Digiday Programmatic Marketing Summit May Recap: How marketers are navigating agentic ad buying

But just as important as how marketers are adopting agentic AI are the rules and guardrails they are putting in place around their AI use, which was a major topic of conversation on and off stage.

Each Digiday event kicks off with the edit team breaking down current industry themes. Here’s where we think we are:

1. AI agents are participating in programmatic ad buys.
2. AI compute costs are a new ad tech tax.
3. Shared standards are necessary to avoid agentic ad buying repeating the problems of programmatic’s past and creating new issues.
4. Demand-side platforms and supply-side platforms are converging into unified ad platforms.
5. Programmatic supply chain consolidation changes power dynamics among intermediaries — including agencies.
6. Questions of transparency and control have never been more necessary and contentious.

Here are the topics we think the industry will spend the next quarter discussing:

Writing the rulebook for agentic ad buying
AI agents can help automate ad buying, but they should not be given full autonomy. Brands and agencies are putting guardrails in place. OMD, for example, sees the potential to allow AI agents to make campaign optimization decisions in the second half of this year, but “there’s still a need for a human to sign off on that work,” said the agency’s managing director of data and technology solutions Emily Proctor. Similarly, Bayer has worked with ad tech and data providers that use AI to optimize campaigns, but the brand has implemented spending caps to keep a leash on the technology, in addition to a rule requiring humans to oversee and sign off on any changes recommended by an AI agent. “Everyone will need to define those guardrails,” said Glenniss Richards, senior director of digital media activation at Bayer.

Agentic ad buying requires QA and homework grading
For everything AI promises in streamlining the ad buying process, it’s not yet as easy as inputting a single prompt. Agency execs cite things like AI hallucinations, such as incorrect CPMs, which could cost the agency and client money. Even a supposedly clear prompt can result in an irrelevant output, prompting quality assurance checks to grade the agent’s homework, execs said. Attendees learned that lesson first-hand when they were asked to negotiate an ad deal against AI agents during an interactive game. Human agency execs had to workshop prompts until it was accepted by the seller agents. On stage, Scott Ensign, chief strategy officer of Butler/Till noted a similar experience when the agency first used AI agents for ad deals. Marketers are using agents — but with training wheels.

Agencies’ agentic AI second brains
Data is only as good as what you can do with it, and AI agents are enabling agencies to do a lot more with the data they have available. KSM Media, for example, has developed a “librarian” agent that “is our coded institutional knowledge about clients” that can be used by employees as well as other agents, according to the agency’s vp and director of analytics and insight Henry Webster. Meanwhile, Huge has used “digital twins” that are modeled from actual consumer behavior and enable the agency to simulate marketing moves without risking alienating actual people, per its svp of data Luis Caballero. And WPP Media’s Open Intelligence platform has enabled the agency to take disparate sets of data and “turn these into models that predict and allow us to invest more securely,” said Amanda Grant, evp and global head of data and technology partnerships at WPP Media.

Transparency problems compounded
Marketers have long groaned over transparency issues. Usually, The Trade Desk is at the center of that belly aching. At DPMS, however, Google and Amazon were also scrutinized for their own transparency issues. The tech behemoths are swapping manual levers for algorithmic automation, pumping AI into already automated tools — The Trade Desk’s Kokai, Google’s Performance Max and Meta’s Advantage+ — that marketers deemed to be a black box. More automation hasn’t led to more transparency. Marketers say it’s unclear how agents make decisions, which could lead to low-quality inventory and wasted media dollars. Behind closed doors at the DPMS town hall, marketers said Microsoft automatically added new video placements without clear notice. On stage, speakers pointed to incorrect CPMs. The platforms’ remedy, agency execs said at DPMS, is to spend more money.

Job security
AI agents are taking up more duties. For example, Ellie Uberto, director of marketing at Duluth, said the brand has started using agents in the ad buying process. And marketers’ fears of AI replacing them in the workplace have been curbed, but not quelled. Many at this DPMS made the case for human oversight to account for hallucinations, quality assurance and prompt engineering. Questions linger, however, of what the role of a programmatic marketer looks like in five years. At the second town hall, one attendee said: “It’s not, you’re not going to lose your job to AI. You’re going to lose your job to the guy who knows how to use AI.”

Here are some of the sessions — and takeaways — we don’t think you should miss:

How one agency is evolving the human touch in programmatic

As agents take over more executions, evolving the role of the human programmatic trader as a “systems architect” has become an important imperative and skill set for agencies. Henry Webster, vp, director analytics & insight, KSM Media, shared what is taking up headspace at KSM in terms of balancing the human and machine in agency business and culture:

• “It’s not just about the buy. Being responsible for outcomes is still a part of it. It’s always been a part of it, but you need to act as the buyer, the architect, the analyst, in terms of performance.”
• “One of the things that’s really worked for us as an agency is a really democratic approach to agent development… so when somebody raises their hand that levels up, it gets a higher level of development, a higher level of scrutiny, and gets into sort of like the product suite of agents that then people can, in theory, begin to access as a group.”
• “I think the exercise that we did collectively yesterday [Digiday’s game] in terms of interacting with the agent to see if we could win — my team did not — was really good in terms of a demonstration. But that ‘show, don’t tell,’ always works a lot better in terms of bridging that language gap [between agencies and clients].”

Creative intelligence: agents at the dynamic edge

Luis Caballero, senior vp, data, Huge, addressed how agents use real-time and data-informed context not only to place ads in the ideal setting but also to influence which creative assets are served. Caballero shared how data can improve audience intelligence and how marketers use agents is evolving:

• “The technology and even the creative, they’re all starting to come together… with the agents that are now either being able to help them with their actual discipline, now they’re able to essentially create a continuous feedback loop to one another.”
• “I think the next exciting area that is starting to develop is understanding the why and the how, which is really getting into contextual signals, understanding, what’s motivating people, what’s either pushing them or pulling them to take notice of things and how you can then be able to engage with them at the right moment.”
• “It got to a point where we were able to really drive down the production cost, because we were identifying the actual creative assets that were already gravitating to the audiences that they had, and then identify the actual audiences that mattered to them as well.” [Discussing Huge’s agentic work with client PlayStation]

The lean machine: How Bayer uses agentic AI to scale global media

AI agents have become part of Bayer’s core toolkit to make ad buys faster and more efficient. Glenniss Richards, senior director of digital media activation at Bayer, talked about how the pharmaceutical brand’s 15-person in-house team leverages agentic AI tools via its DSP PulsePoint and other data providers to quickly pull data and analyze campaign performance — but not without human oversight:

• “We do have our foundational data providers that have tons of data around ROI performance across the entire business. Naturally, a bot would prioritize or optimize more toward those data providers or those particular channels. It actually became counterintuitive or counterproductive to my team where as they’re planning out testing new models or testing new partners, we weren’t spending. Then it actually became a time suck in those areas. That’s pretty much why I’ve come to the conclusion that I don’t want a bot.”
• “We do need some guardrails in place to ensure that we are still able to test and learn and scale new opportunities.”
• “There’s lots of hype where organizations are reaching out to us saying, ‘We can fully customize your journey, and you will only need one person to manage your multi-million dollar campaigns.’ Not true.”

Navigating the noise: Havas Media Network on agents and the supply chain

Transparency was a key theme at this DPMS. Supply side platforms are adding AI layers that promise to make bid requests smarter and more nimble. Greg Langer, vp of programmatic supply at Havas Media Network, argued that AI agents could add more complexities to the black box nature of the programmatic media buying process:

• “Agentic AI still hallucinates. There’s still that issue that if it doesn’t know the answer, it just makes one up.”
• “There’s no human emotion in it. They [the agents] don’t know the relationship of the historical with the client and maybe that partner… You end up calling up your rep and just say ‘Hey, let’s knock this out, instead of going back and forth with an agent.’”
• “It’s all of those guardrails that you historically will tell in the RFP process of all the KPIs — primary KPIs, secondary KPIs. It’s making sure those are in there…You’re only as good as the brief you provide with those guardrails, but also being nimble on that because certain clients do have different guardrails, so it’s not a one size fits all.”

The brand POV on agentic AI in programmatic advertising

Ellie Uberto, director of marketing at Duluth Trading Company, took to the stage to talk about agentic AI and programmatic advertising from the brand perspective. While Duluth hands off the bidding part of the ad buying process to an agency partner, the brand keeps creative — things like ethos and voice — managed internally:

• “Our brand voice is incredibly important to us — the way that we talk to our customer, our sense of humor — that’s something that we hold very close, and that would be a part of our flow.”
• “That [creative] is one part that we’ve held really close, whereas the way that we bid on our ads, the versioning of creative [assets], those sorts of things that free up a lot of time, that’s something that we are much more comfortable handing over to AI.”
• “On the brand side, we’re much more hands on with [AI]. For example, our creative team is leveraging AI on a daily basis as they are building out our creative. So there’s no more sitting down and building out individual versions of ads. That’s something that we entirely — not entirely, but AI is a much bigger part of that now.”

How to put agents to work buying ads programmatically

Butler/Till claims to be the first agency to do an agentic media buy. In this session, chief strategy officer Scott Ensign walked through the agency’s inaugural agentic ad buy in which it worked with PubMatic’s seller agent to source inventory. Then he outlined a second agentic ad buy in which the agency worked directly with Yahoo to test agentic ad buying in a direct-deal context. Now Ensign is looking ahead to how AI agents can help to automate buying processes for which existing programmatic standards are not suited to handle:

• “It was really fun to see the negotiation and interaction between our instance of Claude and [PubMatic’s] seller agent. We were quite impressed how quickly that worked, how well that it worked, and that we were able to create objects in what they’ve rebranded as their AgenticOS buying platform, and we saw really premium supply. We saw lower than expected CPMs, in part because this was a compression of that supply path, buying very directly and in a way that allowed us to get to market very quickly.”
• “People are probably familiar with this process: You go in, you say, ‘This is what I’m looking for.’ You send that out to half a dozen partners. You wait a couple days for the sales person to get back from vacation and email you back, and then you’re like, ‘well, that’s not anything that I was looking for.’ And you kind of go back and forth. This [agentic deal with Yahoo] came back with five ranked packages that we could choose from as quick as you get a response from one of the large language models. And again, that was really high quality contextual inventory.”
• “We want to do things that are addressable, biddable, optimizable, and I think agentic will take over that method of buying. Will it still be programmatic? I think so, because of the broad way that I view and think about programmatic. But I think it will look very different, both in terms of the supply paths, in terms of the pipes, in terms of the platforms, whatever a platform is in the future, based on what we’re seeing today. But I don’t think it will replace programmatic.”

How agentic AI Is rewriting programmatic data strategy

No one really likes working in a database — well, except AI agents. Which is convenient for advertisers and agencies that may maintain multitudes of databases connecting all kinds of data. In this session, WPP Media’s Amanda Grant unpacked how agentic AI is being used to stitch together disparate data sets and create predictive models for planning, activation, and measurement systems via WPP’s Open Intelligence platform:

• “There’s a lot of data, and a lot of data sits in a lot of places, and there’s workflows already that have been managing that historically. But with regulation, with us having our data at WPP Media, clients have their data. We’re working with partners and publishers with their data. We have performance data. There’s all this data everywhere, and it’s really important for us, from just a governance perspective, to have that in a system that’s safe and is data at rest and everyone’s in control of their data.”
• “We do have client-specific models, but we also have models that represent these different cohorts, and those go back to like it could be identity, it could be contextual, behavioral, action-based that the model is a model that kind of exists within a market, but then that’s a model that can be customized for a client need.”
• “Let’s say [you’re trying to market] soap to someone who is a female, 25 through 45 [years old]. You may have some keywords that you want to use to find that person. You might even have some interest targeting. That’s today. So embedding allows us to rethink, not just those very specific parameters, but in the context of these fundamental cohorts, which can be geo and demo, and then plugging in other things like behavioral, contextual, creative, what’s content that that consumer or your potential consumer is consuming that makes them a candidate based off of also a radius that you input and gets you closer to that group.”

OMD’s ‘getting started’ guide for clients applying agentic AI to programmatic advertising

It’s become increasingly clear that AI agents stand to play a role in programmatic advertising workflows. But where to start? And where to go from there? In this session, OMD’s Emily Proctor broke down the agency’s playbook for getting clients on board with agentic AI in their programmatic advertising practices. The lowest-hanging fruit is using AI agents to help speed up planning cycles. But even then, an important aspect of any agentic workflow is keeping humans in the loop:

• “At the end of the day, it’s garbage in, garbage out. So if you don’t have really solid human inputs into the agent or the workflow, it’s going to break down. And so it’s about understanding where to work faster. We still need to be thoughtful about the inputs, but once we have those solid inputs, then the AI can help the rest of the workflow flow faster.”
• “While AI can help us QA a campaign or whatever, it does not negate the need for a human to have eyes on it before it goes into market. So it really comes down to really being responsible in how we use it and making sure that we know exactly where in the process we do want human intervention to make sure that we’re not just completely reliant on it.”
• “Having those manual skills, or at minimum understanding what it takes to get it done manually, because you can’t fully appreciate the benefits of the efficiency that you’re going to gain from these agents if you don’t understand how it’s done manually.”

Couldn’t be there? Here’s our video on how it worked.

Brand and agency executives didn’t just talk about agentic ad buying during the Digiday Programmatic Marketing Summit. They also got hands-on with it.

In a live interactive workshop, attendees were organized into teams, presented with an inventory package and a $6,000 per-team budget, and tasked with using AI agents to win the most ad impressions at the lowest price.

Teams drafted initial prompts to instruct AI buyer agents to create offers to submit to an AI seller agent. The AI seller agent then countered each team’s offer, to which teams had to draft prompts for the AI buyer agents to respond to the counter-offers. The AI seller agent then reviewed the revised offers and decided how impressions should be allocated and at what price.

Congratulations to Team 1 for winning the game. Their prompts — including the buyer agent’s resulting messages to the seller agent and the seller agent’s counter-offer and final response — are below.

Inventory package:

• Seller: Netflix
• Channel: CTV
• Inventory type: 30-second ad
• Available impressions: 300,000
• Floor CPM: $30

Team 1’s initial prompt to buyer agent:

“I’m a media buyer looking to buy inventory with Netflix looking for 300,000 impressions with the CPM under $30. Pursue a deal that beats campaign objectives of completion rate for entertainment enthusiast who are 18 to 49 years old. Thinking behind the offer is to use Netflix platform to achieve the most efficient in scalable reach. We would also like added value.”

Buyer agent’s proposal to seller agent:

“This proposal targets a core demographic of entertainment enthusiasts aged 18-49, utilizing the Netflix platform to achieve high completion rates and efficient, scalable reach. We believe this audience segment is perfectly aligned with Netflix’s premium streaming entertainment inventory. We are looking for a deal that reflects this strategic alignment and includes added value to support our campaign objectives.”

Seller agent’s counter-offer to team:

“The requested CPM is currently below our established floor for premium inventory. To move forward, please adjust your pricing to meet or exceed $30.00 and refine your targeting with more specific consumer intent signals.”

Team 1’s response prompt to buyer agent:

“We would like to target cord cutters, with HHI of $100K and with a CPM of $35. We are willing to make a deal once you accept ours.”

Buyer agent’s revised proposal to seller agent:

“In response to your feedback regarding consumer intent signals and pricing, we have refined our proposal to target high-value cord cutters with a household income of $100k+. To reflect the premium nature of this audience and the Netflix environment, we have increased our offer to $35.00 CPM, which is well above the floor and exceeds your previous suggestion. We believe this revised targeting and premium pricing offer a highly competitive deal for this CTV inventory.”

Seller agent’s final response:

“Your revised CPM of $35.00 and the addition of HHI targeting make this a standout proposal. We are pleased to accept this high-value bid.”

Deal result:

171,428 impressions at $35 CPM

Sourced from town halls, sessions and hallway conversations at DPMS:

• “I feel like our [TTD] reps are outreaching more often and trying to assist on strategy more often, since [the news in March of Dentsu, Publicis and WPP advising clients against spending with The Trade Desk overall or on OpenPath specifically], partly because they want to keep the business from the agencies.”
• “You can pull the data [through TTD]. It is transparent. It’s just really, really complicated to literally see where does everything in my media supply chain go. You have to pull three different reports and pivot them together.”
• “You talk about transparency, it’s putting the [Sellers and Publishers 500+ marketplace] on every campaign by default. And your traders don’t even know it, because Kokai is so shitty that when you look at it, you’re trying to read it and make sure your settings are all good, and you have to uncheck [SP500+] in three different spots. That’s going to annoy people.”
• “Other DSPs now have reached out and kind of said, like, ‘Hey, if you’re looking to shift money, we’re transparent.’ And one thing we’ve challenged them with is every DSP now has their own bid optimizer, and they say it generates X amount of savings, and if it doesn’t work, you don’t get charged for it. Well, that’s fundamentally what a DSP should do to start with. They just added a little button now they can charge more margin on it. They said, ‘Well, trust us. You know, if it doesn’t work, you don’t get charged. All right?’ Can you prove it in logs?”
• “Google, forget it. Google’s never been transparent about anything. They’re always like, ‘let me get back to you on that.’”
• “[PMax] is unsold inventory. They’re giving us the shit inventory, and that’s what we’re buying. So it doesn’t work. And so we’ve not seen any success with it. Clients ask for it because it’s a sexy thing, but no one sees success with it.”
• “It’s a hard no for [Meta’s] Advantage+, because the social platform, specifically Facebook, did this thing where they removed our capabilities of targeting specific type of audiences due to sensitivity, but now you roll out a new capability that says, ‘Hey, we can look at your existing campaigns and optimize and tell you who you should target.’ It’s a hard ‘no,’ and it actually did not perform well when we actually tried to test it against one of our other audiences, to where our costs did go up and the engagements weren’t there.”
• “[Microsoft] did a rollout about three or four weeks ago, and it a video placement within their ecosystem you could not opt out of. It just was out there. So in pharma, that’s actually a big no-no, because it changed up the way our ad looked and a regulatory person brought it up. Now I also know that I did not see any publication, so they did not outwardly tell people that it was done. So our campaign on Microsoft was paused for about two weeks because we found it. They now are giving people an opportunity to opt out of it, but then I don’t know how transparent that is.”

1. Agentic ad buying is not so simple as inputting a single prompt.
2. Quality assurance and campaign reporting are two areas of programmatic workflows most suited to AI assistance.
3. AI governance and controls are crucial; so are contracts.
4. AI is compounding programmatic’s transparency problems.
5. Billing models and employee incentives need to be updated for AI-powered workflows.
6. AI copy-and-paste will get you replaced.

Marketers put up guardrails as AI agents reshape programmatic buying

As AI agents enter programmatic advertising, marketers are adding guardrails to maintain oversight, transparency and control.

Why Duluth trusts AI agents with bidding, but not brand storytelling

Brands like Duluth are in the middle of the agentic debate, leveraging AI for automated ad bidding while keeping brand voice human-led.

‘Google doesn’t care that it’s terrible’: Brand, agency execs air frustrations with The Trade Desk, Google’s Performance Max, Meta’s Advantage+

Think transparency is hard to come by in programmatic advertising? Well, get a bunch of brand and agency executives in a room, and they’ll get super transparent about how opaque the digital ad market has become.

Read all Digiday Programmatic Marketing Summit coverage.