How AI Is Changing Home Appraisals

Bottom Line Up Front: Automated valuation models (AVMs) like Zillow’s Zestimate and Redfin’s Estimate are used by millions of homeowners and buyers every day — but their accuracy varies dramatically depending on the market, property type, and data availability. AI is increasingly being accepted by lenders for certain transaction types, which has real implications for buyers, sellers, and appraisers. This article explains how AVMs actually work, what their published error rates mean in practice, and when to trust — or not trust — AI valuations.

What Is an Automated Valuation Model (AVM)?

An automated valuation model is a software system that uses statistical modeling and machine learning to estimate a property’s market value without requiring a human appraiser to physically inspect the property. AVMs ingest data from multiple sources — public property records, tax assessments, MLS transaction data, neighborhood comparable sales, property characteristics, and increasingly, computer vision analysis of property photos — and produce a value estimate along with a confidence interval.

AVMs are not new. The first generation appeared in the 1990s, and Fannie Mae and Freddie Mac began using them for collateral evaluation in the early 2000s. What’s new is their sophistication. Modern AVMs use deep learning techniques that can identify non-linear relationships between property features and value that traditional regression models miss, process far more data inputs simultaneously, and update estimates in near real-time as new transactions close.

Understanding AVM accuracy matters not just for buyers and sellers checking Zestimates, but for investors doing acquisition analysis and agents advising clients on pricing strategy. For the context of how AI is being used across real estate broadly, see our article on AI for real estate professionals. For how commercial real estate uses AI valuation differently, see our AI for commercial real estate guide.

How Zillow’s Zestimate Actually Works

Zillow’s Zestimate is the most widely used AVM in the United States, with an estimated valuation available for over 100 million properties. Zillow has never fully disclosed the specific algorithms powering the Zestimate, but has described its architecture as a neural network that incorporates hundreds of property attributes alongside geographic, market, and temporal features.

Data Sources the Zestimate Uses

The Zestimate model draws from public tax records and property assessments, MLS listing and sales data (where Zillow has data sharing agreements), agent-reported information, user-submitted property details (which Zillow allows homeowners to edit), satellite and street imagery analysis, neighborhood demographic and economic data, school ratings and district boundaries, proximity scores for amenities and infrastructure, and short-term rental activity where applicable.

The model weights these inputs differently based on their predictive reliability in each market. In dense urban markets with hundreds of comparable sales per month, recent transaction data dominates. In rural or exurban markets with few transactions, automated assessment of property attributes becomes more important.

Zestimate Accuracy: What the Numbers Actually Mean

Zillow publishes the Zestimate’s accuracy metrics on its website and updates them regularly. As of early 2026, Zillow reports a national median error rate of approximately 2.4% for on-market homes (homes currently listed for sale) and approximately 6.9% for off-market homes (not currently listed).

The median error rate means that half of all Zestimate estimates are within that percentage of the actual sale price. But “median” obscures the distribution. The other 50% of estimates are off by more — sometimes significantly more. Zillow also publishes what percentage of estimates fall within 5%, 10%, and 20% of sale price: approximately 79% of on-market Zestimates fall within 5% of sale price, 93% within 10%, and 98% within 20%.

In dollar terms: on a $500,000 home, a 2.4% median error is $12,000. A 10% error is $50,000. The 7% of on-market homes where the Zestimate is off by more than 10% represent genuine misvaluation at a scale that matters for real estate decisions. For the $500K home example, 7% of cases means the Zestimate could be off by more than $50,000.

Where the Zestimate Struggles

Zestimate accuracy is demonstrably worse in specific situations: unique or architecturally distinctive homes with few true comparables, properties with significant renovations not reflected in public records, rural and exurban markets with sparse transaction data, high-end luxury properties where small differences in finishes represent large value differences, and rapidly changing markets where recent sales haven’t yet been absorbed into the model. Zillow itself acknowledges the off-market error rate of 6.9% versus 2.4% for active listings — the model performs significantly better when it has current listing price data as a strong anchor.

Redfin’s Estimate: A Competing AVM

Redfin’s home value estimate is the second most widely used consumer AVM, and it operates on different data than the Zestimate because Redfin is an active real estate brokerage with direct MLS access in all of its operating markets. This means Redfin’s model uses actual MLS data in real time, rather than depending on data sharing agreements or public records that may lag by weeks. For more on this topic, see our guide to AI for real estate market research.

Redfin Estimate Accuracy

Redfin reports a median error rate of approximately 2.08% for on-market homes in markets where Redfin operates — slightly better than Zillow’s 2.4%. The gap is more pronounced for off-market homes in active Redfin markets, where fresher transaction data gives Redfin’s model an advantage. The Redfin Estimate is only available in markets where Redfin has MLS access; it does not have the near-universal coverage of the Zestimate.

One important difference: Redfin updates its home estimates multiple times per day as new data comes in, while the Zestimate updates less frequently for most properties. In fast-moving markets where values are changing week-to-week, more frequent updates matter.

Other Consumer AVMs

Realtor.com’s home value estimate uses a partnership with CoreLogic, one of the largest institutional AVM providers. Chase’s home value tool uses its own proprietary model. Homebot provides monthly home value reports to homeowners through their agents. Each uses similar underlying techniques but different data sources and model architectures, which is why the same property can show different estimates across platforms — sometimes by 5-10% or more.

Institutional AVMs: The Lender Side

Consumer AVMs like the Zestimate are designed for general awareness and search, not for lending decisions. The institutional AVM market — serving mortgage lenders, servicers, and investors — operates on different standards and involves different providers.

Fannie Mae’s Collateral Underwriter and Desktop Appraisal

Fannie Mae’s Collateral Underwriter (CU) is an AI system that reviews every appraisal submitted on Fannie Mae loans and flags potential overvaluations, comparable selection issues, and appraiser adjustments that appear inconsistent with market data. CU assigns a risk score to each appraisal, and high-risk scores trigger additional review or field work requirements. This AI review layer has fundamentally changed how appraisers work — they now write reports knowing an AI will scrutinize every comparable and adjustment. For more on this topic, see our article on whether AI is always right.

More significantly, Fannie Mae has expanded its “Desktop Appraisal” program, which allows appraisals to be completed without a physical property inspection for certain transaction types (typically lower LTV refinances and some purchase transactions). Desktop appraisals use existing data — photos, tax records, previous appraisals — combined with AI valuation models to complete the appraisal without the appraiser visiting the property. This is a meaningful step toward AI-driven loan collateral evaluation.

Property Inspection Waivers (PIWs)

Going further, Fannie Mae and Freddie Mac offer “property inspection waivers” — essentially appraisal waivers — for certain loans where the agencies’ AVM models show sufficient confidence in the value. When a lender submits a loan for PIW eligibility, the GSE’s AI model evaluates the property’s data profile and either approves the waiver (no appraisal required) or requires a traditional or desktop appraisal.

In 2024, approximately 25-30% of eligible purchase loans received PIWs, and over 40% of eligible refinance transactions. This represents AI effectively replacing the traditional appraisal for a substantial portion of conventional mortgage transactions. The AI-driven PIW program has significant implications for the traditional appraisal profession — a topic that connects directly to our analysis of companies trying to replace real estate professionals with AI.

CoreLogic, Black Knight, and Institutional AVM Providers

The professional AVM market is dominated by CoreLogic (which provides AVM technology to most major lenders and servicers), Black Knight (now merged with ICE Mortgage Technology), and ATTOM Data Solutions. These enterprise AVMs are calibrated for lender use, with accuracy measured against stricter standards than consumer tools and confidence interval reporting that lenders use to determine if a full appraisal is needed. Institutional AVMs typically claim median errors of 3-5% on the full property population (including rural and unusual properties) with confidence intervals reported per estimate.

When to Trust an AVM — And When Not To

For buyers and sellers, the practical question is when AVM estimates provide useful guidance and when they’re misleading.

Trust AVMs When:

• The property is a standard home type (not custom, historic, or architecturally unusual) in a neighborhood with frequent sales
• Multiple AVMs from different providers are in close agreement (within 3-5%)
• The property is currently listed (on-market estimates are significantly more accurate)
• You’re seeking a general ballpark for awareness and research, not a pricing decision
• The local market has been stable — not rapidly appreciating or declining

Be Skeptical of AVMs When:

• Multiple AVMs disagree significantly (10%+) — a sign of model uncertainty
• The property has unique features, recent renovations, or characteristics that differ from neighborhood averages
• The local market is moving rapidly in either direction
• The property is in a rural or exurban area with few comparable sales
• You’re making a significant financial decision (buying, selling, refinancing, estate settlement) — in these cases, a licensed appraiser’s on-site analysis remains valuable

For agents helping clients understand pricing, the AVM is a starting point for conversation, not a substitute for a proper comparative market analysis. The AI estimate tells you what similar homes have sold for — the CMA tells you what this specific home should sell for, accounting for its unique characteristics and local market dynamics. For more on this topic, see our AI for renters guide.

For related insights on how AI is being used in real estate beyond valuations, see our article on best AI tools for real estate agents and our guide to AI for real estate lead generation.

The Future of AI Appraisals

The appraisal industry is at a significant inflection point. Fannie Mae’s expanding PIW and desktop appraisal programs, combined with improving AVM accuracy driven by better data (including computer vision analysis of listing photos and street-level imagery), suggest that AI will handle an increasing percentage of routine residential valuations. Industry analysts project that 50%+ of residential mortgage transactions may use some form of AI-assisted or AI-replaced appraisal within the next 5-7 years for standard properties in data-rich markets.

For complex properties, estate situations, disputed values, and commercial real estate, licensed appraisers will remain essential. The AI trend in appraisals follows the same pattern as other real estate roles: routine, data-rich tasks automate first; judgment-intensive, complex situations remain human-driven longest.

Key Takeaways

• The Zestimate’s national median error rate is 2.4% for on-market homes and 6.9% for off-market — in dollar terms, potentially tens of thousands of dollars on a typical home price.
• Redfin’s estimate (2.08% median error) is marginally more accurate than Zillow’s for on-market homes in markets where Redfin has direct MLS access.
• 25-30% of eligible purchase loans in 2024 received property inspection waivers from Fannie Mae — meaning AI effectively replaced the traditional appraisal for those transactions.
• Multiple AVMs in close agreement (within 3-5%) provide stronger confidence than any single estimate; significant disagreement between AVMs signals model uncertainty.
• For major financial decisions (buying, selling, estate settlement, litigation), licensed appraiser analysis remains valuable even when AVMs provide a plausible range.
• The appraisal profession is being restructured by AI, with routine valuations increasingly handled by models and complex, unique, or disputed valuations remaining appraiser-dependent.

Frequently Asked Questions

Is the Zillow Zestimate accurate enough to use for pricing a home to sell?

The Zestimate provides a useful data point but should not be the basis for a listing price decision. On-market homes have a 2.4% median error, meaning half of estimates are within $12,000 on a $500K home — but the other half are off by more. A licensed agent’s comparative market analysis using actual recent local sales, a physical walkthrough of the property, and knowledge of current buyer demand will produce a more reliable pricing recommendation.

What is a property inspection waiver and should I take one?

A property inspection waiver (PIW) from Fannie Mae or Freddie Mac eliminates the requirement for a traditional appraisal on your loan, using the GSE’s AI model to validate value instead. For buyers, accepting a PIW saves $500-$700 in appraisal cost and removes appraisal contingency timing from the transaction. The tradeoff: you’re relying on the lender’s AI assessment rather than an independent appraiser. For standard properties in stable markets, PIWs are generally safe. For unique properties or rapidly changing markets, an independent appraisal may provide protection worth the cost.

Why does the Redfin Estimate differ from the Zillow Zestimate?

The two estimates use different data sources and model architectures. Redfin has direct MLS access and uses fresh transaction data updated multiple times daily. Zillow uses a broader data network but with varying freshness depending on data sharing agreements in each market. Both models weight different property attributes differently, which is why estimates can diverge significantly for the same property — and why checking both (plus Realtor.com) is better practice than relying on any single AVM.

Are AI appraisals accepted by all mortgage lenders?

Not universally. Fannie Mae and Freddie Mac’s PIW and desktop appraisal programs are available through lenders that sell loans to the GSEs — which covers most conventional mortgage transactions. FHA and VA loans still generally require traditional appraisals. Jumbo loans (above conforming loan limits) typically require traditional appraisals regardless of AVM confidence. Portfolio lenders (banks that hold loans rather than selling to GSEs) set their own appraisal policies.

Will AI replace human appraisers?

AI will replace human appraisers for a growing percentage of routine residential transactions in data-rich markets, as evidenced by expanding PIW usage and desktop appraisal programs. Full replacement is unlikely in the near term for complex properties, litigation-related appraisals, estate settlements, and commercial real estate, where physical inspection and professional judgment remain legally and practically necessary. The appraisal profession is likely to see volume reduction but not elimination, similar to other professional roles being reshaped by AI automation.

---

Sources

• Grokipedia: Automated Valuation Models — How AVMs Work and Their Accuracy Limits
• Fannie Mae: Collateral Underwriter and Desktop Appraisal Program Documentation
• Appraisal Institute: AI Impact on the Appraisal Profession — 2025 Research Report

---

Related Reading

Explore more: AI for Real Estate Professionals | Companies Trying to Replace Real Estate Agents with AI | AI for Commercial Real Estate | Best AI Tools for Real Estate Agents 2026 | AI for Real Estate Lead Generation | AI Browser Automations for Real Estate Listings

---

---

Stay current on AI developments in real estate and beyond. Subscribe to the Beginners in AI newsletter.

You May Also Like

• AI Tools Directory
• Best AI Tools for Beginners
• Best Free AI Apps
• Best AI Software
• Best AI Websites