1. Not a Number but a Confidence Interval: The Reshaping of Valuation

Same House, Different Numbers

In 2021, a three-bedroom single-family house in a suburb of Austin, Texas, produced two answers.

One was given by a person. An appraiser with twenty years in the local market walked the house, checked the condition of the roof, the sunlight in the backyard, the height of the fence against the neighbor’s lot, compared it to three recently sold comparable homes, and attached a price tag. The other was given by an algorithm. Zillow’s AVM (Automated Valuation Model) — the “Zestimate” — ran thousands of variables at once and produced a number in seconds. Square footage, orientation, floor level, school district, recent local sales history, even whether a pool showed up in aerial photography.

The two numbers differed. That’s not surprising — human appraisers routinely disagree with each other too. What was genuinely surprising happened next. Zillow trusted its own algorithm’s answer enough to start buying houses directly at that price. The business, called “iBuying,” rested on a simple and audacious proposition: if AI prices it accurately, why shouldn’t we buy and sell at that price ourselves? In the third quarter of 2021 alone, Zillow lost $421 million on this business. Before the year was out, Zillow Offers shut down entirely, and the company laid off a large share of its workforce. The full-year loss figure and the CEO’s own on-record confession are covered in detail in Chapter 8.¹ Around the same period, iBuying experiments in Europe and Asia scaled back or quietly wound down — several UK startups, and the in-house buying operations of a few major Chinese brokerage platforms among them. This wasn’t one company’s misjudgment. The mistake — confusing “an accurate price tag” with “market liquidity to actually buy and sell at that price” — repeated itself across multiple continents.

The Zestimate’s accuracy itself wasn’t bad. Against homes actually on the market, the error rate ran around 2%, tighter than most human appraisers manage.¹ The company still burned through hundreds of millions of dollars in a matter of quarters.

The reason lies in a distinction that runs through this entire chapter. “Knowing what something is worth” and “having the nerve and the timing to actually buy and sell at that price” are entirely different problems. AI is good at the former. The latter remains, and likely will remain for a long time, a domain tangled up with market liquidity, timing, and human judgment. The real shift in real estate valuation over the past five years isn’t the simple story that “AI got more accurate than people.” The question itself changed. Not “what is this building worth,” but “what does the probability distribution of what this building is worth look like.”

From an Age of Precision to an Age of Confidence Intervals

William Poorvu, whose 1999 book The Real Estate Game is a classic of real estate investing, makes an argument worth revisiting: real estate is judged, most of the time, not by an elaborate spreadsheet but by “back-of-the-envelope analysis.” Even as he taught real estate at Harvard, he noted that while academia drifted toward ever more elaborate models, practitioners filter opportunities quickly using just a handful of core ratios — net operating income (NOI), return on equity (ROE), purchase price against replacement cost. The calculation he scrawled on a yellow legal pad while taking a broker’s phone call is, even now, striking in its simplicity: divide the purchase price by square footage to get a price per square foot, compare that against the cost of building new to gauge how safe the price is, then calculate what percentage of invested equity the rent throws off after subtracting operating costs and debt service. He called this analogy “grandmother’s chicken soup” — a result that comes out roughly right every time, with no precise recipe required.

Twenty-five years on, it’s worth flipping that analogy. Real estate valuation in the AI era hasn’t arrived at grandmother’s touch — it has arrived at the opposite extreme. An ultra-precise machine that quantifies thousands of variables, trains on millions of historical transactions, and manages its error rate to the decimal point. And yet what this precision has handed practitioners isn’t “one more accurate number.” Ironically, it’s the question of how much to trust that number in the first place.

The reason is simple. An AVM’s error rate isn’t uniform. In liquid markets with standardized assets, the error rate can fall to 2–3%. A three-bedroom suburban home in the United States is the classic case — the data is abundant.² Move to markets with thin transaction volume, distinctive commercial assets, or an emerging asset class just coming into existence (data centers, discussed later, are the prime example), and the error rate easily exceeds 10%. Data scarcity is the cause. The same model, the same company, produces numbers where “this one is close to gospel” and “this one is a rough guide at best” coexist side by side. An AVM is essentially a comparable-sales appraisal run at extreme speed. Where a person forms an impression from five or six comparables during a single phone call, the machine scans thousands of comparables in the same instant. But in a market with nothing to scan — no transaction record — even the fastest machine has nothing to work with.

One sensibility has quietly, but firmly, taken root in the industry over the past five years. The measure of a good valuation model is no longer “how accurate a number does it produce” but “does it tell you how confident to be in that number.” The confidence interval — a concept borrowed from statistics — has become part of the working vocabulary of real estate practice. Instead of a single flat declaration — “this building is worth $1 million” — an answer like “this building is worth, with 90% confidence, between $950,000 and $1,050,000, though this range rests on just three recent comparable sales and the sample is thin” is now the answer people trust. Not precision but honesty has become the new standard.

The Eight-Minute Underwrite, and Its Price

Nowhere has this shift shown up more dramatically than in commercial real estate loan underwriting. Traditionally, a commercial real estate loan was a bureaucratic process — 30 to 45 days from paperwork to approval, passing through the hands of bankers and analysts multiple times. Reading lease agreements one by one, transcribing tenant-by-tenant rent rolls into spreadsheets, verifying the numbers on operating statements against prior years, building cash-flow models by hand. Just as the back-of-the-envelope calculation described above wrapped up within a single phone call, the underlying logic of this work is itself simple. It’s the scale that’s different. The same arithmetic an individual investor uses to screen a single building, a bank has to repeat on hundreds of deals every day.

Over the past five years, reports of AI underwriting adoption at banks have piled up: 50–75% reduction in underwriting time, cost savings of up to 20%.³ For deals with clean, well-defined terms, cases have emerged where AI reads the documents, builds the cash-flow model automatically, and finishes the underwrite in eight minutes. Work that used to take an analyst several weeks, a machine now handles in the time it takes to finish a cup of coffee.

That speed comes at a price: a growing number of “black boxes” where it’s hard to explain why a given number came out. A human analyst can articulate the reasoning behind a judgment — “I marked this tenant’s renewal probability down because their recent sales have been shaky.” When an AI model reaches the same conclusion, it’s far harder to tell whether that conclusion really traces to the tenant’s sales figures, or to some bias that slipped into the training data by accident. Regulation has begun catching up too — late, but fast — and the point of origin and the shape of the response differ by country. In the United States, the Interagency Rule on AVMs, which took effect in 2024, mandates confidence management for valuation models, safeguards against data manipulation, and conflict-of-interest protections.⁴ In Europe, the body that sets appraisal-industry standards, through its 2025 revised standard, cemented the principle that “an AVM cannot substitute for a formal appraisal on its own — it must be combined with a site inspection and expert judgment.” Five years ago, competitive advantage meant “how fast can you adopt AI.” Today, it means “how defensibly can you explain that AI’s judgment” — a new competitive axis that cuts across every continent.

Thirty People Buy a Giant

One more event captures this five-year arc. In 2021, JLL — one of the largest real estate services firms in the world — acquired Skyline AI, a data startup based in Israel and New York with a headcount of just over thirty. What exactly this company built, and how it was used after the acquisition, is covered in detail in Chapter 2.⁵

The question this acquisition raised was simple and cutting: can a small team of data scientists predict an asset’s future more accurately than an organization built over decades of brokers and appraisers? The industry’s answer was the acquisition itself. This wasn’t an isolated way of acquiring capability — over the following five years, major brokerages and asset managers repeatedly chose to buy rather than build. Asia took a somewhat different approach. A large Chinese real estate brokerage platform, rather than acquiring, built a large in-house data-engineering team and developed its own AI valuation-assistance features in-house — including image-recognition tools to read a listing’s condition from photographs. Buy versus build diverged, but the conclusion — that AI capability has become a core competitive asset for real estate services firms — held across continents. JLL’s own supporting statistics (number of AI use cases, institutional pilot-adoption rates) are also covered in Chapter 2. Valuation is no longer the experiment of a handful of early adopters. It has become standard industry workflow.

An Algorithm Built Collusion It Didn’t Even Know About

Not everything in this five-year arc is a bright story. The most dramatic reversal is the case of RealPage, a rent-setting software. The software pooled confidential pricing information from multiple landlords and recommended an “optimal rent” for each building. The logic sounded reasonable — the same premise as an AVM: more data yields a more accurate price.

The problem was that multiple landlords in the same city used this software simultaneously. No individual landlord ever discussed collusion with another. But because everyone fed their own building’s pricing data into the same algorithm, and the algorithm synthesized that information into the same whispered conclusion for each of them — “everyone else is using the same software too, so the market can bear a coordinated rent increase” — a structure emerged that functioned like collusion without anyone explicitly agreeing to collude. According to the U.S. Department of Justice’s complaint, one landlord began raising rents within a week of adopting the software and had raised them more than 25% within eleven months.⁶ The U.S. rent-algorithm lawsuit effectively charged this as facilitating collusion, and in November 2025 the software company settled with the Justice Department — without admitting wrongdoing — agreeing to discontinue the rent-recommendation feature that used confidential competitive information.⁶

This case matters beyond any single company or country. It’s the first major legal rebuttal of the narrative that “AI makes markets more efficient.” A similar concern has already surfaced across the Atlantic — not in real estate but in gasoline retail in Europe, where a wave of stations switching to the same pricing algorithm was followed by a joint rise in margins, drawing the attention of competition authorities. Different industry, same structure. Set aside any individual country’s statutes and look at the underlying principle: the moment multiple competitors run the same pricing algorithm simultaneously, that algorithm sits on the line between an efficiency tool and a collusion mechanism. Each party behaves rationally on its own, yet the outcome functions collectively like a cartel — a new species of market failure. This is the most unexpected lesson of AI valuation’s first five years, and it isn’t confined to any one country.

The Appraiser Isn’t Disappearing — They’re Relocating

So what happens, at the end of all this, to the professions of appraiser and broker? The five years of data point to an answer that’s “relocation,” not “extinction.”

There’s real work AI has clearly taken over: pulling and listing comparables, calculating prices for standardized assets, reading rent rolls and building cash-flow models. These repetitive, quantitative tasks are now faster for machines than people — and, for standard assets, more accurate. That’s exactly why major brokerages bought companies like Skyline AI whole instead of building in-house. Repetitive-calculation capability is no longer a blank a human has to fill by hand; it’s a component you buy off the shelf.

For precisely that reason, the character of the work left to people has changed. The question an appraiser now wrestles with isn’t “what did a comparable building recently sell for” — a machine already answers that in seconds. Instead, people have moved toward the questions data still doesn’t capture: How do you value an asset with no meaningful comparables for the model to draw on? Is there a shift brewing in this neighborhood’s tenant composition that the data hasn’t picked up yet? Is the seller concealing a desperate situation? The European appraisal-standards body’s insistence that “an AVM must be combined with site inspection and expert judgment” sits in the same vein — it institutionalizes a structure where the machine’s answer and the human’s answer sit side by side, and a person points out where they diverge.

The broker’s role has shifted in a similar way. Telling someone the going rate has lost its scarcity value now that anyone can check it on a phone in seconds. What earns value instead is reading the other side’s real situation across the negotiating table, and navigating variables that don’t reduce to numbers — regulation, community relations, partnerships. The new division of labor these five years have produced is this: the machine calculates “what is it worth” where data is abundant; the person judges “should this number be trusted” where data is thin or was never reducible to numbers in the first place. Not a lost job. A relocated one.

What Comes After the Confidence Interval

So which parts of this five-year arc will still hold true three years from now, or in a different country? Three questions stand out.

First, the structural limit that AI valuation accuracy scales directly with data density isn’t going away. For standardized assets in liquid markets, AI will keep out-answering people with finer-grained precision. For rare, specialized assets and newly emerging asset classes, human experience and intuition will keep the edge. That boundary line will keep shifting as AI models improve, but it won’t disappear entirely.

Second, “prediction” and “the decision to actually deploy capital on the strength of that prediction” will remain separate. Zillow lost over $400 million not because its model was wrong, but because turning an accurate prediction into an actual transaction ran into market liquidity and timing — variables a statistical model struggles to handle. That gap is unlikely to close no matter how sophisticated AI becomes, because real estate isn’t an asset that sells the moment you want to sell it.

Third, the structural risk that arises when multiple competitors use the same tool simultaneously will keep following this industry regardless of how regulation evolves. Valuation tools have already crossed over from an aid to individual judgment into infrastructure that directly shapes price formation across the whole market. Who oversees this infrastructure, and how, is a question that has only just begun.

What the past five years have left behind isn’t a simple win-lose story of “AI replaced people.” The very way we ask about real estate’s value has changed. It used to be that a single number, signed off by a single appraiser, was the answer. Now, a question mark follows that number as a matter of course: how much can this number actually be trusted? The ability to answer that question mark honestly has become the new qualification demanded of both people and machines working in real estate in the AI era.

Now it’s time to carry that question mark into the next question. Why is AI so intent on calculating real estate’s value with this much precision? And where, exactly, does the AI itself live?

---

Rule of the Game

A model summarizes the market; it doesn’t make the market. AI is only as smart as the data is deep, and knowing the accurate number doesn’t mean having the nerve and the timing to actually put capital behind it. So the side that wins isn’t the side with the more sophisticated model — it’s the side that knows exactly how far that model can be trusted, and where a human has to take over.

---

Sources

Footnotes

1. Zillow ran an iBuying business (Zillow Offers) from 2018, using its own AVM “Zestimate” to buy and sell homes directly, but withdrew after losing $421 million in Q3 2021. The Zestimate’s error rate against on-market listings has been reported at around 2%. (The Close, “Zillow Estimates Ultimate Guide”; Best Practice AI, “Zillow provides real estate price estimates”) ↩ ↩²

2. On AVM error rates — industry materials broadly cite 2–3% for standard residential assets and 5–15% for non-standard or commercial assets. (PatSnap, “AI property valuation technology landscape 2026”; BusinessWire, “PropStream Announces New AVM & AI Innovations”) ↩

3. Industry vendor/consulting materials (Blooma, GrowthFactor, Alpaca, etc. ) — banks adopting AI underwriting report 50–75% reduction in review time, up to 20% cost savings, and cases of sub-8-minute underwriting for well-defined deals. Note significant variance by institution and sample; not standardized industry statistics. ↩

4. The U. S. Interagency Rule on AVMs (effective 2024) mandates confidence management for valuation models, safeguards against data manipulation, and conflict-of-interest protections. Additional regional compliance requirements, such as Colorado’s 2026 AI law, have also taken effect. ↩

5. JLL acquired Skyline AI, an Israel/New York-based commercial real estate AI startup, in 2021. Full specifications, post-acquisition usage, and JLL’s AI-adoption statistics are covered in Chapter 2. (JLL Newsroom; AI Business; PitchBook) ↩

6. The U. S. Department of Justice charged rent-pricing software company RealPage in August 2024 with facilitating algorithmic collusion, citing a case where one landlord began raising rents within a week of adoption and had raised them more than 25% within eleven months. In November 2025, RealPage settled with the DOJ without admitting wrongdoing, agreeing to discontinue rent-recommendation features that used confidential competitive information. (DOJ official announcement; ProPublica; NPR; Holland & Knight) ↩ ↩²