82% of agents use AI.
46% say it changed nothing.

Two numbers from the same industry, twelve months apart.

82% of agents now use AI. Ninety-seven percent of brokerage leaders say their agents do, up from 80% in 2024. Non-adoption at the brokerage level has fallen to around 4%, and only about 2% say they have no plans to adopt in 2026. Whatever argument there was about whether this technology was coming, it is over.

46% report no noticeable impact on their business. Seventeen percent report a significant positive one, thirty-three percent a moderate one. That is NAR's own 2025 Technology Survey — 49,233 active Realtors invited, 1,241 usable responses. It is not a fringe result and it is not a contrarian read of the data. It is the data.

Near-total adoption. Half the users reporting the tool changed nothing. Hold both numbers at once, because the gap between them is the entire subject of this issue.

It is not a technology problem

The obvious explanations are all wrong. The models are not too weak — they got dramatically better across exactly this period. The tools are not too expensive; most of what agents use is free. It is not a training gap in the usual sense either. NAR's own frequency split: 20% of agents use AI daily, 22% weekly, 27% a few times a month, and 32% have not actively tried it. Four in ten are in it every week.

The answer is in what the industry pointed it at.

Every leading use case is content generation. Listing descriptions grew twenty-four points in two years, which makes it the fastest-adopted application in the industry — and it is also the one with the lowest switching cost, the least defensibility, and no measurable effect on whether a file closes on time.

That is the mechanical explanation for the 46%. The industry deployed AI where it was easiest to deploy, not where the money leaks. Nobody's year got better because the listing copy arrived forty minutes sooner.

The agents themselves say as much, from the other direction. Asked for their top concerns about AI, they name accuracy of outputs at 63%, compliance or legal issues at 49%, misinterpretation of market data at 47%, and Fair Housing at 28%.

The control group is one industry over

Property management ran the same eighteen months and made a different choice: it pointed AI at operations rather than marketing.

Half the adoption rate of brokerage, and roughly the inverse impact profile. Same industry, same calendar, same models available to both. The variable is not the technology, the vendor, or the budget.

The variable is whether the AI was pointed at content, or at a workflow with a deadline and a penalty attached.

MIT put a number on the same divide

This is not a real-estate phenomenon. MIT's Project NANDA published The GenAI Divide: State of AI in Business 2025 in July 2025 — 52 organisations interviewed, 153 senior-leader survey responses, and a review of 300+ public initiatives — and found that 95% of enterprise generative-AI pilots produced no measurable business return against $30–40 billion of spend.

The stated cause was not model quality. It was data readiness, workflow integration, and the absence of a defined outcome before the build started. Tools that never entered the workflow they were bought to change.

The surviving 5% shared three traits:

• They bought externally rather than building in-house.
• They attacked back-office friction rather than customer-facing flash.
• They measured workflow change rather than licence adoption.

Read that list next to the table above and the property-management result stops looking like a coincidence. Leasing automation and maintenance triage are back-office friction. Listing descriptions are customer-facing flash.

What this looks like here

In the New York and New Jersey metro, back-office friction is unusually well-defined, because most of it is written into law with a date attached.

Regulated units in New Jersey rent-control municipalities register annually — and the deadline is set by the municipality, not the state, so it moves from town to town. Miss it in Hoboken and the city's own guidance is that a property may forfeit eligibility for future rent increases. That is not a fine you pay and move past. That is the compounding number, gone for the year, on a filing.

New York City's FARE Act moved broker-fee liability to whoever hired the broker and has been in force since June 2025, reshaping how every rental file is papered and invoiced. And on the sale side, New Jersey standardises exactly one thing statewide — the smoke, carbon-monoxide and fire-extinguisher certificate. Whether the town also wants a full certificate of continued occupancy, which form, what fee, how far ahead the inspection must be scheduled: all municipal, all different. Bergen County alone is seventy municipalities.

None of that is a marketing problem. All of it is rule-shaped, deadline-bound, and penalised — the exact profile software is good at. And almost nobody has built for it, because no national vendor will ever maintain town-level rulesets for a single state. The economics do not work at their scale.

They work fine at yours. You do not need 564 municipalities. You need your footprint.

The test, and you don't need me for it

Two columns, this week.

Left column: the tasks AI genuinely finished faster for you in the last thirty days. Be strict — "faster" means you can name the task and the time.

Right column: the recurring work that still runs on one person's memory and a phone call to a municipal office that stops answering at three.

The right column is your actual bottleneck. It is usually longer than people expect. And it is never listing descriptions.

If the exercise is useful and you never speak to me, it has done its job. That is roughly the standard I am trying to hold this newsletter to.