Most marketing teams have adopted AI. Far fewer are letting agents run the work. After a year in which every major platform shipped autonomous agents, the surveys landing through 2026 tell a consistent story: using AI is now close to universal, but scaling agentic AI — handing real tasks to systems that act on their own — remains the exception rather than the rule.

The numbers vary by who is counting and how they define an “agent”, but the shape is stable. McKinsey’s latest State of AI survey puts AI use at 88% of organisations in at least one function, up from 78% the year before, yet finds only 23% actively scaling AI agents, with another 39% still experimenting. The hype says agents are everywhere. The data says most teams are still trying them out — and that the marketer pacing carefully is in the majority, not behind.

How wide the gap actually is

McKinsey’s figures capture the split cleanly: near-universal adoption of AI, a minority putting agents into production, and a much smaller group — around 6% — that the firm classes as genuine high performers getting outsized returns. Gartner found something similar from the technology side, reporting in late 2025 that just 15% of IT application leaders were deploying fully autonomous agents. Salesforce’s State of Marketing 2026, drawn from 4,450 marketers, puts marketing AI adoption at roughly three-quarters while noting that the highest-performing teams behave noticeably differently from the rest.

Sources disagree on the exact percentages, partly because “agentic AI” means different things in different surveys. What they agree on is the pattern: adoption is no longer the story, because almost everyone has adopted something. The story is the distance between having AI in the building and having it do work that pays off.

Why scaling is the hard part

IBM’s research found that 45% of executives see a lack of visibility into how agents make decisions as a barrier to adoption — and it is only one of several brakes. The clearest is proof: teams struggle to show that an agent’s output is actually better than the manual baseline it replaced, which makes expanding from a pilot hard to justify. The concern bites hardest in regulated work, where someone has to be able to audit what the system did and why.

Underneath both sits data. An agent is only as good as the customer records, product feeds and definitions it draws on, and most teams know their foundations are shakier than they would like. Costs add a final wrinkle: consumption-based pricing means a workflow that loops or retries can run up an unpredictable bill. The cautious pace many teams have taken is a reasonable response to real risk, not a failure of nerve.

What it means for a marketing team

Salesforce’s data points to the more useful conclusion. Its highest-performing marketers were more than twice as likely as the rest to have optimised for AI search — a sign that the teams pulling ahead are doing the foundational work of being found and usable by machines, not switching on the most agents. The lesson is not “move faster” but “get the unglamorous parts right”.

That reframes the to-do list. Cleaning up data, defining what a good outcome looks like, building a simple way to test whether an agent beats the old method, and deciding which tasks genuinely warrant autonomy — this is what separates the teams getting value from the ones running expensive experiments. None of it makes a keynote. All of it compounds.

It also takes the pressure off the marketer who has not yet handed a campaign to an agent. The data is clear that they are in the majority, and that being early is not the same as being ahead. The teams that win the next year will be the ones that picked their spots deliberately and could prove the result — not the ones that adopted the most, the fastest.

The adoption question, in other words, is settled. The value question is wide open. For a marketing team, the one worth asking is no longer whether to use AI, but where it actually compounds — and whether the foundations are in place to tell.