Observability, not compute

Fannie Mae’s Desktop Underwriter reached production in June 1995 and rendered a credit decision in under fifteen seconds. In the same country, in 2026, an institutional office building is valued by a human appraiser, on an annual external rotation, engaged and paid by the party whose fee is a function of the resulting number. The technology was ready for the harder problem long before it was applied to the easier one.

That gap is thirty-one years wide and it is not a technology gap. Both problems are regressions on comparable sales. The office building has more money at stake, better-funded counterparties, and a far smaller universe of assets to model. If computing power were the binding constraint, the expensive problem would have gone first.

The constraint is observability: how often an asset’s price is recorded, how publicly, and in how standardized a form. Automation diffuses in proportion to it and is largely indifferent to available computing power. Where a dense public price record exists, decisions get ceded to models. Where it does not, valuation stays with a human whose incentives are set by whoever pays for the number.

Residential is automated because it is observed

The residential price record is dense, public, and standardized to a degree that has no analogue elsewhere in real assets. Sales are recorded, aggregated, and republished within days. Three national portals compete to restate the same facts. The consequence is that residential valuation has been quietly ceding decisions to models for three decades.

Desktop Underwriter reviewed 62% of conventional closed loans market-wide from 2018 to 2024. Not 62% of Fannie Mae’s own book. 62% of the market.

Fannie Mae, on HMDA data, 2018–2024

The appraisal itself is now optional on a large minority of loans. In March 2026, 28% of GSE acquisitions carried an appraisal waiver, split 25% at Fannie Mae and 32% at Freddie Mac; the share peaked near 49% in March 2021. A waiver is the clearest statement of institutional confidence available: the buyer of the loan has decided the model is good enough that nobody needs to visit the house.

AEI Housing Center, March 2026

Regulation followed, which is the reliable tell. Six federal agencies finalized quality-control standards for automated valuation models effective October 1, 2025. Regulators do not write model-governance rules for tools that are not making decisions.

Interagency AVM quality control standards, effective 2025-10-01

And the profession contracted. Appraiser credentials fell from 120,551 in 2008 to 95,731 in 2017.

Appraisal Subcommittee registry, via CSBS, 2008–2017

None of that required a breakthrough in computing. It required a price record dense enough to regress on.

Institutional real estate is not, because its record is produced by interested parties

One tier up: the same asset class, the same arithmetic, and almost none of the automation. The institutional price record is thin, slow, and produced by people with a stake in the answer.

The frequency is annual per asset. Common practice is a third-party appraisal of roughly 25% of a fund’s investments each quarter — each asset gets an outside look once a year, on a rotation.

Common non-traded fund practice, as of 2026

The independence is weaker than the label suggests. Surveying ten non-traded NAV REITs, DLA Piper found the adviser performs the NAV calculation for three of them, with the independent valuation advisor performing a reasonableness review rather than a determination. A reasonableness review is a different instrument from an independent number: it can reject, and only at the margin.

DLA Piper, survey of ten non-traded NAV REITs, 2023

And the standardization is absent by design. FINRA Rule 2310 requires an annual valuation with third-party “material assistance or confirmation” and expressly declines to prescribe a methodology.

FINRA Rule 2310, as of 2026

No model is going to be trained on that. Not because the model is not good enough — because there is not enough record to train on, and what record exists was produced by someone whose fee depends on the number.

The lag, and its asymmetry

If appraisals were merely slow but unbiased, the lag would be the same length in both directions. It is not.

Green Street’s CPPI peaked in March 2022. The NCREIF NPI market value index peaked in the second quarter of 2022, one quarter later.

Green Street CPPI; NCREIF NPI, 2022

At the bottom the distance was three quarters. Green Street was down 22% from peak by November 2023 and calling a bottom. The NPI did not stop falling until the third quarter of 2024, after nine consecutive declining quarters.

Green Street CPPI; NCREIF NPI, November 2023 to Q3 2024

One quarter behind at the top and three quarters behind at the bottom is not latency. Latency is symmetric. This is the signature of a number produced by someone who is comfortable being slow to mark down and much less comfortable being slow to mark back up.

Still open, four years on

In the first quarter of 2026 the NPI reported a 4.57% appraisal cap rate against 5.27% on properties actually sold out of the index.

NCREIF Property Index, Q1 2026

The last clause is the one that matters. This is not an outside critic’s estimate of appraisal bias. It is the number the index reports about itself, against its own realized sales, and it has not closed.

The people doing the work say it is data, not compute

The obvious objection is that this is a story about models that are not ready yet, and that enough compute will eventually arrive and settle it. The practitioners do not describe it that way.

Deloitte’s 2026 Commercial Real Estate Outlook surveyed more than 850 senior executives. 1% reported transformative impact from AI.

Deloitte, 2026 Commercial Real Estate Outlook, 2026

Mercer fielded 131 asset managers in February 2026. 55% have integrated AI into at least one investment process. Only 5% to 6% have granted it autonomous or semi-autonomous decision authority. And 69% name data quality or access as the binding barrier — ahead of regulation, ahead of cost, ahead of talent.

Mercer, survey of 131 asset managers, February 2026

Nobody in either survey is citing compute.

Against the thesis

What would make this wrong

Four ways. The first three are evidence that already exists and cuts against the argument. The fourth is the explanation most likely to be right if the argument is wrong.

Observability without standardization does not help

Publishing a record only creates a reference point if everyone means the same thing by it. For July 2026, Trepp reported an overall CMBS delinquency rate of 7.86% with office around 11%, while Fitch reported 3.49% overall and 8.89% for office.

Trepp; Fitch Ratings, July 2026

Both numbers are correctly computed. They cover different rated universes and use different delinquency definitions. A market can be observed continuously and still fail to produce a usable price record, which means observability is doing less work in this argument than it first appears.

Mandated transparency can reduce activity, not just cost

Asquith, Covert and Pathak find that after TRACE, costs fell across all bond types but trading activity did not rise. Trades fell 11.7% overall and 71.1% in the last-disseminated, least-liquid tranche, as dealers withdrew from intermediating the least observable paper.

Asquith, Covert and Pathak, 2019

If publishing a record can drive the intermediary out of the market it describes, then assembling the record and publishing its existence is not automatically good for that market. It may be good for us and bad for them, which is a worse business than it sounds.

Automation may consume information rather than produce it

Weller finds the information content of prices decreases 9% to 13% per standard deviation of algorithmic trading activity.

Weller, 2018

At the limit this inverts the argument. If automation degrades the price record that made it possible, the observation layer is not a durable position but a resource being drawn down by the thing it enabled.

The alternative explanation most likely to be right

Transaction size. A $200 million building can support human underwriting in a way a $400,000 house cannot. On that reading, automation arrives wherever the per-transaction cost of a human exceeds the cost of building a model, and observability is a coincidence of which assets happen to be cheap enough to need one.

The evidence against it is the split below. It holds transaction size fixed entirely — the same houses, the same housing stock, the same day, the same model — and varies only whether a price was published.

The error is four times larger on the unobserved population. Transaction size cannot explain that, because it does not vary.

The claim in its defensible form

Observability is necessary, not sufficient. A dense, public, standardized price record is a precondition for automation, and its absence is a reliable predictor that valuation will stay with a human. It does not follow that observability alone produces a good market, that publishing a record is always welfare-improving, or that a data moat is permanent once it exists.

What does follow is narrower, and enough to build on: where the record does not exist, whoever assembles it first is the only party who can see the market, and that position accrues at the observation layer rather than the computation layer. Compute is a commodity. Observation is not.