Gary Neumann has appraised homes in Bucks County, Pennsylvania, for 31 years. He drives a 2019 F-150 with a measuring wheel in the bed and a laser distance meter in the center console. He charges $550 per appraisal, inspects two homes per day, and grosses about $220,000 in a good year. In a bad year, which is most years lately, he grosses closer to $140,000 before expenses, insurance, and continuing education fees. He is 62 years old, and his two apprentices both quit within eight months because the 2,000 hours of supervised fieldwork required for licensure paid less than driving for DoorDash, and the housing market was too slow to generate enough assignments to fill a full-time schedule.
Gary is not unusual, nor is his situation surprising to anyone who has watched this profession hollow out from the inside while the lending industry looked the other way.
On May 5, 2026, a company called ATTOM launched an AI-powered automated valuation model covering 98 million properties across the United States with a claimed median absolute percentage error of 2.9 percent. It processes a valuation in seconds, does not drive an F-150, does not need 2,000 hours of supervised training, does not need a license, and unlike Gary, it scales infinitely across every market in every state without hiring a single person.
(National Mortgage Professional)
A Workforce Walking Off a Cliff
In 2007, the United States had roughly 101,000 active licensed and certified residential appraisers. Today that number sits below 78,000, a decline of more than 23 percent in less than two decades, and the trajectory shows no sign of reversing because the economics that drove people out have only gotten worse while the regulatory barriers to entry remain exactly as onerous as they were a generation ago. National Mortgage Professional reports only 8,000 to 9,000 appraisal trainees remain nationwide, a pipeline so thin that retirements outpace new entrants by a widening margin every year.
Why? Because becoming an appraiser is one of the worst career propositions in real estate, a profession that demands thousands of hours of poorly compensated apprenticeship before granting the credential that allows you to compete for a shrinking pool of work. In most states, a trainee must complete 2,000 hours of supervised work under a licensed appraiser before sitting for the licensed residential credential, and 2,500 hours for the certified residential credential. During those hours, the trainee earns whatever the supervising appraiser decides to share, which is often nothing, or a nominal per-report split that works out to less than minimum wage when you account for windshield time between inspections. A HousingWire investigation found that 30-year-old appraisers are struggling to find stable employment in the current market, caught between a shrinking volume of work and a profession that never designed its pipeline for economic rationality.
Meanwhile, the people who do the hiring have decided they no longer need as many appraisals. In October 2024, FHFA expanded appraisal waiver eligibility from 80 percent loan-to-value to 90 percent for Fannie Mae and Freddie Mac loans, and to 97 percent LTV for inspection-based waivers. Fannie Mae announced that appraisals are "no longer the default option," and claimed the program has saved borrowers over $2.5 billion since 2020.
Read that sentence again, because the framing matters: not "appraisals are sometimes unnecessary," not "we've improved the process," but appraisals are no longer the default protection standing between a borrower and a catastrophic overvaluation.
What 2.9 Percent Means in Dollars
ATTOM's new AI valuation model reports a 2.9 percent median absolute percentage error across 98 million US properties, with more than 80 percent of valuations falling within 10 percent of actual sale price. On its face, 2.9 percent sounds precise, the kind of number that ends arguments in conference rooms where nobody in attendance is personally buying a house with their own savings. CEO Rob Barber called it "a complete re-architecture of the AVM using advanced AI," built on 30 years of time-adjusted transaction history rather than the comparable-sales approach that traditional AVMs depend on.
But percentages obscure what matters to the person writing the check. A 2.9 percent median error means half of all valuations miss by more than 2.9 percent. And in dollar terms, those misses look different depending on where you live:
| Home Value | 2.9% Error | 10% Error (tail) | Human Appraisal Cost |
|---|---|---|---|
| $350,000 | $10,150 | $35,000 | $500–$600 |
| $500,000 | $14,500 | $50,000 | $500–$700 |
| $750,000 | $21,750 | $75,000 | $600–$900 |
| $1,200,000 | $34,800 | $120,000 | $800–$1,200 |
| $2,000,000 | $58,000 | $200,000 | $1,000–$1,500 |
A human appraisal on a $750,000 home costs roughly $600 to $900. An AVM error of 2.9 percent on that same home is $21,750. At the tail end, where 20 percent of valuations fall outside the 10 percent window, you are looking at a $75,000 miss. The insurance policy, the human walking through the house with a tape measure and a trained eye, costs less than half a percent of the error it might catch.
Nobody in the lending industry frames the arithmetic this way, because cost savings for borrowers get celebrated in press releases while error budgets stay buried in technical white papers that borrowers never read.
What an Algorithm Cannot Smell
Gary Neumann once appraised a four-bedroom colonial in Doylestown where the listing photos showed gleaming hardwood floors, fresh paint, and a kitchen renovation that added $40,000 to the asking price. He pulled into the driveway and noticed the soil was wet along the foundation on a day with no rain. Inside, the basement had a faint musty smell that the sellers had covered with plug-in air fresheners. He pulled the carpet back from one corner and found black mold spreading across the subfloor, fed by a cracked foundation wall that channeled groundwater into the finished basement during every storm.
His appraisal came in $55,000 below the contract price. Saved the deal. Saved the buyer from six figures of remediation they did not know they were about to inherit. Without that walk-through, an AVM would have pulled comparable sales from the neighborhood, noted the kitchen renovation, and assigned a value that assumed the foundation was intact, because foundation condition does not exist in any public dataset that feeds an automated valuation model, and no algorithm trained on tax records and deed transfers can detect the smell of mold behind drywall.
This is not an edge case. It is the norm. Illegal additions, unpermitted conversions, deferred maintenance, environmental contamination from a neighboring property, drainage problems visible only at ground level during specific weather conditions: these are the things that separate a building from a data point, and they represent a category of risk that no amount of historical transaction data can capture because they are not recorded anywhere until a human being stands on the property and looks with their own eyes at what the spreadsheet cannot contain.
AEI Housing Center data shows GSE appraisal waivers now account for roughly 23 percent of all GSE loan valuations, elevated from historical norms though still below the pandemic-era peak of nearly 50 percent. The Working RE 2026 market update reports that waivers cluster around lower-risk, standardized transactions, particularly refinances and loans with low LTV ratios. But "lower-risk" is defined by the data available to the algorithm, and the algorithm's data does not include mold, cracks, or the smell of something wrong that makes a 62-year-old appraiser crouch down and pull back the carpet.
Strongest Counterargument
AVMs do not need to be perfect. They need to be better, on average, than an overworked, aging human workforce operating in a constrained market with declining volume and deteriorating economics. A 2019 Fannie Mae study found that traditional appraisals had their own significant error rates, with roughly 10 to 15 percent of appraisals requiring revision after underwriter review. Human appraisers bring bias: NAR and FHFA research has documented persistent racial disparities in home valuations, with homes in majority-Black neighborhoods appraised 12 to 23 percent below comparable homes in white neighborhoods. An AVM, whatever its flaws, applies the same mathematical model to every property regardless of who lives there.
And the workforce crisis is real. If the pipeline cannot produce enough new appraisers to replace the ones retiring, and if remaining appraisers are concentrated in high-volume markets while rural and exurban areas go underserved, then AVMs are not replacing humans by choice. They are filling a void that the profession's own economics created. Demanding a human appraisal on every transaction is a position that requires enough humans to do the work, and the numbers say there are not enough, with the gap widening every year.
This argument is strong, and parts of it are true. But "better than a declining workforce on average" is a different claim than "adequate protection for the individual borrower." A median error of 2.9 percent means half of all valuations are worse than 2.9 percent. For the borrower in the worse half, the algorithm's overall accuracy is irrelevant. What matters is whether their specific home, with its specific foundation crack or illegal addition or drainage problem, got caught. Averages comfort lenders. Outliers bankrupt homeowners, and the homeowner never gets to choose which side of the distribution they land on.
What You Should Do
If you are buying a home and your lender offers an appraisal waiver, understand what you are accepting. You are trading a $500 to $1,200 inspection by a licensed professional for an algorithmic estimate built on public data that cannot see inside the house. On a $500,000 purchase, the appraisal costs 0.1 to 0.14 percent of the transaction. The median AVM error is 20 to 30 times larger than the cost of the human check.
Request a full appraisal even when one is not required. Pay for it yourself if the lender will not. This is especially critical for homes with any of the following characteristics: older than 30 years, renovated without permits, located in a flood zone or near environmental hazards, priced significantly above or below neighborhood comps, or new construction where comparable sales data is thin.
If you are selling, know that an AVM-based transaction means no one with a license and a legal obligation to assess market value will physically examine your property. Any condition issue you have disclosed, or failed to disclose, sits outside the valuation model entirely. Your legal exposure does not change just because the lender decided a human inspection was unnecessary.
If you are building a new home, AVMs perform worst on new construction because the model depends on historical transaction data that does not exist for a house that was not there six months ago. Insist on a traditional appraisal. Construction lenders generally require one anyway, but if yours does not, the gap between what the model thinks your house is worth and what it actually cost to build can be enormous.
Limitations
ATTOM's 2.9 percent median error figure is self-reported from internal out-of-sample testing across a decade of residential sales. No independent third-party audit of this claim has been published. AEI waiver data tracks GSE-backed loans only and does not capture private portfolio lending, which operates under different valuation standards. Appraiser workforce counts vary by source and methodology: National Mortgage Professional, HousingWire, and state licensing boards all report different totals because they count active versus inactive credentials differently. The error cost table above uses the median, but actual error distributions are not symmetric and may skew higher in low-liquidity or rapidly appreciating markets. Regional variation in both AVM accuracy and appraiser availability is massive: the national average obscures the reality that some counties have three working appraisers and some have three hundred. We do not have data on how often AVM valuation errors result in measurable borrower losses post-closing, because that data would require tracking loan performance against initial valuation across millions of transactions, and no public dataset currently enables that analysis.