Advertising is entering an agentic phase, where AI agents can negotiate, transact, and increasingly attempt to execute campaigns across the supply chain. Standards bodies are actively building this future. IAB Tech Lab is mapping existing ad standards into “agentic” extensions and creating an Agent Registry intended to improve trust and transparency.
Meanwhile, Ad Context Protocol (AdCP) is trying to standardise how AI agents talk to ad platforms, so the same agent can discover inventory and place or update media buys across different sellers without needing a bespoke integration for each one.
This all sounds sophisticated and exciting, especially if you are trying to drive more efficiency into advertising operations, but there’s one glaring problem.
AI agents are probabilistic by design. A language model is literally a probability distribution over sequences of tokens, and it generates outputs by selecting what comes next based on those probabilities. That means you cannot guarantee exactly what it will do, or that it will do it the same way every time.
That is fine for low-risk tasks like reporting and insights. It is not fine for high-risk tasks like placing a buy, changing budgets, or optimising a live campaign.
My argument is simple. Sell-side agents should support reporting, insights, and forecasting. But they should not support executing, or optimising media campaigns directly.
The only distinction that matters
Most of the excitement around agentic advertising is about efficiency. Fewer manual steps. Faster optimisation. Less operational drag. And on paper, it sounds inevitable: if AI agents can negotiate and transact, why wouldn’t it also execute?
But there is a line here that I don’t think should be crossed, at least for large clients or agencies managing large budgets and high volumes of campaigns. An AI agent that recommends is one thing. An AI agent that can change a live campaign is something else entirely. The moment an AI agent can place a buy, reallocate budgets, alter targeting, or change bidding strategies, it stops being a helper and starts making decisions that directly shape delivery.
This matters because campaigns are not just settings. They are the execution of a strategy that has been crafted, signed off, and often contractually committed to. Once an AI agent can change a live campaign, you are no longer executing the plan you agreed. You are continuously rewriting it, and accountability gets blurred. And when accountability gets blurred, trust breaks.
Where the risk actually lives
The risk is not that AI will sometimes be wrong. Everyone knows that. The real risk is that the industry makes it easy for AI agents to take actions that materially change delivery, at speed, across lots of campaigns, and across lots of platforms.
That is exactly what the emerging standards are trying to enable. IAB Tech Lab’s agentic roadmap is about scaling agentic execution by extending existing standards with modern protocols, so systems can coordinate and transact more efficiently without rebuilding everything from scratch.
But once you accept that direction, you also have to accept what comes with it. If an agent can place or update buys, it can also make the wrong change, at the wrong time, in the wrong place, and do it repeatedly. This is about edge cases, and the fact that generative systems can be unreliable or inconsistent under different conditions.
When you are running big budgets at scale, small inconsistencies become real money, real brand risk, real governance problems, and real liability issues.
The solution is not to slow down the standards. Standardisation is helpful because it makes workflows more consistent and, in principle, easier to govern. The question is where you draw the execution boundary.
A practical line for large advertisers is this. Let sell-side agents help you understand and plan: insights, recommendations, campaign setup, and reporting performance. Let them propose changes. But do not let sell-side agents put campaigns live or directly change them once they are live.
Instead, campaign execution must remain buy-side controlled. The buy-side owns the plan, the budgets and holds the liability if things go wrong.
The rule set we need
None of this is an argument against agentic advertising. It’s an argument for being explicit about where agents should sit in the operating model.
Standards like the IAB Tech Lab’s agentic roadmap are useful because they move us towards common workflows and shared language. They reduce integration friction and make it easier for systems to coordinate. That is good for the whole ecosystem.
But as we make campaign operations easier to automate, we also need to make governance easier to enforce. The moment a probabilistic system can change what is live, the question starts becoming “who is accountable?”
So I think the industry needs a simple set of rules.
Sell-side agents can advise, propose, and explain. They can help with planning, forecasting, and setup. They can surface opportunities and risks. But they should not be allowed to activate, change, or optimise live campaigns.
Buy-side agents can automate execution, but only through deterministic interfaces with clear parameters, human-validation before changes go live, and audit logs that make every action attributable.
If we get that boundary right, we get the upside of agentic workflows without creating a gap where the buyer carries the liability but loses control. And if we get it wrong, we will normalise a world where campaign delivery can be altered by systems the buyer does not govern, and then we’ll spend the next few years trying to bolt accountability back on after the fact.
“The future is already here, it’s just not evenly distributed.“
William Gibson