Omnicom says AI agents are already buying ads, and the middlemen should worry

Omnicom is testing whether AI can cut a very old tax out of advertising: the fees and friction that pile up between the agency buying an ad and the publisher selling it. On Monday's earnings call, the holding company said it has already executed real client media buys with AI agents talking directly to other AI agents. That matters only if some of the middlemen between buyers and inventory really become optional.

The more interesting part is what Omnicom said next. Chief executive John Wren framed direct publisher relationships as an explicit objective, not a side effect, while the company tied the push to Acxiom, its data business, and to a still-unfinished acquisition of Interpublic Group. In plain English: Omnicom is not just trying to automate media buying. It is trying to own more of the data, decision-making, and transaction path at the same time.

That makes this more than another earnings-call cameo for "agentic" software. According to The Motley Fool's transcript of Omnicom's first-quarter earnings call, Wren said the company had "executed real media buys for several clients using our agent framework, doing agent-to-agent buying." He also said Omnicom was "first to market" with AdCP, short for Ad Context Protocol, which he described as a standard evolving around this kind of buying.

AdCP is not just a slogan on a slide deck. The project's GitHub repository describes it as an open standard for tasks such as buying media, building creatives, activating audiences, and managing accounts across ad platforms. The repo says those tasks can run over two other agent plumbing standards, MCP and A2A. For non-specialists, the important point is simpler: AdCP is meant to give software agents a shared language for ad transactions, instead of forcing every buyer and seller to wire up a custom integration.

There is also a public example that the protocol's organizers say used real money months ago, though it is much narrower than proof of routine live buying. The AdCP site says the first agent-to-agent media buy ran on Oct. 16, 2025 using real money and inventory from LG Ads with human oversight. MediaPost reported that Omnicom described that event as a live transactional demo. But Digiday's later reporting on LG Ad Solutions said LG's own agents were not handling money. That does not disprove the October event. It does mean the public evidence is still organizer-described and murkier than Omnicom's earnings-call phrasing might suggest.

A launch coalition announced by Samba TV in October included more than 20 companies. That suggests Omnicom is attaching itself to an actual protocol effort rather than inventing a fresh acronym for ordinary automation. It does not independently confirm scale, savings, or how much of the workflow now runs without the usual intermediaries.

The pressure lands on the ad-tech tollbooths. If large agencies can route more buying logic, approvals, and inventory discovery through shared agent protocols and direct publisher links, some of the platforms that profit by sitting in the middle look less essential. That is especially awkward timing because, as Digiday reported, holding companies are already reassessing their dependence on The Trade Desk and other intermediaries.

Omnicom's case for why it can pull this off leans heavily on data quality. AdExchanger reported that Wren pointed to Acxiom's work in heavily regulated sectors such as finance and pharmaceuticals as evidence that the company has the accuracy and identity spine needed for AI-enabled buying. That matters because agent software that buys ads with bad customer data is just a faster way to waste money.

Still, most of the evidence here is Omnicom talking about Omnicom. The company has not named the clients using agent-to-agent buying, disclosed spend volume, or shown what share of the workflow actually bypassed existing ad-tech layers. The October AdCP example helps establish that the protocol effort is real. It does not establish broad production use.

So the story is not that advertising suddenly has autonomous media buying nailed down. It is that one of the largest agency groups is now openly treating agent-to-agent infrastructure as a way to squeeze the middle of the ad stack, while the public proof remains narrow and partly self-described. If that claim holds up, the next thing to watch is not whether more companies add "agentic" to earnings remarks. It is whether publishers, rival agencies, and the intermediaries collecting the tolls start naming what breaks when machines can buy from machines more directly.