86% use AI for marketing. They rate its performance 2.76 out of 5.
IAB Europe released its 2026 Impact of AI on Digital Advertising Report at DMEXCO on September 24, and the headline number is the one everybody expected. Eighty-six percent of respondents say their organizations use AI for marketing. Among those who answered the investment questions, 72 percent report higher AI investment than in 2025. None of that will surprise anyone who has sat through an agency pitch this year.
The number I keep coming back to is a different one. Asked to rate AI's current performance in their advertising operations on a scale of one to five, respondents averaged 2.76.
That is below the midpoint, and it comes from the people who chose these tools, paid for them, and put them into their workflows. I don't read it as a verdict on the technology.
I read it as a verdict on what was supposed to be built around the technology and mostly wasn't, the measurement that tells you whether it works and the rules that tell people how to use it.
One caution before leaning too hard on any of these percentages. The survey rests on 50 respondents from agencies, advertisers, publishers, ad tech companies, and other organizations, so a single respondent moves a result by at least two points, and some of the splits rest on much smaller groups. This is a directional read, not a census, but the direction is hard to miss. Adoption is close to universal, investment is climbing, expectations for agentic buying are running high, and the performance score is sitting under three.
When the industry explains how it judges AI, it mostly talks about speed. Operational efficiency was the most common measure of AI effectiveness, chosen by 69 percent of those who answered, followed by CPM improvement at 48 percent and faster or more accurate reporting at 45 percent. That lines up with where the tools actually live. Reporting is the most common AI-supported workflow, named by 76 percent of those who described their deployments, well ahead of programmatic optimization and media planning.
I understand why efficiency comes first. It is visible, it is easy to count, and it gives a team something to show for the first year of spend. The trouble starts when it quietly becomes the entire case.
A report that arrives in half the time doesn't make the media decision built on it any better, and an automated bid adjustment says nothing about whether the budget created demand that wasn't going to happen anyway. You can get very efficient at work nobody has proven is worth doing.
At some point the investment has to connect to something a CFO would recognize, whether that is incremental revenue, margin, retention, or media that finally stopped being wasted. Which one depends on the use case, but there has to be one. I suspect that is part of what the 2.76 is picking up. The people rating these tools can see the activity going up without being able to see the value that was supposed to come with it.
This is the part of the report I care about most. Seventy-eight percent of respondents have a formal owner for AI governance, but only 48 percent have AI guidelines written specifically for advertising and marketing. On a slide, those two numbers sit next to each other and look like progress. In practice, the distance between them is where most of the risk lives.
Naming an owner matters. It means somebody is accountable, and AI is no longer parked as an innovation experiment. But an owner doesn't tell a media buyer whether a tool may touch customer data, whether generated creative has to be reviewed before it runs, how a vendor's performance claims get tested, or what has to be kept on record after an automated decision. Those are the questions teams run into every day.
Without specific rules, the owner turns into an escalation point, and people either wait for answers, make their own calls, or keep using tools nobody has reviewed.
I'd go further than the report does. A governance owner and a policy document are what governance looks like before anything runs. The real test happens at runtime, when a buyer drops a customer list into a tool on a deadline or an agent shifts budget overnight, and again afterward, when somebody asks what was decided and why. Advertising makes that test harder than most functions, because decisions are made fast and spread across a long chain of vendors, markets, formats, and audiences. Enterprise-wide principles don't survive contact with that environment unless someone translates them into the language of a media plan.
The industry seems to know it. Sixty-two percent of respondents said they would benefit from AI marketing guidelines developed by an industry association, and 38 percent are still working out what the EU AI Act means for them.
Security ranks as the top AI concern and privacy second, while a lack of expertise or training remains the most common barrier to adoption.
That doesn't describe an industry that doesn't care about governance. It describes one that has assigned the responsibility without yet writing down how the work is supposed to be done. Good governance makes the approved path the fastest one, and that only happens when the rules are specific enough to act on.
Expectations for agentic ad buying are running well ahead of what is deployed. Fifty-eight percent of respondents expect it to reach operational use or scale within the next year, and 30 percent expect it to become a main trading method in at least some markets over the same period. At the other end of the room, 8 percent don't expect any value from their AI investments in that time at all.
The deployment data tells a quieter story. IAB measured agent autonomy on five levels, from agents acting on a user's directions and decisions at Level 1 to agents operating on their own under monitoring at Level 5. Most respondents either have no agentic systems or run systems that people still direct or plan with, and only a couple claimed Level 5. Size makes a difference. Among those who answered, 86 percent of respondents from organizations with more than 500 staff said their most advanced production agent had reached at least Level 2, the stage where people and agents plan, delegate, and execute together, compared with 48 percent at smaller organizations. That gap makes sense, since larger companies tend to have the data infrastructure, the specialists, and the control functions. But Level 2 of five is still a long way from letting an agent spend real money unsupervised.
The difficult part begins after an agent proves it can place a bid. Someone has to decide which calls it may make, what data it may use, how much money it can put at risk, and when a person has to step in. And when performance slips or a control fails, someone has to be able to reconstruct what the agent did and why. Until a company can do that, agentic buying is a forecast, whatever the pilot deck says.
If I were advising a CMO on where to take this, I wouldn't start with how much AI to deploy. I'd start with how much authority each use of it has actually earned, because that authority has to be built up in sequence. It begins with experiments that have a clear purpose and a small risk surface, and with the one discipline most teams skip, which is measuring the baseline before anything gets automated. Without a before, every later claim of improvement is a guess, and a company running dozens of pilots with no consistent way to evaluate them looks busy while learning very little.
The next step is moving AI out of individual habit and into governed workflows, where approved tools, data restrictions, review duties, and escalation paths are part of the process rather than a policy PDF nobody opens. Those rules have to separate risks, because drafting internal copy with an approved tool is not the same exposure as feeding customer data into automated targeting. Only once that is in place does it make sense to hold AI to the standard of any other investment, with agreed success criteria and a way to separate the system's effect from changes in media, creative, pricing, or the market. That works in both directions.
A tool that saves labor but weakens campaign quality is not a clean win, and a tool that raises technology costs can still be worth it if the commercial gain is larger and provable.
Autonomy comes last, and it gets granted one use case at a time. An agent can be ready to adjust bids within a narrow range and still be nowhere near ready to move budget across markets, choose sensitive audience segments, or approve creative claims. Each added authority needs its own spending limits, monitoring, and a condition that stops it, designed in before the agent acts. Companies that skip the earlier steps can reach technical autonomy without ever having operational control, which is the most expensive way to find out where the gaps were.
Nobody in advertising needs another headline announcing that AI use is widespread. The more useful question is what each use of it is allowed to do and who can account for it. For any significant use case, a CMO should be able to say who owns it, which rules apply, what the baseline was, which business outcome it is meant to move, how much autonomy it has, and what would make the company pull it back. If those answers don't exist, the use case is less mature than the technology behind it suggests.
IAB Europe's own Data & Innovation Strategist, Dimitris Beis, framed the leadership question in the report's announcement as "who is accountable when something goes wrong?" That is the right question, and I'd argue it is the one the 2.76 is really about. If a company can't answer it for every system touching its media budget, 86 percent adoption is just 86 percent exposure.