A number is going around: 90.3% of marketing teams use AI agents. It shows up in trade coverage, in vendor decks, and in posts engineered to make you feel late. It comes from Martech for 2026, the annual report by Scott Brinker and Frans Riemersma, and it is transcribed accurately.
It is also the least informative figure in the report, and the report says so more plainly than most of the people quoting it.
Two notes before the argument. Martech for 2026 is a sponsored report; its sponsors are GrowthLoop, Hightouch, Intuit Mailchimp, MetaRouter, Progress, SAS, and Treasure Data. That does not make its numbers wrong, but it is worth knowing. And this publication is produced with support from Docket, which sells software in this category. The argument here is ours; weigh it knowing that.
One question, five answers
The 90.3% and its sober companion come from the same survey question, about how AI agents are currently deployed in the respondent's marketing or sales organization. The answers split five ways.
Full production, meaning integrated into core processes or customer-facing experiences: 23.3%. Limited production, used in specific cases but not widespread: 35.9%. Experimentation only, covering testing, pilots and proofs of concept: 31.1%. Not in use yet but planned within six months: 7.8%. Not in use, no short-term plans: 1.9%.
The 90.3% is what you get when you take the last two rows away from the whole. It is not a measure of adoption; it is the complement of the 9.7% who have not deployed an agent at all. Anyone who ran a proof of concept once, on a Tuesday afternoon, is inside it.
The report's own summary is blunter than the coverage it generated: most respondents, 67%, were using agents only in pilots, proofs of concept, or limited production.
The authors already published the caveat
Immediately after the 90.3%, the report volunteers that its participants skew, in its own words, "more tech-savvy than the average marketing team," and that this cohort is almost certainly ahead of the curve. That is not a hostile reading. It is the authors' own.
It matters because the sample is not a random slice of marketing. These are martech and marketing-operations people running stacks at large B2C and B2B brands, the population most likely to have tried this already. Read correctly, 90.3% describes the front of the field. Quoted as "marketing teams," it becomes a claim about the company down the road that the survey never made.
There is a second gap the coverage tends to close silently. Martech for 2026 is dated December 2, 2025, and its figures come from the AI & Data in Marketing Survey, 2025. A number describing late 2025 is being used in the back half of 2026 to describe now. In a year when the underlying tooling turned over more than once, that is not a rounding error.
The report also does not publish a sample size for that survey, while it does publish one for a McKinsey source it cites, n=1,927. The omission is conspicuous rather than sinister. But without an n, you cannot weigh this finding against anything else you read.
What the 90.3% is genuinely good for
The number is not junk, and the case against citing it loosely is not a case against the research. It settles one argument outright: among the people who run marketing infrastructure, nobody credible is sitting this out. Only 1.9% reported no short-term plans.
So if you are still building an internal case for looking at agents at all, the figure to bring is not the 90.3%. It is that 1.9%, evidence that the wait-and-see position has almost no constituency left in this cohort. Adoption is not the contested question any more. Deployment is.
Five questions to ask any adoption statistic
This is not a Brinker problem. Their report is unusually well-caveated; the failure is in the citation chain. It is also fixable with a short habit. Before you repeat a number, ask what the question actually was: adoption, usage, deployment, and production are four different questions with four different answers. Ask how the stages were defined, because "in production" means nothing until someone writes down whether a pilot counts. Ask who answered, since a tech-savvy cohort is a legitimate sample and an illegitimate proxy for everyone. Ask what the sample size was, and say so when it is not published. And ask when the survey was fielded, which is not the same as when the report shipped.
Every one of those was answerable for this report in about twenty minutes with the PDF open. That is the entire cost of not being wrong in public.
The 23.3% is the interesting cohort
If you want the figure that should change what you do this week, it is 23.3%. That group has agents wired into core processes or touching customers, which means they have crossed from trying something into running something, whether or not anyone is watching on a given morning.
That is precisely the population we argued has an unfilled seat: the operating row. Who edits the rulebook, who approves the output, who reads the exception queue, who can turn it off. A pilot without an owner is an experiment. A production agent without an owner is an incident with no one rostered to notice it.
The distance between 90.3% and 23.3% is not a maturity curve that resolves itself with time. It is 67 points of work nobody has done yet, and most of that work is governance rather than modelling.