Every financial decision that touches fine art rests on a valuation someone must trust. A premium and a claim. A net-worth statement. An estate settlement or a matrimonial division. A charitable deduction or a collateral advance. Each one consumes the same input, and the input is only as good as the process behind it.
The gold standard for that input, in the absence of an actual sale, is a qualified human appraisal. It is also hard to scale. Appraisal is expensive, slow, and logistically heavy, so the temptation to automate it is old and understandable. Automated valuation models promise portfolio-scale revaluation at almost no marginal cost, a promise they partly delivered in residential mortgage lending.
We wanted to know whether that promise transfers to fine art. So we did something unusual for a company that would benefit from the answer being yes. We tried to build one.
This post summarizes what we found. The full argument, with the evidence and the citations, is in our white paper, Model-as-Trigger, Appraisal-as-Arbiter.
Download the white paper (PDF)
What an AVM actually needs
An automated valuation model is the market approach industrialized. It predicts what a market would bid for an asset, in the absence of any bid, by learning from a large body of comparable transactions. That machinery runs on three conditions at once.
The assets have to be similar enough that their measurable characteristics explain most of the price. They have to trade often. And those trades have to be recorded in a machine-readable dataset. Combine the three and you get the single number that governs whether a model can work: comparable transaction velocity, the rate at which arm’s-length, digitized sales occur within an asset’s comparable set.
Residential real estate clears that bar. A tract home belongs to a comparable set of thousands of near-substitutes, several of which sell every week, all publicly recorded through deeds and listing services. The individual house almost never trades and the model never needs it to. The population trades on its behalf.
Why art sits below the chart
Fine art fails all three conditions, categorically.
It is unique. A painting is not a unit drawn from a population of substitutes. It is, in the limit, its own population. Two works by the same artist, same year, same size, same medium can sell an order of magnitude apart based on subject, provenance, condition, and freshness to market. The comparable set is one.
It is illiquid. A metro house sells in weeks. A significant artwork commonly takes twelve to twenty-four months to sell at fair value. Repeat sales of the same work are rare and separated by decades. Three centuries of the Western art market produced only a few thousand repeat-sale pairs. The US housing market records more comparable transactions than that before lunch on an average Tuesday.
It is barely digitized. In 2025, global art sales reached an estimated $59.6 billion. Public auctions, the only segment that produces the data a model trains on, accounted for a minority of it. Roughly two-thirds of the market by value transacts with no public price record at all. The most price-relevant knowledge in this market is held by the people who profit from keeping it undisclosed. A market can lag on digitization by inertia. This one withholds by design.
Commercial real estate makes the point cleanly. CRE assets are far more standardized than art, yet institutional CRE is still valued by appraisal, because there are not enough transactions to mark a portfolio to market. If quantifiable office towers cannot support a model, unique paintings never had a chance.
Even an accurate model would not produce a value
Suppose a model could estimate a work within a tolerable band. It still would not produce an official value, for reasons that have nothing to do with statistics.
The standards each draw the same line. USPAP holds that a model’s output is not, by itself, an appraisal, because an appraisal is an opinion of value and only a person can form one. The RICS Red Book, the International Valuation Standards, and US Treasury regulations independently require that a qualified appraiser be a person.
Behind the standards sits a harder requirement no model can meet: accountability. An algorithm cannot explain why it weighted one comparable over another. It cannot answer an examiner’s challenge, testify in court, or bear liability for a conclusion that exposes an estate to penalties or an insurer to a bad-faith claim. When two tools disagree, no mechanism decides which is right, because neither can give an account of itself.
The stakes are documented. The IRS Art Advisory Panel reviewed 195 items in fiscal 2023, on $795.5 million of claimed value, and adjusted 47% of them, in both directions. When the most scrutinized valuations in the market are tested, nearly half do not survive intact. A wrong number carrying a confident label is worse than no number.
The proof already exists
The clearest evidence comes from the one asset class where AVMs reached scale. Zillow made its own estimate the live cash offer for eligible homes, bought roughly 7,000 properties on those figures, and shut the business down within the year. The home-flipping operation lost $881 million in 2021.
A median error that is tolerable in a marketing widget is fatal when you transact at the estimate thousands of times on thin margins. Display-grade accuracy is not transaction-grade accuracy. This happened in the asset class with the highest comparable transaction velocity on earth. Fine art has almost none.
What works instead
The productive output of our research was not a model. It was an architecture.
If an algorithmic estimate in this asset class cannot be a valuation, what can it responsibly be? A trigger. Our design converged on a clear division of labor. The model monitors reference clusters of comparable works and estimates whether an object’s value has probably drifted since its last qualified appraisal, with an uncertainty band. It then does the one thing a statistical estimate is qualified to do. It classifies. Green means the value likely holds. Amber means review it. Red means revalue. A red flag routes to a qualified human appraisal, and the appraisal sets the value.
We call it Model-as-Trigger, Appraisal-as-Arbiter. Every objection above is answered by where the decision sits. The model never issues a valuation, so the standards are satisfied by construction. Every value has a human author, so accountability holds. And the appraiser’s scarce time is spent where drift is probable, so accuracy arrives at a cost sized to the premium.
That reframes the scaling problem. It was never a modeling problem. It is a routing and logistics problem, and that is the infrastructure Title Collections builds: matched appraiser selection, secure data rails that deliver an expert everything they need at once, and outcome capture into a system of record so values stay live and auditable. Underneath runs a health diagnostic that scores each object’s documentation and brings it to appraisal-ready before a claim, a loan, or a settlement forces the question.
What this means for you
If you underwrite, advise, or litigate around these assets, the questions are the same. What share of your scheduled values still carry the purchase price, and how old is the oldest figure? If a claim, a division, or a catastrophe demanded defensible values on ninety days’ notice, how many objects could deliver one? When a vendor shows you error statistics, have you seen the tail of the distribution, or only its center?
Give your clients current, defensible values without a schedule-wide revaluation every year. Title monitors for drift, routes the appraisal, and captures the result, so the expensive work happens only where it is needed.
The full white paper walks through the evidence, the regulatory floor, and the architecture in detail.