Agentic advertising is best defined by one shift: the moment an advertising transaction stops passing through a human and starts passing directly between two machines. Most definitions describe something narrower, AI that plans, buys and optimises a campaign with little human involvement, which is really programmatic with a faster engine. The genuinely new thing is not that a machine can buy media. It is that a buyer’s agent and a seller’s agent can agree a deal with nobody confirming it, and that agreement creates a demand the old model never faced: a single record of what was agreed that both sides can hold.
In short (August 2026). Agentic advertising is the shift from human-approved media transactions to deals agreed directly between AI agents. The capability is no longer in question: in December 2025, Anthropic’s Project Deal showed autonomous agents negotiating and closing deals unaided. The unresolved problem is agreement after the fact, because when each agent keeps its own record of a deal, a 2026 study run with WPP found the two records disagreed in 95.3% of 90,202 simulated transactions. What defines agentic advertising, and decides whether it becomes a market, is a shared record both sides write to.
What does agentic advertising actually mean?
The word does too much work. “Agentic” now covers a campaign that sets itself up in a fraction of the old time and two machines closing a deal with nobody watching. Both get filed under the same term. Only one of them is new.
Even the people setting the standards say so. Anthony Katsur, chief executive of the IAB Tech Lab, told The Current in October 2025 that “agentic is a shiny penny”. He was not dismissing it. He was warning against treating one broad, shiny label as though it named a single thing.
The plain definition is the place to start, because it is not wrong, only incomplete. Agentic advertising means AI that can plan, buy and optimise a media campaign with little human involvement. Not software that recommends and waits for a click. Software that reads the brief, builds the plan, launches, adjusts pacing and reports back. When Butler/Till bought CTV for Clubtails through PubMatic in December 2025, the brief went in through an AI, the agents built the strategy, and the humans stayed on the creative. Campaigns like it are live now, and the money is moving through them.
Agentic advertising also sits on a spectrum, and the single word refuses to say where. At one end, AI does the repetitive work a person used to do and still signs off. At the other, a buyer’s agent and a seller’s agent negotiate directly and close, with no human confirming the terms. Most of what ships today sits near the first end. The industry keeps pointing at the second. The distance between them is the whole story.
How is agentic advertising different from programmatic?
Agentic advertising is easy to confuse with programmatic, because programmatic already automated most of what the plain definition describes. A bidding engine sets a price against a counterparty’s mechanism thousands of times a second, faster than any human trader could manage, and nobody ever built an approval step around it. That is an agent setting a price, and it has run for years under a duller name. So “an agent can buy media” cannot be the line that separates agentic from what came before.
The line is not what one agent can do on its own. It is what two agents can do to each other. An agent optimising a campaign answers to a human who set the goal and will read the outcome. An agent agreeing a deal with another agent answers to nobody in the room, because there is no room. That is the shift: not autonomy of action, which programmatic already had, but autonomy of agreement.
The capability itself is proven. In December 2025, Anthropic ran a study it called Project Deal: 69 agents closed 186 deals worth roughly $4,000 in a week, in a marketplace built on chat. The agents negotiated, countered and closed with real strategic skill. When stronger models traded against weaker ones, the stronger side quietly extracted more on every sale and paid less on every purchase, and the humans on the losing end never noticed they had been beaten. Agents can deal. The open question is not whether two agents can agree, but whether they can agree on what the deal was after it is done.
Why do two AI agents disagree on a deal they both closed?
Two agents can close a deal and still hold different accounts of it, and the reason is structural rather than a matter of one agent being wrong. When one agent acts and a human checks the result, one record matters: the human’s. When two agents agree and both walk away, there are two records, one each, and nothing forces them to match. They do not match. Not through malice or a bug, but through the ordinary mechanics of two systems each writing down its own version of the same event.
A study published with WPP in May 2026, Agree. Transact. Verify., modelled this across 90,202 simulated agent-to-agent deals. Under separate record-keeping, where each agent kept its own account of what was agreed, the two records disagreed on at least one term of the deal in 95.3% of transactions. Under a single shared record that both agents wrote to, disagreement fell to 0.19%. Same agents, same market, one change: whether the truth of the deal lived in one place or two. (These are simulation figures; the full methodology and codebase are available on request.)
The gap does not announce itself. In the same simulation, run to ninety days at holding-company scale, 677 million impressions reached settlement with no agreed record of what had been delivered. Buyers paid against their numbers. Sellers invoiced against theirs. Nothing flagged, because from inside either system nothing looked wrong.
This is what separates a market from a negotiation engine. A negotiation engine closes deals. A market remembers them, the same way for everyone, so that a price means something, a track record can be trusted, and a dispute has a fact to resolve against. Take away the shared record and agents can still close all day. What they cannot build is a market, because a market is a shared memory before it is anything else.
So how should agentic advertising be defined?
Agentic advertising, defined usefully, is the point at which the transaction stops passing through a human and starts passing between machines. Everything before that point is automation: useful, real, and continuous with programmatic. Everything after it is new, and it arrives carrying a requirement the old world could ignore, a record of what was agreed that both sides hold and neither can quietly rewrite.
That requirement doubles as a test to carry into any agentic pitch. When a vendor shows two agents shaking hands, the question that matters is not how fast they closed. It is where the deal is written down, who can read it, and whether the buyer can prove, six weeks later across four hundred deals, what the agent was actually told to do.
Most of the market answers the first half of that and goes quiet on the second. It sells the handshake and stays silent on the record. The handshake was always going to be the easy part.