An automated valuation model is not a faster, cheaper appraiser. It is a price predictor, which is a different thing. A working map of which parts of the appraisal a machine can absorb, which it cannot, and where the line moves over the next five years.
The defining mistake in the public conversation about automated valuation models is that it treats them as a substitute for the appraiser. They are not. They are a substitute for price prediction, and a price prediction is not an appraisal.
A price prediction is a statistical estimate of what a property would trade for, expressed as a number with a confidence band, derived from prior sales of properties a model treats as comparable. An appraisal is a defensible opinion of value that answers a specific question, for a specific user, under a specific definition of value, as of a specific date, and carries the reasoning to survive challenge. The first is an output. The second is an informed and educated opinion.
Almost every claim that AVMs are "close to replacing appraisers" rests on blurring those two. A Zestimate is good at predicting what a standard house in a high-volume subdivision will sell for next month, given the last two years of sales around it. That is useful, but it is not what an appraiser does. An appraiser first decides which question is being answered (market value, insurable value, assessed value, value in use, going-concern value), under what conditions (cash equivalent, subject to financing, as if vacant), and for which user (a lender, a court, an assessor, an owner) and then defends an opinion against a counterparty paid to break it. The first half of that work is not statistical. It is definitional, contextual, and adversarial.
This distinction is not academic. It explains why AVM accuracy benchmarks improve every year while AVMs absorb a smaller share of real appraisal work than the headlines imply. It explains why the federal AVM Quality Control rule, effective October 1, 2025, governs only AVMs used in credit decisions on a consumer's principal dwelling and leaves the rest of the real property universe outside its scope. And it explains why commercial AVMs, despite serious data and capital behind them, have made only modest inroads into the work an MAI does for a contested matter, even as consumer AVMs dominate the kitchen-table conversation about home values.
This issue maps the field. Part One surveys the AVM products you already recognize. Part Two walks the entire appraisal process, task by task, and rates each one for how much a machine can already do it. Part Three draws out what the map means for the people doing the work. The piece commits to a forecast that is specific through 2029 and directional through 2031. Issue 03 takes up the bifurcation thesis the map points toward; Issue 04 will commit to dated, on-the-record predictions.
There is no single thing called an AVM. There are at least five product categories, separated by who pays for them, what data they consume, and what accuracy they claim. The categories matter because they carry different consequences for appraisers, and lumping them together is how most commentary produces consistently wrong forecasts.
Zillow Zestimate, Redfin Estimate, Realtor.com, the bank home-value tools. These are the AVMs nearly everyone has seen. Both leaders cover roughly the entire housing stock, and both are free at the point of use because the business is advertising and lead generation, not the sale of the valuation. The accuracy numbers reward a close read. Zillow reports a national median error around 1.9% for homes that are actively listed and around 7% for off-market homes; Redfin's figures are similar. The on-market number looks impressive until you see why: once a home is listed, the model has the list price, which is itself a market-tested estimate produced by a working agent. The on-market AVM is, in part, copying the agent's homework. The off-market number, which is the situation in which an independent estimate is actually needed, is where these tools fall apart, and the fall is structural, not a tuning problem.
A peer-reviewed study in Advances in Consumer Research (September 2025) compared Zestimates against New York City's own 2024 market-value figures for a clean sample of properties and found the two disagreed by a median of roughly 17.5%. The study, and the coverage of it, read that as Zestimate error. Anyone who understands market value and works with assessed figures knows better than to treat the city's number as ground truth; what was actually measured is the distance between two automated systems, either of which can be wrong on any given property. That is the sharper point. When two machines sit 17.5% apart on the same house, the question that matters is which one is right, and neither machine can answer it. Answering it, with evidence, on the record, is the appraisal.
Cotality (formerly CoreLogic, rebranded March 2025), ICE Mortgage Technology, HouseCanary, Veros, Clear Capital. This is the category appraisers underestimate. Lending-grade AVMs are sold to banks and AMCs, are tested with statistical error metrics rather than a single flattering median, and now fall under the federal Quality Control rule. The honest tell is in how each grade reports itself. Consumer AVMs quote a median error, which hides the big misses. Lending-grade AVMs report a confidence score and a forecast standard deviation, an estimate of how wrong the number might be, precisely because their buyers need to know where not to trust the model.
Two structural points matter more than the marketing. First, lending-grade AVMs do not value every property; they return a value only where confidence clears a threshold, and the properties they decline are systematically the complex ones: odd lots, atypical improvements, recent renovations the data cannot see, thin markets, properties with no recent transfer. The properties the model refuses to price are exactly the properties where appraiser judgment is worth the most. Second, the concession is on the record. Arguing to federal regulators for broader AVM use, HouseCanary estimated that roughly 40% of homes can be served by a fully automated valuation, another 45% by an AVM plus an inspection, and about 15% still require a traditional appraiser. Even the case for the technology concedes a hard 15%, and that 15% is the complex tail.
HouseCanary, Cotality, ATTOM. A "pre-list" valuation is one generated before a property is listed for sale, which means there is no list price for the model to anchor to, so it is a more honest test of the model than the flattering on-market number. This category sells to investors and to iBuyers, the instant-offer buyers such as Opendoor that make an immediate cash bid on a home, lightly improve it, and resell, all on the strength of an AVM. (When Zillow ran its own iBuyer and the model mispriced its inventory in 2021, the unit was shut down. The cautionary tale is built into the category.) Pre-list is the residential AVM that competes most directly with a real appraisal: when a lender swaps a pre-list AVM in for a full report, that is the substitution event. When a homeowner glances at a Zestimate, that is a marketing event. Trade press conflates the two constantly.
Green Street, CompStak, Cotality Commercial, RealPage, Moody's. This category is different in kind, and the difference is the heart of where the profession is going. Commercial AVMs exist: Green Street publishes single-digit error rates for individual properties and tighter rates for portfolios of twenty-five or more, though those claims hold mainly for institutional-grade assets in active markets. CompStak runs a crowdsourced lease- and sale-comp database covering millions of properties, built from a network of brokers and appraisers. RealPage provides multifamily data and analytics.
The limit on commercial AVMs is not the algorithm. It is the data the market produces. Commercial transactions are sparse, heterogeneous, and often private in ways residential transactions are not. A regional mall does not have ten close comparables that traded in the last six months. A net-leased drugstore's value turns on its specific lease structure, tenant credit, and remaining term, none of which the public record holds. A going-concern hotel is valued on operating performance the public record never sees. A better model cannot manufacture transactions that did not happen, so the commercial AVM hits a low limit on accuracy no matter how good the math gets, while the residential AVM keeps climbing because the data underneath it is deep. The residential AVM and the commercial AVM are not the same product at two stages of maturity. They are different products with different limits, and the commercial limit is lower because the data is thinner.
Class Valuation, Solidifi, AppraisalWorks, Reggora. Strictly these are not AVMs, but they are the product class displacing the most traditional residential appraisals, so they belong here. In a hybrid appraisal a third party collects the property data on site and a licensed appraiser does the analysis from the desk. Desktop appraisals became a permanent GSE option in March 2022; the hybrid path has grown since 2023 under Fannie Mae's Value Acceptance plus Property Data framework. The reason this matters for the map:
Hybrid and desktop products are not appraisal automation. They are appraiser-supervised automation. A licensed appraiser still signs the report and still carries the liability. The labor input shrinks; the appraiser-of-record does not disappear. This is the operating model the residential side of the profession is quietly being rebuilt into, and it changes what the license is actually being paid for: less the production, more the signature.
Fannie Mae's Collateral Underwriter and Freddie Mac's Loan Collateral Advisor are routinely miscalled AVMs. They are not. They are appraisal review systems: they take a finished appraisal, compare it against the agency's own database, and return a risk score that decides whether the report earns representation-and-warranty relief or needs a human second look. They compete with appraisers in one narrow place, and it is an important one. The senior reviewer who a decade ago read every report now reads only the ones the system flags. The middle of the appraisal labor market, the review and quality-control layer, has been automated more quietly and more thoroughly than the production layer ever was. And every report filed through the agency portal becomes training data for the next version of the system. UAD 3.6, mandatory November 2, 2026, accelerates this because the new dataset is structured for machine reading in a way the old forms were not. The residue of appraiser work is being used, at scale, to build the systems that grade appraiser work.
| Task | Machine-capable today | By ~2029 | Offshored already | What stays the appraiser's |
|---|---|---|---|---|
| A · Define the assignment | ||||
| Scope and problem definition | ●○○○○ | ●●○○○ | No | Deciding which value question is being asked, for whom, and under what definition of value. |
| B · Data and description | ||||
| Subject identification and public records | ●●●●○ | ●●●●● | Yes | Catching when the public record is simply wrong. |
| Site description (zoning, utilities, easements, flood) | ●●●○○ | ●●●●○ | Partial | Judging how zoning, access, and easements actually move value. |
| Improvements and property-data collection | ●●●○○ | ●●●●○ | Partial | Condition and functional-utility judgment a photo set cannot carry. |
| Neighborhood and market-area description | ●●●○○ | ●●●●○ | Yes | What the trend means for this specific property. |
| C · Analysis | ||||
| Market analysis (supply, demand, exposure time) | ●●○○○ | ●●●○○ | Partial | Reading the current market, especially at a turn. |
| Highest and best use | ●○○○○ | ●●○○○ | No | The entire analysis. |
| D · The approaches to value | ||||
| Sales comparison - comp identification | ●●●●○ | ●●●●● | Yes | Which comps a reviewer or a tribunal will accept. |
| Sales comparison - adjustments | ●●●○○ | ●●●●○ | Partial | Defending the adjustment for this property, not the average one. |
| Cost approach (cost-new and depreciation) | ●●●○○ | ●●●●○ | Partial | Functional and external obsolescence. |
| Income - direct capitalization | ●●●○○ | ●●●●○ | Partial | Cap-rate selection. |
| Income - discounted cash flow | ●●○○○ | ●●●○○ | Partial | Defending the assumptions behind the model. |
| E · Conclusion and delivery | ||||
| Reconciliation and final value opinion | ●○○○○ | ●●○○○ | No | Defending the weighting. |
| Narrative and report production | ●●●●○ | ●●●●● | Yes | The reasoning the narrative expresses; the signature and the liability. |
| Appraisal review (CU / LCA, secondary review) | ●●●●○ | ●●●●● | Partial | Adjudicating the genuinely hard files the system flags. |
| F · Defense | ||||
| Expert testimony and litigation support | ●○○○○ | ●○○○○ | No | All of it. Reserved, by statute, for a credentialed human. |
Figure 01 walks the full assignment, not a cherry-picked task. Read it and a shape appears, and the shape is the whole argument.
Two clusters, with a hinge in the middle. Defining the assignment, highest and best use, reconciliation, and testimony sit at the floor, near 1, and barely move by 2029. These are the judgment tasks, and a machine cannot do them because they are not prediction problems. Subject data, descriptions, comp identification, and report production sit at the top, near 4 and 5. These are the production tasks, and they are largely done already. The approaches to value sit in between, and the reason is precise: the calculation automates completely while the selection does not. A model will compute a value off a cap rate instantly; choosing the right cap rate for this asset, and defending it, is the job.
The offshoring tell. Look at the offshored column and then back at the capability columns. Every task already shipped offshore is also a high-automation task. That is not a coincidence; it is the same fact twice. For more than a decade, production work has been sent to India and the Philippines at a fraction of domestic cost: lease abstraction at roughly five to twenty-five dollars a lease, ARGUS modeling and underwriting support, comp data entry, narrative typing, and increasingly the data-collection and back-office roles themselves. A task cheap enough to send to Manila was already codified enough for a model to learn. Offshoring was the dress rehearsal; AI is the performance. And the tasks that never offshored, the ones no firm trusted to a remote contractor reading from a script, are the same ones the model cannot do now: highest and best use, cap-rate selection, reconciliation, testimony. The boundary that protected those tasks from offshoring is the same boundary protecting them from automation.
The single most important row is the last one. Expert testimony does not move, this decade or next, because a tribunal cannot cross-examine an algorithm. It is the profession's strongest protection, and it is the one the profession itself talks about least.
The commercial overlay. Now lay a typical assignment over the map. A refinance appraisal on a standard suburban house spends almost all its hours in the production rows, the rows getting eaten. A commercial assignment, a struggling shopping center, a value-add multifamily deal, a contested assessment, spends most of its hours in the durable rows: the income approach selection, highest and best use, reconciliation, and, often enough, testimony. That is the entire reason the two sides of the profession are pulling apart, and it is the subject of the next issue.
Three conclusions follow from the picture, and each one lands differently depending on where you sit, so I have split each into what it means inside a large national or AMC-fed shop and what it means for a small or solo independent.
"AVMs replace appraisers" is wrong at the level of the unit. AVMs absorb tasks inside the workflow and leave others alone. What happens to you depends entirely on which tasks fill your billable hours, and most appraisers, asked to guess, get their own mix wrong.
National / AMC shop: throughput per appraiser is about to rise sharply, and the organization will expect more files per head, not the same files at higher margin. The analyst and trainee rungs that did comp pulls, data entry, and first-draft narrative are the first line item cut, which solves a cost problem this year and creates a pipeline problem in ten.
Small / solo: the production tasks you bill by the hour are the exact tasks getting cheaper. If your fee is built on producing the report, you are about to be competing with a fifteen-dollar offshore seat and a model that drafts the narrative. Reprice around the judgment, or get repriced by it.
The data conditions that let residential AVMs work, frequent standardized public sales, do not exist in commercial markets and will not. No algorithmic gain closes a gap that is structural rather than computational. This is the asymmetry Issue 03 builds on.
National / AMC shop: the residential desk industrializes and competes on cost and turn time; the commercial and complex desks stay human and gain pricing power. Staff, train, and invest as if those are two different businesses, because they are becoming two different businesses.
Small / solo: if you are residential-only, the squeeze is real and close. If you do commercial, litigation, eminent domain, or special-purpose work, your protection is the data scarcity itself, not how hard you work. Lean into the assignments a model structurally cannot price.
The federal AVM rule, effective October 2025, applies to AVMs used on a principal dwelling. It does not reach commercial AVMs. The commercial de minimis appraisal threshold has sat at $500,000 since April 2018. The residential side is being industrialized inside a regulated AVM-and-hybrid market; the commercial side, for now, sits outside it.
Everyone: do not assume that line holds forever, and do not panic that it has moved when it has not. The waiver expansion and the AVM rule are residential events. The commercial threshold has not changed in seven years. Watch the threshold, because a move there would be the real signal, and treat anyone selling commercial-appraisal panic today as ahead of the evidence.
Across all three, the durable play is the same: stop letting your judgment evaporate at the end of each day. The appraisers who do well this decade turn their own market knowledge into an asset that compounds, a private comp record, their own adjustment logic, a working model of the assessors, lenders, and panels they deal with, so the thing a machine cannot replicate gets deeper every year instead of starting over each Monday. You do not need a new subscription. You need to treat your own situated knowledge as the product and the tools as the leverage on it.
Specific through 2029, directional through 2031. Issue 04 will commit to dated forecasts I expect to be held to; this is the field those forecasts sit inside.
This issue avoids telling you what to do; Issues 04 and 05 will get specific on forecasts and on the paths off the appraisal chair. What is fair to say here is the reasoning the map asks of you: write down where your billable hours actually go, then mark each task against Figure 01. The hours sitting in the production rows are getting cheaper whether you like it or not. The hours sitting in the judgment rows are getting more valuable. Most appraisers discover their hours are not distributed the way they assumed, and the ones who do well are simply the ones who move their time, and their fee, toward the rows a machine cannot reach. Issue 03 takes up the bifurcation thesis in full.
Product details and accuracy figures draw on the published accuracy pages of Zillow and Redfin, the peer-reviewed Zestimate benchmarking study in Advances in Consumer Research (September 2025), provider and regulator documentation from Cotality, ICE Mortgage Technology, HouseCanary, Veros, Clear Capital, Green Street, and CompStak, the American Enterprise Institute's GSE waiver prevalence reports (most recent September 2025), Fannie Mae and Freddie Mac UAD 3.6 materials, the WorkingRE 2026 market update on appraisal volume and property-data collections, the interagency Quality Control Standards for Automated Valuation Models (Federal Register, August 2024; effective October 1, 2025), and public materials from commercial-real-estate offshore-services providers on the scope and pricing of outsourced production work.
The dot ratings in Figure 01 are my own assessment, not a measured benchmark, and are offered as a considered opinion. Where I have forecast or taken a position, I have said so. Where the data is thin, I have said that too. I welcome correction from readers who know a specific corner of this better than I do; my email is in the masthead.
- J.R.C.