Using machine learning to automate and optimize ads is familiar to publishers. For years, this tech has helped publishers tune floors and serve the best ads behind the scenes.
Now, publishers are exploring automation and optimization through agentic AI. Many publishers have identified a similar use case: sales agents who actively make decisions in the auction on behalf of the publisher.
“Essentially, a sales agent is powered by an LLM model, and it effectively sells our inventory on our behalf to buy-side agents,” said Vikesh Chevli, director of ad tech and programmatic strategy at News Corp.
Industry collaboration is already underway to move forward with this idea. Following AdCP’s donation of a sell-side agent code to Prebid in January, publishers now have access to an open framework to test agents for themselves.
“We’re still quite early,” said Chevli. “We’re in the design phase and scoping out how all of that looks.”
One possibility is that AI could handle smaller clients, giving them the kind of hand-holding that’s difficult to support for customers spending a small amount of money.
Agents could handle smaller, highly segmented deals that traditional sales teams often ignore due to high transaction costs, said Patrick McCann, senior vice president of research at Raptive and a participant of the Prebid sales agent collaboration with AdCP.
“There are countless opportunities that never get executed because they’re too small or customized,” McCann said. By automating discovery and execution at scale, agents could reduce friction for both publishers and buyers, potentially limiting revenue lost to intermediaries.
As companies build these initial sell-side agents, some are using AdCP’s source code and some are not. Here’s how some early adopters, from publisher Weather Company to data clean room Optable, are testing our seller agents.
How The Weather Company built a sell-side agent
The Weather Company built the first version of its sell-side agent using the open framework developed through AdCP, said Dave Olesnevich, VP of weather data and advertising products at The Weather Company.
“The goal of the first release was to test whether automated agents could help media buyers more easily discover available ad placements and assemble campaign proposals,” Olesnevich said. The publisher is heavily involved in industry working groups to advance agentic protocols, such as the Prebid Agentic Task Force. For its first Sales agent, it worked with Scope3 and its own ad product, ad tech and ad ops teams.
To keep the rollout controlled, the first version of The Weather Company’s Sales Agent includes only a small set of the company’s top-performing ad formats and audience segments. Access is restricted through authentication and permissions so only approved partners can explore this inventory through the agent.
“We didn’t try to boil the ocean on day one,” said Olesnevich.
The publisher selected a small set of top-performing, high-viewability ad formats and paired them with a defined set of premium audiences. Afterward, they plugged that data into the agent. And when a buy-side request comes in, a human validates the availability and pricing given by the sell-side agent.
“I have premium ad formats. I have premium audiences. I have fantastic creative,” Olesnevich said. “That often gets lost in an open auction.”
The goal is to make those assets discoverable to buy-side agents so planning becomes more effective and activation faster, he added.
Previously, much of that premium inventory would have been exposed – and often buried – inside standard open-auction programmatic flows through SSPs and GAM. Now, Olesnevich is testing a sell-side agent that aims to make those offerings explicitly discoverable in a more request-based flow.
How Optable tested sell-side agents
Ad tech companies that work with publishers are also testing out sales agents. Data clean room Optable, for example, is building its own agents directly into its existing data platform. Optable is using a sell‑side agent for its data‑owner customers – often publishers and media companies, but not exclusively – to package their audiences and inventory for buyers who access those products through agentic workflows.
Inside Optable, that work starts with the company’s audience agent. A salesperson can, for example, upload an advertiser’s RFP and ask the system to suggest audiences that fit the brief.
Anthropic’s Claude is the front end for that process. Optable connected Claude into its data environment so when someone drops an RFP into Claude and asks for a plan, Claude hands the request off to Optable’s tools. Those tools route it to an AI model running inside Optable that understands how each publisher’s data is structured.
The agent returns with a mix of new audience ideas and existing segments it thinks might fit. Then, at the user’s request, it can create and activate those audiences and sync them with a publisher’s preferred ad platform, like Google Ad Manager.
That audience‑building layer is the foundation on which the sales agent sits.
The sales agent focuses on packaging and selling that data. Inside Optable’s dashboard, publishers can see a sales agent section where they can define specific products, such as a CTV package built from a publisher’s first‑party audience segments. Then, publishers configure how these data and media packages are priced and presented to buyers inside the agent. They can also approve which buyers are allowed to interact with their data through the agent.
For instance, Optable Co-Founder and Chief Product Officer Bosko Milekic demoed a potential campaign, acting as the role of the buyer, where he described for the agent a fictional brand, who the brand wants to reach, preferred media types, and set a $1 million budget. Claude, connected to Optable’s sell‑side agent, turned that brief into a plan that listed which sellers and networks were relevant, how much would go toward data fees, how much toward media and how those campaigns would be executed.
When he asked to focus only on connected TV, the agent reworked the proposal around that constraint.
“Now it’s coming back with a refined proposal on how to allocate my million bucks,” Milekic said. “It’s saying, ‘Here’s the Family Network and FreshCard Media that have relevant products for you. Here’s the cost of the data, here’s the cost of the media.’ In this case, it’s proposing to execute these products through deals.”
Using the buyer’s existing demand‑side platform and ad server accounts, the system can move beyond a proposal and actually set up the programmatic deals in the exchange, so they’re ready for the buyer to start running campaigns. On the publisher side, those deals appear inside Optable as a list of incoming opportunities associated with their data.
While Optable counts itself as a supporter of Prebid, which now owns the sell-side agent code, it didn’t use the source code for this specific project. The data clean room platform Optable, for example, which works closely with both Prebid and AdCP, did not use the source code.
“We are involved in the Prebid initiative,” Milekic said. “They’re assessing the code base, and there’s been some discussion around whether to start from scratch or to try to incrementally improve what was donated.”
Where do sell-side agents fit in the stack?
One of the biggest open questions for publishers is where the sell-side agent will do its work: in direct buys or programmatic ones?
“I kind of see it as a third channel. It transacts within the two existing channels – direct and programmatic,” said News Corp’s Chevli. For publishers, that means the agent could insert and manage direct deals within the ad server or facilitate programmatic deal negotiations across SSPs.
For most publishers, the revenue hierarchy remains familiar: sponsorships at the top, direct deals below, programmatic guaranteed and nonguaranteed next and open marketplace at the bottom. Chevli expects agents to eventually touch all of those layers.
“A lot of the focus has been on how we get an agent to work with Google Ad Manager [GAM],” he said. “But I see it covering all of those bases.”
At The Weather Company, experimentation is happening alongside its existing stack.
“For instance, the agent can tap into GAM for inventory and forecasting data, using that information to shape its recommendations and negotiations,” said Olesnevich.
He added that the Weather Company agent stops short of a full end-to-end integration with GAM, leaving the last leg of campaign setup and approvals to human operators for now.
Sell-side agents also have the potential to add a conversational layer to programmatic, said Milekic from Optable. He explained that traditional programmatic pipelines optimize for price and scale but lose nuance in buyer-seller intent.
With agents, buyers could communicate campaign goals or audience needs in natural language, and the agent could decide how to respond in real time. This intent-driven layer is what makes experimentation promising, added Milekic.
While still a long way from standard operating procedure, sell-side agents are at the beginning of a pilot stage as they aspire to become part of the everyday machinery of how ads get sold.
