMs systematically misprice. There are too few comparables and the value drivers are qualitative.

Thin markets. A villa on a Greek island that trades every 15 years is not a statistics problem. There is no "signal" in the data, only noise.

Cross-border and emerging markets. AVMs require data: MLS records, clean public registries, verified transaction histories. In countries where sale prices are under-declared for tax reasons, where registries are fragmented, or where off-plan sales dominate the new-build market, AVM confidence intervals widen dramatically.

Regulatory and program-driven premiums. A property that qualifies for a Golden Visa threshold, a citizenship-by-investment program, or a Special Economic Zone carries a premium AVMs rarely capture, because the premium is policy-driven, not market-driven.

Condition and renovation. This is the evergreen blind spot. Public records show a three-bedroom house was three bedrooms in 2007; they do not show that the kitchen is now marble and the roof is new. Appraisers walk the property. AVMs don't.

Structural market breaks. Zillow's concept drift problem wasn't unique to Zillow. Every AVM is trained on history. When history stops being a good guide, geopolitical shocks, interest-rate regime changes, pandemics, sudden capital flight, all AVMs get worse at the same time, precisely when accurate valuation matters most.

The hybrid model that's actually winning

The most sophisticated institutional investors have stopped treating "AI vs. appraiser" as a binary. The operating model that's quietly won is:

• AI for speed, scale, and screening. Portfolio monitoring, deal filtering, preliminary offers, loan-to-value calibration.
• Human expertise for the deciding input. The final valuation on any specific, meaningful asset is done by, or at minimum signed off by, a qualified appraiser or broker with direct knowledge of the property and local market.
• AI as second opinion, not first. The AVM is used as a benchmark against human judgment, flagging when the two diverge, rather than replacing judgment.

The 15% rule that quality-control auditors use in U.S. lending captures this well: if an appraised value falls within 15% of the AVM, the appraisal gets a light review. Outside 15%, a human takes a harder look. The AVM doesn't decide, it allocates attention.

What this means if you're buying across borders

For investors evaluating property in a country you don't live in, AVMs are especially tempting and especially dangerous. You're trying to compress geography, language, and a totally different market into a single number on a screen. The models have been built on someone else's data, calibrated to someone else's market conditions.

A reasonable mental framework:

1. Trust AVMs for direction, not for decisions. If the AVM says €280,000 and the asking price is €420,000, that's a signal worth investigating, not a conclusion.
2. Widen confidence intervals in thin markets. For a property in an emerging market, a Golden Visa zone, or a unique asset class, assume the real uncertainty band is at least 2–3× what the platform advertises.
3. Pay for a human appraiser on anything over roughly $250,000. The cost is small relative to the decision. In most jurisdictions it's also required for mortgage financing anyway.
4. Ask what data the model is trained on. A U.S.-trained AVM applied to Turkish coastal resort property is not doing what its confidence score implies.
5. Beware when the market is changing fast. During inflection points, rate cuts, sanctions regimes, post-conflict reconstruction, every AVM gets less reliable simultaneously. These are exactly the moments investors reach for them.

The verdict

AI is not killing real estate valuation. It is bifurcating it.

For liquid, data-rich, standardized segments, median-priced U.S. suburban homes, portfolio-level commercial monitoring, mass appraisal, the appraiser of the 2010s is being replaced by a machine learning model with a human reviewer. That transition is already well underway and it's probably healthy.

For unique assets, luxury properties, thin markets, and cross-border investment, which is to say, most of what serious international investors actually buy, the appraiser remains irreplaceable. Not because humans are smarter than the algorithm, but because the data the algorithm needs to be smart doesn't exist.

The investors who will do best over the next decade are the ones who understand this split and position accordingly: use AI to filter a thousand listings, then use a human to evaluate the five that matter.

Zillow already paid $880 million to learn this lesson. You don't have to.

Frequently asked questions

What is an AVM?
An automated valuation model is software that estimates a property's value using public records, tax assessments, recent comparable sales, property characteristics, and machine learning trained on millions of transactions. The output looks like a single clean number, but under the hood it is a probability distribution with a confidence score.

How accurate is Zillow's Zestimate?
As of 2025, Zillow's published nationwide median error is roughly 1.8–2.4% for on-market homes and 7.0–7.2% for off-market homes. On a $500,000 off-market home that is a typical miss of about $35,000–$36,000, and because it is a median, half of all estimates are off by more.

Why did Zillow's iBuying program fail?
Zillow shut down Zillow Offers in November 2021 with estimated losses of more than $880 million. The algorithm did not break mathematically, it failed from concept drift as the 2021 market shifted, seller adverse selection that concentrated its portfolio in mispriced homes, unmodeled features like condition and neighborhood feel, and operational realities like renovation costs.

Should I rely on an AVM when buying property abroad?
Use AVMs for direction, not decisions. AVMs perform worst exactly where cross-border investors put their money, unique and luxury assets, thin markets, and emerging markets with fragmented registries. Trust the AVM to filter listings, then pay for a human appraiser on anything over roughly $250,000.

JanusHermes tracks cross-border real estate investment across 50+ countries, with market intelligence built for international buyers, Golden Visa seekers, and relocators. Explore listings on JanusHermes →

A note on the numbers: where no source is named, the market figures in this article (prices, yields, costs) are indicative estimates compiled from publicly available market data and industry reporting at the time of writing. Markets move and rules change, so treat them as a starting point and verify current figures with official sources before acting on them.