The coal question for appraisersWill more efficient valuation technology shrink the appraisal profession, or grow it?

We can now check how Brynjolfsson’s candidates did, along with older cases that have run their full course. The results differ, and the differences line up with two things: whether a licensed person still signs the work, and whether people bought more of it as its price fell.

Radiology is the case most often cited for Jevons. Average radiologist pay reached $571,000 in 2025, according to Medscape figures reported by Fortune, and caseloads rose 25% from 2018 to early 2025.⁶ Two caveats apply: applicants fell from 2,014 in 2023 to 1,741 in 2026,¹¹ and researchers attribute 73% to 88% of projected imaging growth to an aging, growing population, a projection others have publicly disputed.^8,9,10 The narrower lesson is that imaging got less expensive, volume rose, and radiologists kept the licensed read.

Tellers show the second outcome. ATMs reduced tellers per urban branch from 20 to 13 between 1988 and 2004. Banks opened 43% more urban branches, so teller jobs still grew, about 2.0% a year after 2000.^12,13 Mobile banking then replaced the branch. The Bureau of Labor Statistics counted about 608,000 tellers in 2006 and 339,200 in 2025, and projects a further 13% decline by 2035.^158,14 James Bessen found the same arc in textiles, steel and autos: demand was elastic early in each industry’s life and inelastic later, so employment rose and then fell.²⁴

Accountants and bookkeepers split inside one field. Accountants and auditors grew from 39,000 in 1910 to 1,795,000 in 2000 as tax rules, regulation and managers’ demand for data expanded. Bookkeeping clerks declined after 1970 as software took over the entries.¹⁵ BLS projects bookkeeping clerks down 6% from 2025 to 2035, accountants up 5% and financial analysts up 7%.^16,17,157

Law split the same way. Software took over most document review, yet lawyers and judges grew from 272,000 in 1970 to 927,000 in 2000, and BLS projects lawyers up 5% to 2035.^18,15,20

Translators were on Brynjolfsson’s list, and they show what happens when the output itself is automated. Carl Benedikt Frey and Pedro Llanos-Paredes link the spread of Google Translate to about 28,000 fewer translator jobs than expected from 2010 to 2023. Translator employment grew about 3% while all jobs grew about 13%.²¹ BLS projections show the same decline for typists and travel agents, whose output software also took over.^22,23

Economists who study automation find the same rule. Jobs grow when tools take over tasks and people move to harder ones, and they shrink when the tool replaces the whole output.^25,26,27 Appraisal has both kinds of work in it. On a simple loan, ACE and waivers replaced the output itself, the value, much as machine translation replaced the translation. Measurement, data collection, research and drafting are tasks, and tools make them faster while the appraiser still signs the opinion. That is the radiology pattern, and it is the one complex property work could follow.

ACE and waivers work where a model has enough sales of similar homes. The examples below show what a model, or a rushed appraisal, has to answer when it does not. Each layer is a separate question with its own data source, and the answers change one another. The hillside parcel in the illustration is invented, but every card beside it describes a real place and a real rule. This is the complex property work from chapter 01, and it is where Brynjolfsson’s third test could pass: someone has a reason to pay for each answer, and better tools make each one more cost-efficient to produce.

Each layer comes from a different record: title and tract history, the municipal code for the parcel, a neighbor’s permits, rent comparables, ordinances and pending lawsuits, feasibility studies, a site visit, assessments and insurance quotes. A brokerage in Yorba Linda reports that the standard appraisal form has no fields for stalls, arena footing or trail access, and describes a barn costing about $150,000 that appraised at $40,000 to $70,000; that is one brokerage’s account, without sales data behind it.⁷⁰

Every property carries its own thumbprint

The layers also act on each other. Card 03 shows one case, where a neighbor’s ADU can make a horse use nonconforming. A short-stay rule changes what an added unit can earn, and so what the land under it is worth. Each home ends up with its own combination of rights, rules, units, views and costs, a thumbprint that the subject and every comparable sale carry separately. Two homes can match on bedrooms, square footage and lot size and still differ on several layers, so each comparison means working out how one combination moves value against another. A paired-sales adjustment for one layer assumes the others hold still, and on these properties they often do not. I see that as the next technical frontier in residential valuation, and the work where better tools would help most.

Models show their limits on the same properties

When few sales share a home’s thumbprint, a model has little to learn from. Automated valuation models (AVMs), the computer estimates behind tools like the Zestimate, struggle with the same properties. Zillow reports a median Zestimate error of about 1.94% for homes on the market and 7.06% for homes off it, meaning half of off-market estimates miss by more than 7%. The Close notes that once a home is listed, the list price becomes an important input, which helps explain the gap.⁸⁶ Condition shows the same limit. A 2024 Cityscape study of Atlanta and Memphis found mean AVM error 4.35 points higher for homes in poor condition than for homes in good condition.⁸⁹

Brynjolfsson’s third test asks whether buyers would order enough more of the work as its price falls. Nobody has measured that for complex property work, so this chapter looks for the next best signs: a rule that creates a new analysis requirement, a buyer who already pays for a report, and a buyer who settles for a cheaper substitute. The fire insurance case below shows the third, where an insurer’s rebuild estimate stands in for a replacement cost analysis. The research found ten such cases. I add an eleventh from my own experience, and I think it is the strongest.

Among the documented cases, ADU feasibility and rent is the clearest. Fannie Mae now lets borrowers count rent from a backyard unit toward the income they qualify with, up to 30% of it, on a one-unit home they live in, and its loan-approval software began accepting it on March 21, 2026. Neither announcement says how appraisers should report the unit or find rent comparables for it.^77,155 Freddie Mac’s guide requires at least one ADU comparable for a zoning-compliant unit and two for a non-compliant one, allows older and out-of-area sales, and says these assignments take additional work.⁷⁶ California permitted 30,354 ADUs in 2024, more than 26.6% of all homes permitted statewide, and in Los Angeles an ADU raises sale price 7% to 9%.^72,74

DSCR and short-term rental income is next. A DSCR loan is approved on the property’s rent rather than the borrower’s income, and it belongs to non-QM lending, the loans made outside the standard mortgage rules. Investor and DSCR loans rose from 22% of non-QM production in August 2022 to 35% in August 2026. Moody’s found that 30% of the DSCR programs it surveyed let borrowers use the higher of appraised rent and actual rent, with no cap.¹²⁷ AirDNA counted 1.73 million available short-term rental listings in November 2025.¹³⁰ On average, rental rules move value only a little: in Los Angeles County, house prices and rents fell about 2% to 3% just inside the borders of cities that restricted short-term rentals.^83,154 For a single property, though, the rules decide whether the income exists at all, as New York’s short-stay rule in chapter 05 showed.

Replacement cost has the largest gap and the least clear buyer. Of 675 households surveyed by United Policyholders after the 2025 Los Angeles fires, 69% of those with total losses said their coverage would not rebuild their homes, with average underinsurance of about $247 per square foot.¹³² Policies on the California FAIR Plan, the state’s insurer of last resort, reached 645,987 in September 2025, up 169% in four years.⁹¹ The San Francisco Fed found that distance from fires added only about 2% to the value of an average home from 2018 to 2021, so wildfire risk shows up more in insurance and carrying costs than in price.⁹²

Layered development matters most for housing and is the hardest to open up, because institutions control four of the six barriers in its way. Private and non-QM lenders set their own scope of work outside agency forms, which makes them the likeliest early buyers of deeper analysis.

Renovation and after-repair value (ARV) analysis has a large base: 297,045 flips in 2025, about 112,000 of them financed (computed). Gross returns were 25.5%, the lowest since 2008, and thinner margins limit what investors will pay.¹²⁶

The research behind this article found eleven barriers between lower-cost analysis and new demand for complex property work. Lenders, appraisal management companies (AMCs, the middlemen that order appraisals for lenders), the GSEs, regulators and legislatures control most of them. The one appraisers control is their own use of technology, and residential appraisal has been slow to adopt it for a long time.

When I asked the companies that make appraisal report software, they told me that fewer than half of their residential users use the mobile inspection app, even though many of those apps come free or bundled with the software. The rest sketch the house and take notes on paper, then type it all in again at the office. By my estimate, fewer than half subscribe to market analysis software that helps derive adjustments and build a solid work file. Neither figure comes from a published survey. AI use is wider. In a McKissock poll with no stated sample size, 40% said they use AI for narrative and 23% had not used it.⁵³ Menlo Ventures estimates that only about 3% of people who use AI pay for it,¹⁶² and I expect the share among appraisers is lower, which suggests most of that use is free chat tools that help with writing.

The standards no longer stand in the way. USPAP, the profession’s national standards, now includes Advisory Opinion 41, which adds no new requirements and sets out how to evaluate any tool, including AI. The National Association of Appraisers called generative AI another tool and noted that its members range from advanced users to non-users.^47,48

That gap is the opportunity, and it is Brynjolfsson’s first test, the one appraisers control. Appraisers who invest in technology and well-built systems can solve hard valuation problems in less time, which lowers what the work costs to produce. On the elastic side of the market, a lower cost brings in clients who could not justify the fee before, so those appraisers can grow their business by growing demand in their own markets.

Some of that growth needs no one else to move. Legal work, which I rate on my own judgment in chapter 08, faces no institutional barrier: executors and attorneys already order it, and a stronger report at a lower fee is worth ordering in more cases. Tool adoption blocks five of the eleven opportunities from chapter 08, more than any other barrier, and for legal work it is the only one. The other ten also need lenders, AMCs, the GSEs, regulators or legislatures to move. Brynjolfsson’s second test sits with them, because a lower cost becomes a lower price only if the saving reaches the client, and FHFA cites GAO that AMCs keep at least 30% of the fee.⁵⁷ Chapter 11 sorts those changes by who can make them.