The one number he won't let the model touch

It’s 9:40 on a Tuesday and an acquisitions analyst is rebuilding a number somebody already handed him. The offering memorandum says net operating income of $2.4 million. Behind that figure sits a trailing twelve carrying three months of the prior owner’s accounting, a management fee that changes definition halfway down the page, and a rent roll exported on a date nobody wrote down.

A chat window is open on his second monitor. He’s used it twice tonight, once to pull the renewal options out of a 140-page lease and once to tidy the market paragraph he’ll paste into the memo.

He will not let it near the NOI.

That refusal is what the survey numbers are actually describing. The CRE Industry Pulse Check that DealGround fielded with First American Data & Analytics between 31 March and 8 April 2026, among 255 professionals in brokerage, lending and capital markets, development and asset management, found 66 percent using AI weekly or daily and 5 percent who trust it enough to inform a real deal decision.

The gap between those two numbers has less to do with how good the models are than with where the answer came from. An analyst cannot put a figure in front of an investment committee that he can’t trace to a page and a date. So anything that produces an answer without showing its source gets used for drafting and never for deciding.

Look at where the usage sits. In the same survey, market research and comps led at 20 percent, lease abstracting and document review at 15 percent, marketing and listing copy at 14 percent. Deal sourcing, the application closest to money, finished last at 5 percent, and 53 percent said they use these tools for support only, keeping them out of final decisions entirely.

The stated barriers point the same direction. Thirty-two percent of the professionals DealGround surveyed named output trustworthiness as the barrier; 5 percent named cost or unclear return. Nobody is waiting on the price to come down.

The industry mostly reads this as immaturity, a trust curve that will bend upward on its own. We read it as a correctly priced verdict. The analyst isn’t being timid; he’s doing arithmetic on who carries the number.

Here is the moment that decides what’s worth building. A partner is three pages into the memo and stops on a rent growth assumption of 3.25 percent. Where did that come from? The answer that survives the next ninety seconds is never that the model suggested it. It’s three signed comps in the submarket, the quarter each was executed, the source each came from, and a sentence about why the fourth one was excluded.

Anyone who can’t produce that hasn’t made an argument. They’ve made an assertion with a decimal point on it.

Which points the work somewhere unglamorous. A lease abstraction that returns $45.00 per square foot is a draft. A lease abstraction that returns $45.00 per square foot, page 3, section 4.1, from the executed lease dated 14 February, with eight months of free rent captured separately, is evidence. Same model underneath, completely different object at the top of a memo, because one can be interrogated and the other can only be believed.

That distinction is also why the honest obstacles are honest. Comps genuinely disagree with each other. CompStak’s net-effective-rent guide shows why the definition matters: starting rent is required, while net effective rent is calculated from concession data, and not every comp includes those concessions. Two figures can therefore answer different questions. The workflow has to preserve which definition it holds.

The reasonable objection is that this is a waiting problem. Models improve every few months, accuracy climbs, trust arrives on a lag. Build nothing unusual and be patient.

That fails on the shape of the cost. An analyst who is wrong on one line of a committee memo doesn’t lose a percentage point of accuracy — he loses the room, and he carries it personally for a year. A tool that is right ninety-seven times out of a hundred and cannot tell you which three are wrong has an accuracy number that means nothing at the table. Excellent for the draft, disqualified from the decision, which is exactly where the survey found it sitting.

The deeper trouble with waiting is that accuracy isn’t well defined on this data to begin with. If two sources report different rents for the same lease because one is face and one is net effective, there is no single true number for a model to converge on. Better prediction cannot settle a definitional disagreement. Showing which definition you’re holding can.

The one corner of this business where somebody has written the expectation down landed in the same place. On 23 April 2026 the Appraisal Standards Board adopted Advisory Opinion 41, Use of Technology in an Appraisal or Appraisal Review Assignment. The Foundation says AO-41 addresses appraisers’ USPAP responsibilities when they use automated valuation models, statistical software, or generative AI. For the workflow, that translates into a source-linked record of the tool output, inputs, instructions, and the appraiser’s decision.

Advisory Opinions are guidance rather than enforceable USPAP requirements, and they speak to appraisers, not to brokerages or acquisitions desks. Nobody should stretch this into a rule it isn’t. But read that workfile description as an engineering specification instead of paperwork and it is the most precise product requirement anyone has written for this industry.

A valuation is a signature. Signatures need records.

When we build provenance into an underwriting workflow, we spend the first stretch of it nowhere near a model. We spend it on the assumption tab whose source is a conversation from March, the comp set copied out of a deck two deals ago, and the rent roll with no export date — the same defect as an uncited model output, sitting in the file the whole time.

Most desks find their own spreadsheets can’t answer where did that come from either. The tooling was never the reason.

The payoff shows up in the disagreements. Two rents that differ because one is face and one is net effective, sitting side by side with their definitions attached, tell you something real about the concession environment in that submarket. Averaged together they tell you nothing and quietly move a valuation in a direction nobody can reconstruct later.

Somewhere a desk is about to hand a committee a memo where every assumption opens into the document it came from. It will take longer to build than the one before it. It will also be the first one in the room that nobody has to take on faith.