Advertisers Are Ahead on Agentic AI. Publishers Need a Control Layer

The buy side is further into agentic AI than the sell side. The data behind that claim has limits, and the right response for publishers is control, not volume.

The gap in one survey

In a Digiday survey sponsored by Optable and published on 13 May 2026, 89% of advertisers said they were using or building agentic AI, while only 29% of publishers said they were actively deploying it. Another 29% of publishers were in early planning or piloting, and 42% were not prioritizing agentic advertising or had no long-term plans for it. The survey covered 180 agencies, publishers, brands and retailers.

Read the numbers carefully before building a strategy on them.

  • They answer different questions. "Using or building" (advertisers) is a broader bucket than "actively deploying" (publishers). The fair comparison for publishers is 29% deploying, or 58% if you include planning and piloting.
  • This is one snapshot. The survey measures where each group stood when it answered, not how quickly either group is moving. It does not support claims about relative speed.
  • The method is thinly disclosed. The public report does not give segment sizes within the 180 respondents, the countries surveyed or the fieldwork dates.
  • The sponsor sells into this market. Optable describes itself as an agentic audience platform for publishers and is a founding member of the AdCP protocol. Its head of marketing is quoted in the report.

Even with those caveats, the direction is clear. The buy side has moved further into agentic AI than the sell side. According to the same report, industry investment in agentic AI is focused on media buying optimization (70%) and measurement and reporting (58%), and 66% of advertisers plan to increase spending on agent-enabled buying over the next 12 to 18 months.

The counterpoint: buy-side agents mostly advise

"Advertisers are automating buying" is too strong. In October 2025, Digiday reported that agents in ad tech were mostly limited to advisory and information-retrieval roles. Media agents were "advising rather than placing spend". One agency's agents suggested audience segments, ran competitor analyses and proposed keywords, but none executed media buys. LG Ad Solutions cut campaign report compilation from two days to five hours, and its CTO said none of its agents would handle money.

The 2026 survey points the same way. Advertisers' top concerns about agent-driven buying are brand safety (63%) and loss of control (61%), and 43% say agent-driven buying is hard to integrate with their existing stacks. The advertiser lead is real, but much of it is still analysis, recommendations and reporting rather than autonomous spending.

The wider enterprise data agrees. In a Gartner survey of 360 IT application leaders (May–June 2025), only 15% were considering, piloting or deploying fully autonomous AI agents, while 75% were piloting, deploying or had deployed some form of AI agent. Only 19% had high or complete trust in their vendor's hallucination protection, and only 13% strongly agreed they had the right governance structures in place.

Hlinor's view: the gap is about control, not volume

This section is our analysis, not a survey finding.

When the side that holds budgets adopts agents first, leverage moves toward whoever runs the agents and the rails they transact on: platforms, agencies and ad-tech vendors. Publishers that wait risk having their inventory described, priced and selected by someone else's agent, using someone else's data.

The answer is not to match advertisers on automation volume. Publishers do not seem to be trying. Only 5% of respondents said building or enhancing AI agents was a top publisher priority for the coming year, while 49% said publishers were improving first-party data and signals for AI agent discovery. The barriers publishers name are technical complexity (68%) and organizational readiness (51%), not lack of interest.

That points to a different investment: a control layer. This means proprietary data that agents can read, workflows that decide what an agent may do, and attribution that makes every agent action auditable. It matters on both sides of a transaction. A publisher exposing inventory to buyer agents needs to know which agent asked, what it was allowed to see and what was committed on its behalf.

The same survey suggests the industry is under-investing here. Across respondents, agentic AI investment in governance (56%) and human-in-the-loop controls (45%) trails investment in buying optimization (70%).

What a control layer looks like in practice

The rules below come from practitioners who publish their operating guidance. None of them depends on a particular vendor.

1. Classify tasks by impact and risk, not by channel. WeArePicante uses an impact × brand-risk matrix: high-impact, low-risk work is automated fully, while for high-impact, high-risk work the agent drafts or recommends and a human approves. Bolta puts it more simply: "If the task is repetitive, well-defined, and low-risk, automate more of it."

2. Choose the autonomy tier on purpose. PostSider describes three levels of agent autonomy. In the middle level, which is the realistic production setting, the agent drives the workflow and a human approves key outputs. It cites admove's view that "level three does not exist in production today."

3. Ground every output in one source of truth. Machina lists four things that kill broadcast deployments: approval that does not fit the production clock, outputs generated from separate lookups that drift apart, brand voice that collapses at volume, and no defined behaviour when data is missing.

4. Design exception handling before automation. TFSF Ventures calls exception handling "the primary concern" in broadcast automation. It sorts exceptions into data, system, rights and quality failures, each with its own escalation path.

5. Log, limit and keep a kill switch from day one. Column Content lists logging, monitoring, audit trails and permission systems as essential for deploying agents.

6. Start narrow and expand on evidence. Okara's first step is "Pick One Job, Not the Whole Channel." Machina ranks match previews and recaps first among the outputs that show value quickly, because they are high-volume and low-risk.

What the publisher side is already proving

There is evidence that a governed deployment pays off on the content side too. At an INMA town hall, HT Media said its RankAI product, built from four agents, improved SEO coverage by 50% and user traffic by 20–25%. It also said its newsroom tools reduced time to publish deep explainers and research stories by 40%, with twice as many multimedia formats. These figures are self-reported by the company.

Vendor data deserves the same caution. Gartner found that 45% of martech leaders with AI agents in pilots or production say vendor-offered agents do not meet expectations, and half say they lack the technical and data stack readiness to deploy them.

A Monday-morning checklist

  1. Pick one repeatable, low-risk, high-volume job.
  2. Write down what the agent may read, what it may change and what always needs approval.
  3. Run it drafts-first, with human approval, an audit log, rate limits and a kill switch.
  4. Make your first-party data and inventory machine-readable for the agents that will query it.
  5. Measure time saved for human judgement and the error rate, not the number of outputs.
  6. Expand scope only when those numbers hold.

The advertiser lead in agentic AI is real, even if the survey behind it is vendor-sponsored and the gap is smaller than the headline suggests. Publishers will not close it by automating more. They can make sure that every agent touching their data and inventory runs inside rules they set and can audit.

Sources