Mythbuster: Here’s what ‘agentic’ AI actually means for advertisers, agencies and publishers

Though the tech is clearly still in a larval stage, there’s real implications for brands, agencies and publishers relating to it. But in order to see through chatbots masquerading as something more, it’s important to understand the differences between agentic systems, AI agents and the various use cases already being applied at agencies, brands and publishers.

Wait, aren’t these just AI agents?

Not quite. An AI agent uses generative AI, typically via a large language model (LLM), to perform a task. OpenAI, the company behind ChatGPT, has been developing an agent called “Operator,” which promises to help users make restaurant reservations and book travel arrangements. 

“The world of generative AI, [of] ChatGPT, was about questions and and answers. Now we are in the world of action,” said David Raichman, executive creative director at Ogilvy Paris, where the agency network operates a special AI unit.

The difference is one of scale and process. Agentic AI describes a situation where multiple AI agents work together to complete complex tasks, with minimal oversight or intervention from a human user.

“Agents are AI models with a job. Agentic means those agents can work independently of humans for a certain length of time,” said Wesley ter Haar, co-founder and chief AI officer at Monks.

It’s also different from automated workflows (which only perform tasks in a prescribed manner), because the AI agents are able to write the sequences they’ll use to carry out their objectives. To quote the definition offered by Anthropic, the company behind LLM Claude, they’re “systems where LLMs dynamically direct their own processes and tool usage.”

“You give it a task, it works out the steps it needs to take to do that task, and then generates a result based on that,” explained Peter Gasston, innovation lead at VCCP’s AI-devoted Faith division.

So how might they be used in marketing?

A recent, experimental project produced by Monks, via a partnership with Nvidia, for apparel brand Puma provides the clearest demonstration of agentic’s applications. The agency produced a 30-second brand film in a “fraction” of the time normally required, using up to five agents working only from the client briefing document.

“It’s written by AI agents. It’s mood-boarded by AI agents. The director of photography is an agent, and those agents have worked together to pull that script together, pull that mood board together, and give that back to the people that are in the loop,” said ter Haar. The level of human intervention in such a system could differ, depending on a client’s legal compliance requirements, he added.

Still, Adobe’s battery of agentic tools included audience and data insights applications, including one that can create entire audience segments for a planner. Agentic use cases for media planning are not as far along as those in creative – although ter Haar said Monks is making frequent use of AI audience profiling tools.

The agency had developed a media mix modelling tool, dubbed “Clarity”, ter Haar explained. “It can do marketing mix modeling at massive scale by having thousands of AI agents predict, and then the overlap of the highest performing predictors becomes your media plan,” he said, adding that Monks has yet to roll out Clarity “at scale” for its clients.

What about publishers?

The uptake of agentic AI among publishers remains nascent, to say the least. For many newsrooms, the risk of hallucinations is still too great for it to be deployed widely. But there are some interesting applications cropping up.

For instance, an “SEO agent” can be used on the back end of a publisher’s content management system (CMS) to automate suggestions on how to get an article search-ready. At the click of a button, a journalist can see alternative headlines, descriptions, and keywords. 

Meanwhile, a “drafter agent” refers to an application that allows a journalist to take all the raw materials they’ve assembled when reporting a story – audio transcriptions, video or text – and spin it into an initial draft, per whatever prompts they’ve used. If the writer is happy with the results, they can then continue adding prompts to bring in other data or context until the article starts to take shape. They can then use the SEO agent to finesse it, and create multiple versions for different sites or social platforms. All the information is kept within the publisher’s domain, so sensitive source material isn’t shared externally.

Where’s this all ultimately leading?

Jargon aside, agentic AI’s promise is an optimist’s view of an AI-enabled future – one where time savings are reinvested back into higher order tasks, allowing creatives and media planners to spend more time thinking and less time grinding at their keyboards. But that’s a long way off. 

Ter Haar told Digiday he expected media planning to eventually make heavy use of agentic systems. “Let’s call it a year, year and a half, maybe,” he predicted for when it might take serious root.  

While publishers are interested in discussing the potential for agentic AI, significant hurdles remain — especially as they grapple with unresolved copyright disputes and shifting referral traffic dynamics. And until AI’s kinks and hallucinations are ironed out, scalable use cases will remain theoretical. 

“There is a trust gap to be solved before this really gets deployed,” said Jon Roberts, chief innovation officer of Dotdash Meredith. 

Trust issues might prove sticky, but the central promise of agentic will likely tempt brands, agencies and media owners to keep pushing further into this new terrain. 

Said VCCP’s Gasston: “You can delegate a task to it, leave it to run and go away and come back, and it’s given you a result … I’m not sure how close we are yet to that being a reality.”