What happens when agency execs negotiate against AI agents - Digiday

The simulation runs on Google Gemini. One agent is the buyer. Another is the seller. Somewhere in between, a real ad deal is supposed to happen. Tim Peterson, a reporter at Digiday, built the thing to ask a question that none of the campaign metrics can answer: what does it actually feel like to sit across the table from an AI agent trying to cut a deal?

The exercise — reported by Peterson at Digiday (Digiday+, subscription required) — put two practitioners through the scenario. Skyler McGill, head of video and programmatic at Wpromote, an independent digital marketing agency, and Ryan Lammela, VP of channel activation at Butler/Till, an integrated marketing agency. Both are launch partners on PubMatic AgenticOS, the programmatic ad platform company's agent-to-agent buying infrastructure, and both have already run live agentic media buys. This is not theoretical for either of them.

That last part is what makes Peterson's subjects interesting. In January 2026, as MediaPost reported, McGill called agentic programmatic "the biggest transformation in programmatic since RTB" — real-time bidding, the auction-based system that became the backbone of display advertising in the early 2010s. Both agencies had already been testing agentic tools: as Sam Bradley reported at Digiday in December 2025, Butler/Till was testing a Scope3 buying agent built on Claude against a PubMatic counterpart agent using AdCP, the emerging agent communication protocol, while Wpromote ran parallel experiments. These were the people Peterson chose to sit across from his agent.

The further proof of concept was Butler/Till's Clubtails campaign. As Sam Bradley reported at Digiday, Butler/Till completed a live agentic media buy for Geloso Beverage Group's Clubtails brand between December 2025 and January 2026, running a Claude-based buying agent against PubMatic's counterpart selling agent. The reported results: an 82 percent supply chain cost reduction, 40 percent more impressions, a 30 percent CPM reduction, a 98 percent video completion rate, and a less-than-1 percent made-for-advertising inventory rate. Lammela, who is VP of channel activation at Butler/Till, is one of the two people in Peterson's simulation. He has seen the numbers in production.

The dependency chain on that Clubtails campaign is worth spelling out: AdCP runs over PubMatic AgenticOS, which connects Butler/Till's Claude-based buying agent to PubMatic's selling agent, bypassing the traditional DSP layer entirely. The DSP bypass is the headline performance number. It's also where an accountability gap opens up. When the demand-side platform — traditionally one of the audit layers in a media buy — disappears from the chain, responsibility for decisions gets murkier. One structural detail that most coverage buries: Butler/Till's campaign ran as a private marketplace and direct buy, not open-auction RTB. That matters because the hardest version of the latency problem — LLMs responding within the millisecond windows that open-auction bidding requires — wasn't actually tested.

Peterson's simulation is trying to surface the qualitative friction that metrics can't capture. Every other piece of coverage in this space is infrastructure reporting: AdCP spec updates, AgenticOS launch partner lists, benchmark numbers from live buys. Peterson is asking something different. When a human sits on one side of a negotiation and an AI agent sits on the other, what happens to the deal-making instincts that experienced media buyers have spent years developing? Can a buyer tell when an agent is making a bad trade? Do they trust themselves to catch it in real time?

The counterargument here is well-documented. At Digiday's Programmatic Marketing Summit in New Orleans in December 2025, agency executives gathered under Chatham House rules — meaning the discussion was not for attribution — and reached a rough consensus, reported by Tim Peterson at Digiday: agents should stay away from the actual transaction point for now. The concerns were specific — hallucination risk, the fundamental speed incompatibility of large language models with RTB's millisecond auction windows, and an accountability gap that existing contracts and SLAs don't address. The town hall's rough consensus on timeline: three to four years before agents handle actual transaction execution without a human in the loop.

Christopher Francia, founder of Attention Arc, an ad tech consultancy, made the case most plainly in a December 2025 piece by Peterson: LLMs can't operate at RTB speeds, and when something goes wrong with an AI-driven buy, nobody in the existing accountability structure knows whose fault it is. Opaqueness as a feature is a reasonable critique of current LLM systems; opaqueness in a supply chain that moves hundreds of millions of dollars is a different kind of problem.

The governance infrastructure needed to close that gap is still being built. Ronan Shields at Digiday reported in January 2026 on the emerging standards debate around AdCP and the Agentic Advertising Organization (AAO), whose board includes Ruben Schreurs of Ebiquity, a global media auditing firm. Anne Coghlan, COO of Scope3, a supply chain transparency company, and John Goulding, chief strategy officer at MiQ, a programmatic media company, were among the participants. "You can't have AI agents negotiating multi-million-dollar strategies without systems that track decisions and outcomes with true precision," Schreurs said, according to Shields' reporting.

Peterson's simulation — set to be playable live at the Digiday Programmatic Marketing Summit in Palm Springs, May 6 through 8 — is a pressure test run against exactly that gap. It doesn't prove agents can negotiate well. It asks whether people who do this work can tell when an agent is making a bad deal, and whether they trust themselves to catch it. That's a harder question than whether the pipeline runs.

The honest read on where this stands: PubMatic AgenticOS shipped real infrastructure and ran real campaigns with real launch partners. Butler/Till's Clubtails numbers are not marketing copy — they are reported results from a completed campaign. But the results came from the easier version of the problem. PMP and direct deals sidestep the latency issue entirely. Open-auction agentic buying at scale, with full accountability trails, remains unsolved. Peterson's simulation is trying to measure the human side of that unsolved problem. The question his experiment is actually asking: are the practitioners who ran the campaigns confident enough in agent behavior to let agents run the negotiation too, in real time, with real money on the line?