PubMatic AI Agents Run 14% of Revenue. Nobody Can Say Whose Side They Are On
PubMatic just told investors that AI agents are now running 30 live ad campaigns and handling 14% of its revenue. Nobody outside the company can explain whose interests those agents are optimizing for.
That announcement came in an earnings call on May 7, when CEO Rajeev Goel framed the milestone as the start of the third transformation in advertising technology, after real-time bidding and the mobile shift. The question the call did not answer: what happens when an autonomous agent inside a live ad auction makes a decision that a buyer or publisher wants to audit? No brand, publisher, or agency has yet published a clear answer in public. According to the earnings call, more than 30 campaigns were running on autonomous agents by May 7, alongside more than 1,000 AI-powered deals transacted and over 20 operational AI agents in production.
What the industry has not built is a shared answer to who is accountable when an AI agent makes that call. Under typical agency law, when an AI acts on behalf of a retailer, the retailer bears responsibility, not the technology provider. But ad tech agents operate across a chain of intermediaries— PubMatic's platform sits between the company placing an ad and the publisher selling the space—and that chain has no agreed rulebook for what happens when something goes wrong. Buyers and publishers have no common reference point for what audit rights they actually have, and no regulator has stepped in to define what "acting in the interest of the party who deployed me" even means in this context.
PubMatic is not alone in needing to answer that question. The company is the most advanced publicly-tracked example of a pattern playing out across the programmatic ad stack: AI agents are being inserted into the millisecond decision chain between a brand deciding to advertise and a publisher getting paid. AgenticOS, PubMatic's agentic platform, sits between buy-side and sell-side systems— the software buyers use to decide where to spend ad money and the systems publishers use to offer their inventory— executing campaigns autonomously under goals set by humans. Its agents (Audience Discovery, Inventory Discovery, Campaign Insights, Media Activation) operate on their own timelines inside a proprietary GPU environment. Butler/Till, a connected TV buyer, reported a single agentic campaign that cut buy-side fees by more than 80%, delivered 40% more impressions than planned, and drove a 30% lower effective CPM, according to PubMatic's investor announcement. Brkthru, another buyer, said it now manages media for more than 1,000 brands across 235 agencies using agents that execute campaigns across 25 buying platforms simultaneously, according to the same investor announcement.
Those numbers are real. The accountability structure behind them is not clearly defined anywhere the industry can point to.
"Most of the market still does not understand how agentic advertising works in practice," according to The Current, which surveyed agency and publisher sentiment in late April. "The conversation is noisy because the industry is squabbling over standards, nearly every company started claiming agentic capabilities without credible specifics." An industry veteran quoted by PPC.land put it more bluntly in October 2025: the Ad Context Protocol, one of the competing standards for how agents talk to each other, was "the tail wagging the dog," pushing automation before fixing structural problems in the supply chain.
The structural problem, as Women in Retail and The Lyon Firm both describe it, is that autonomous agents in ad tech create a gap between who is responsible and who is actually acting.
"When something goes wrong, regulators and plaintiffs will ask: What did the AI decide?" Women in Retail noted in a May 7 analysis. The Lyon Firm, a product liability practice, frames the question across three available tracks: design defect, manufacturing defect, or failure to warn, as categories that apply to software agents when something goes wrong at scale.
The gap, as Women in Retail and The Lyon Firm have both noted, is that contracts in the current market are largely silent on AI agent scenarios. When an agent bids on brand-unsafe inventory, or reallocates budget in ways that violate insertion order terms, most agreements don't specify who is liable: the buyer, the seller's platform, or PubMatic or whoever built the agent layer. The silence defaults to the advertiser.
What nobody has published, in any of the press releases or earnings calls or vendor pitches for agentic advertising, is a clear answer to what audit rights a brand or publisher actually has when an agent makes a decision they want to review. Can a buyer pull the decision log for an individual impression? Can a publisher see which agent bid on its inventory and why? The sources reviewed for this article did not turn up a standard answer.
The standards picture is not helping. PubMatic published the first open agent-to-agent specification in September 2025, then helped launch the Ad Context Protocol in October alongside five other companies. The IAB Tech Lab has its own competing framework, ARTF. The two groups have different views on how agents should discover each other, negotiate, and report back. Buyers and publishers trying to evaluate whether an agent is acting in their interest have no common reference point, and no regulator has yet stepped in to define what "acting in the interest of the party who deployed me" even means for an ad tech agent.
Amnet France ran what PubMatic described as the first agentic campaign across a major holding company's systems, according to the April 27 investor announcement. The details are thin: geographic targeting, some brand-safe parameters, a media plan executed automatically. Nobody has published the decision tree the agent used. Nobody has published whether a human reviewed the outcome and, if so, what they found.
PubMatic's earnings gave the category its cleanest revenue validation yet: 80% growth, 14% of revenue, 30 live campaigns, 20 agents in production. Two questions the category still has to answer: who these agents work for, and who is accountable when the decision is wrong.
What to watch: whether any brand or agency publishes a post-mortem on an agentic campaign that did not go as planned. That document, if it surfaces, will tell the industry more about how the accountability gap actually works than any press release has so far.