Between 2024 and 2026, artificial intelligence went from a minority experiment in residential real estate to something close to a default. Over the same period, the share of agents reporting that AI made no noticeable difference to their business stayed close to half.
Both of those statements are supported by industry survey data, and together they describe a gap worth taking seriously. This piece sets out what the published evidence actually shows, where it is thin, and what we think explains the discrepancy.
What sellers say they are buying
Start with demand, because it is the least ambiguous part of the record. In NAR's 2025 Profile of Home Buyers and Sellers, 91% of sellers used a real estate agent — matching the highest share on record — and for-sale-by-owner transactions fell to 5%, an all-time low.5
More pointedly, when sellers are asked what they want from the agent they hire, marketing the home to potential buyers ranks at the top of the list, ahead of pricing and ahead of speed of sale.5 That is the product being purchased. Not access to the MLS, which is commoditised, and not paperwork, which is invisible until it goes wrong.
of home sellers used an agent in 2025 — tying the highest share on record — and ranked marketing the home as their top criterion when choosing one.
NAR, 2025 Profile of Home Buyers and Sellers
So the marketing of a listing is simultaneously the agent's headline deliverable and the thing they are most often judged on. That makes it the natural first target for automation, and the industry has treated it that way.
Adoption is effectively complete
Delta Media Group has surveyed brokerage leadership annually for three years. In its 2026 survey of more than 100 brokerage executives, 97% reported using AI, up from 87% a year earlier.3 Delta's own multi-year comparison puts brokerage non-adoption at 4%, down from 22% two years prior.4
The perceived importance of AI moved just as fast. Asked to rate how important AI is to their business on a ten-point scale, brokerage leaders averaged 4 in 2024, 6 in 2025, and 7 in 2026 — and predicted 8 for the year ahead.3
How important brokerage leaders say AI is to their business
Source: Delta Media Group, 2026 Real Estate AI and Leadership Survey 3
The 2027 value is respondents' own forecast, not an observation, and is drawn hollow for that reason. Ratings are means on a 0–10 scale from a self-selected sample of 100+ brokerage executives.
View the data
| Survey year | Mean rating (0–10) |
|---|---|
| 2024 | 4 |
| 2025 | 6 |
| 2026 | 7 |
| 2027 (predicted) | 8 |
A caveat that matters: Delta's respondents are brokerage leaders, self-selected, roughly a hundred of them. That is a sentiment reading from the top of the org chart, not a probability sample of practising agents. It tells you what leadership believes and buys. It does not tell you what happens on a listing.
Measured impact did not follow
For that, the better instrument is NAR's 2025 REALTORS® Technology Survey, fielded in July 2025 to a random sample of 49,233 active members.1 It asked agents not just whether they use AI, but what difference it made.
32% had never used AI in their business; 20% used it daily, 22% weekly, and 27% a few times a month.1 Call that roughly 68% adoption2 — high, and consistent with Delta's direction of travel, if well short of the leadership figure.
The impact question is where it gets interesting. 17% of agents reported a significantly positive impact on their business. 33% reported a moderately positive impact. And 46% reported no noticeable impact at all.1
Reported impact of AI on agents' business
Source: National Association of REALTORS®, 2025 REALTORS® Technology Survey 1
Shares are of all respondents, including those who have not used AI. Figures do not sum to 100% in the published summary.
View the data
| Reported impact | Share of agents |
|---|---|
| No noticeable impact | 46% |
| Moderately positive | 33% |
| Significantly positive | 17% |
Set the two datasets beside each other and the shape of the problem is clear. Near-total adoption at the top. Roughly two-thirds usage among practitioners. And fewer than one agent in five able to point to a significant effect on their business.
The step that got automated was the cheap one
Look at what AI is actually used for. In NAR's survey, AI-generated content was used by 46% of agents — behind eSignature at 79% and social media at 75%.1 The dominant tools were general-purpose assistants: ChatGPT at 58%, Gemini at 20%, Copilot at 15%.1 Delta's leadership survey points the same way, with content creation and writing listing descriptions the two most-cited applications for both brokerages and their agents.3
In other words, the industry standardised on general-purpose chat assistants pointed at the writing step.
Writing is a real task. It is also, in the production of listing marketing, among the least expensive. Drafting a property description is perhaps twenty minutes. What surrounds it is not:
- Laying that copy into a brochure and typesetting it so it survives contact with a long description and an awkward vertical photo.
- Producing a listing website, a social set at several aspect ratios, an email announcement, and a print flyer from the same facts.
- Keeping every one of those surfaces correct when the price changes on a Thursday.
- Making all of it look like the same brokerage when twenty-four agents each produce their own.
A chat assistant improves the first task and touches none of the others. That is a coherent explanation for why adoption can be near-universal while measured impact sits at 17%: the tooling was applied to the part of the workflow that was never the bottleneck.
Meanwhile, the cost base moved the wrong way
If AI were absorbing production cost, one place it might show is in what agents spend to run their business. NAR's Member Profile reports a median of $8,450 in annual business expenses for 2023 and $8,010 for 2024.6 For 2025, that median rose to $9,530.7
Median annual business expenses per REALTOR®
Source: HousingWire, reporting NAR’s 2026 Member Profile, NAR 2026 member profile shows Realtors more experienced 7
Totals are self-reported medians across all members and cover all business expenses, not marketing alone. NAR reports vehicle operation as the largest single category ($1,580 in 2025). Because these are medians, categories do not sum to the total, and this series cannot be read as a marketing-spend trend.
View the data
| Calendar year | Median expenses |
|---|---|
| 2023 | $8,450 |
| 2024 | $8,010 |
| 2025 | $9,530 |
We want to be careful with this one, because it is the figure most likely to be over-read. It is a median across all business expenses — vehicle, affiliation fees, education, technology, marketing — and NAR reports vehicle operation, not marketing, as the largest single category.7 It is not a marketing-spend series and should not be quoted as one.
What it does establish is narrower and still useful: through the period of fastest AI adoption in the industry's history, the median agent's cost of doing business rose rather than fell. Whatever AI has delivered so far, it has not shown up as visible relief in the cost base.
On a figure we did not publish: we wanted to show marketing spend broken out by brokerage size and US region. That data is not in the public literature — the granular brokerage financial benchmarks live in paid industry reports, and the per-agent marketing-spend numbers circulating on content-marketing blogs (commonly "$10,600" or "$12,000" a year) trace back to no identifiable primary research. We left the chart out rather than repeat an unsourceable number.
Three explanations, and which one we find persuasive
There are at least three readings of the adoption–impact gap.
It is too early. Diffusion takes time, measurement lags, and 2025 was simply early. This is plausible and partly right — but the NAR question asked about impact on the respondent's own business, which is the kind of thing a daily user should be able to feel. 20% of agents were using AI daily in July 2025.1
Agents are using it badly. Also partly right, and the usual conclusion of vendor commentary, which tends to arrive at "you need more training" — conveniently, from a training vendor. But an explanation that requires 46% of a profession to be uniformly incompetent at a tool they use weekly is doing a lot of work.
The tools address the wrong step. This is the reading the usage data supports most directly. The industry standardised on general-purpose chat assistants,1 applied them to content creation and listing descriptions,3 and left the production, layout, distribution, and consistency work exactly where it was. A tool that saves twenty minutes of writing and none of the four hours around it should produce roughly the impact distribution NAR measured.
There is corroborating evidence in what leaders say they want next. Delta found the value brokerage leaders place on all-in-one marketing platforms with AI and automation at its highest level in three years of surveying.4 That is a market articulating that the assistant layer, on its own, did not finish the job.
What would close the gap
If the diagnosis is right, the fix is not a better writing model. It is moving the automation boundary outward from the sentence to the artefact — which requires a system that holds three things the chat assistant cannot.
- One source of truth per property. The website, brochure, social set, and client portal have to be views of a single project rather than exported copies of it, so a price change propagates instead of being re-entered four times. Every stale-price flyer at an open house is this failure.
- Brand as a system, not a template. Assets generated from an agent's or brokerage's own typefaces, palette, spacing, and print margins — so output is consistent across twenty-four agents by construction rather than by everyone picking the same template.
- Production, not just prose. Typesetting that survives a 400-word description, real print output, correctly cropped images at every aspect ratio. This is the four hours.
None of that removes the agent's judgment, and it shouldn't. Two things in this workflow are professional obligations rather than machine tasks: asserting facts about a property, and deciding what a client sees. Facts sourced automatically belong in a draft an agent reviews, with their provenance visible; publishing stays a deliberate human act. We have written about where we deliberately refuse to go faster for exactly this reason.
What we would still like to know
The public evidence has real holes, and we would rather name them than write around them:
- No public dataset breaks marketing spend out by brokerage size or region. Everything granular sits behind paid industry research.
- No published study we could find measures agent hours spent on marketing production as distinct from lead generation. The time-allocation figures in circulation come from blog posts without stated methodology.
- NAR's impact question is self-reported and unipolar — it does not distinguish "no effect" from "cost me time."
- Delta's sample is small, self-selected, and drawn from leadership, so it measures buying sentiment rather than practitioner outcomes.
The honest summary is that the industry has good data on adoption, thin data on impact, and almost none on production cost — which is precisely the variable this argument turns on. We are working on our own measurement of that last one and will publish it here, including the parts that do not flatter us.
This is the analysis Mareto is built on: that the constraint in listing marketing was never the writing, and that a system producing every asset for a property from one project and one brand system is what turns adoption into measurable time back. See how it works, or read the practical version in what AI listing marketing actually does.
