AI in Real Estate Valuation 2026: AVMs After the iBuyer Reckoning
What automated valuation models actually do well in 2026 — and what the Zillow Offers wind-down taught the industry about the limits of price prediction.
Automated valuation models are the oldest production AI in real estate. Zillow’s Zestimate launched in 2006; Redfin’s estimate followed; HouseCanary, CoreLogic, and Quantarium built institutional-grade AVMs for lenders and investors. By the early 2020s most large lenders had AVMs in their workflow somewhere, and several iBuyers — Zillow Offers, Opendoor, Offerpad, RedfinNow — were betting the company on their valuation models being good enough to buy houses sight unseen.
In November 2021 Zillow shut Zillow Offers down and took a roughly USD 881 million writedown across Q3 and Q4. RedfinNow followed in 2022. Opendoor survived but spent 2022 and 2023 retrenching and is a smaller business than it was at peak. By 2026 the iBuyer thesis as originally pitched is dead, and yet AVMs are more important to the industry than ever. That tension is the subject of this post.
What an AVM actually is#
An automated valuation model takes a property — address, characteristics, location — and returns an estimated market value. The model is usually a gradient-boosted ensemble (XGBoost, LightGBM, CatBoost) or in some shops a deep learning model, trained on years of sales records, MLS feed data, tax assessor records, and characteristic data scraped or licensed from public sources.
The output is a point estimate plus a confidence interval. The confidence interval is the part most consumers ignore and most institutional users care about more than the point estimate. A Zestimate that says “USD 740,000 plus or minus 8 percent” is a very different signal than “USD 740,000 plus or minus 22 percent,” and the second one is the reason no lender lets an AVM stand alone for a mortgage decision on a thin file.
The Zillow Offers retrospective#
The Zillow Offers failure is worth understanding precisely because the popular version of the story is wrong. The popular version says “their AI was bad.” The actual story is more interesting.
Zillow Offers was buying tens of thousands of homes a quarter at peak. The Zestimate-derived offer model worked reasonably well on average. It failed on three things at once: the model lagged a rapidly rising and then falling market in 2021; the operational pipeline (inspection, renovation, listing) could not absorb the volume Zillow was buying; and the labor and materials cost of fixing each house spiked at the same moment.
The valuation model was a contributor but not the only failure. The bigger lesson for AI practitioners is that a 2 percent error on average across a portfolio is fine until you are levered, until your operational pipeline is at capacity, and until the market turns. AVM error tolerances acceptable for a comparative market analysis are not acceptable for a balance-sheet buyer of homes. That is the deeper iBuyer reckoning.
The residential AVM landscape in 2026#
Zillow’s Zestimate is still the most visible consumer AVM. Zillow publishes a median error percentage by market; in stable urban markets it is typically in the 2 to 3 percent range for on-market homes and worse for off-market.
Redfin Estimate is the closest competitor on the consumer side and tends to perform comparably; the two estimates often differ by 5 to 10 percent on the same home, which is itself informative.
HouseCanary sells AVMs and a broader analytics platform to lenders, investors, and capital markets participants. Their pitch is institutional-grade precision and audit trails suitable for valuation in lending and securitization workflows.
CoreLogic has been in the AVM business for two decades and operates several models (Total Home Value, PASS, ValueScout) targeted at different lender use cases. Their data moat — tax, deed, MLS, and listing data across most of the US — is the real product.
Quantarium is the third major institutional AVM, known for explicitly modeling property characteristics with computer vision over listing photos in addition to tabular features.
Black Knight (now ICE) and Veros round out the institutional vendor list that lenders typically evaluate.
Commercial real estate AVMs#
The commercial side has different physics. Comparable sales are sparse. Buildings are heterogeneous. Income capitalization matters more than comparables. The vendors that matter:
Cherre is a data infrastructure company first and an analytics layer second. They ingest property, tenant, transaction, and market data and let asset managers and lenders build their own valuation and underwriting workflows on top.
Reonomy (acquired by Altus Group) focuses on commercial property intelligence with ownership data, transaction history, and predictive analytics aimed at brokers and investors.
CompStak is a comparable-leases marketplace; not an AVM, but the data feeding commercial AVMs frequently.
Moody’s Analytics (after the REIS acquisition) and CoStar publish CRE valuations and forecasts that institutional underwriters lean on.
For commercial AVMs the honest accuracy bar is much lower than residential, and almost no institutional buyer will treat any commercial AVM as more than a screening tool.
What AVMs are actually used for in 2026#
Lenders use AVMs in three workflows. The first is pre-application — a borrower-facing rate quote that needs a fast value. The second is portfolio monitoring — quarterly revaluation of a book of mortgages for risk and capital purposes. The third is HELOC and home equity origination, where the loan amount is small enough that a full appraisal is uneconomic and a hybrid AVM-plus-desktop-appraisal workflow has taken over.
Investors and asset managers use AVMs for portfolio mark-to-model, acquisition screening, and disposition pricing.
Insurers use AVMs as one input to replacement cost estimates. Tax assessors use AVMs to support mass assessment.
What AVMs are not used for in 2026 is balance-sheet purchase decisions on individual homes at scale, because that is what the iBuyer reckoning settled. Opendoor still uses AVMs heavily but pairs them with much more conservative pricing, in-person inspection, and tighter operational discipline than the 2020 to 2021 vintage.
Where the models still break#
AVMs degrade predictably in a few situations. Off-market homes with stale characteristic data degrade because the underlying inputs are wrong. Renovated homes where the renovation is not in the assessor record degrade because the model does not know about the new kitchen. Unique homes — historic, custom, very large or very small for their market — degrade because there are no comparables. Rural homes with sparse transaction volume degrade because the training data is thin.
The mature 2026 pattern is not to fight this with a bigger model but to expose the confidence interval honestly and route the low-confidence cases to humans. Several lenders have wired this routing into their loan origination systems directly — high-confidence AVM passes through to automated underwriting; low-confidence routes to a desktop or full appraisal.
Where AI-augmented valuation is heading#
Three trends to watch.
Computer vision over listing photos and street view imagery is becoming a standard model input. The signal is meaningful — interior condition, recent renovation, deferred maintenance, view quality — and the technical work is largely solved.
Large language models are starting to appear in valuation workflows for unstructured inputs: listing descriptions, MLS remarks, appraiser narratives, neighborhood reviews. The use case is narrow but real for explanation and adjustment reasoning, not for the core point estimate.
Regulatory attention is rising. The CFPB and federal banking regulators issued joint guidance on AVMs and bias mitigation; the EU has been clearer that AVM-driven decisioning falls under high-risk categories of the AI Act. Lenders are spending more time on AVM governance than they did three years ago and that is a healthy direction.
Where pdpspectra fits#
Our AI and data engineering practice builds valuation and decisioning systems for real estate platforms, lenders, and institutional investors — model training, feature pipelines, monitoring, and the governance scaffolding that makes models defensible.
Related reading: real estate operations data platforms, real estate lead scoring, and AI credit underwriting explainability.
AVMs are mature, AVMs are useful, and AVMs are not iBuyers. Talk to our team about valuation and decisioning systems for your real estate or lending platform.