Breaking down how to be ‘better safe than sorry’ with agentic AI in marketing

In November, the AI startup Anthropic disclosed that a group of hackers had used its Claude agent to target approximately 30 companies and government agencies, describing the incident as what it believed to be “the first documented case of a large-scale cyberattack executed without substantial human intervention.”

As brands continue to adopt Claude and other AI agents, many marketers are thinking through how to operate them in a safe, low-risk manner. Incorporating agents securely into workflows will require marketers to get their hands dirty playing with agentic AI and navigating the different ways it can be used to assist a campaign, according to experts Marketing Brew spoke with. At the same time, marketers will have to have thorough conversations about how agentic AI can be used in their campaigns.

Compared with GenAI, “agentic [AI] is a little bit newer, and I think there’s still a mix of enthusiasm and fear about what that might mean for the future of work as well as the future of marketing,” Andrew Frank, VP and distinguished analyst at Gartner, said.

Get down and dirty

Marketers need to be actively involved with AI overall to understand how to use it responsibly, said Nick Coronges, global CTO at R/GA. That involves building the right prompts and AI custom models to play with; agents can be used to check the output of a generative model. But building efficient agentic workflows requires design and marketing teams to be given control over outputs, Coronges said, while also learning to delegate to agentic tools.

“In the previous paradigm, you’d be able to go through and approve every single line of copy and read through every single asset,” he said. “Now you’re almost operating more like a game studio…To see how the game plays, you have to play it over and over and over again and see that it creates the right outcome. Each time you play it, it’s going to be something slightly different.”

R/GA has been deploying custom models for specific client campaigns using offerings already on the market that offer robust customization like Adobe Firefly and Black Forest Labs, he said. The agency also deploys bespoke AI strategies for various clients by only using AI in specific parts of the creative process. Customization allows for the agency to ensure the creative aligns with what clients are looking for while having control over the output, Coronges said.

For one client, AB InBev's beverage delivery platform Tada Delivery, R/GA used a hybrid agentic strategy to build a campaign narrative using a LLM and imported audio from AI voice generation firm ElevenLabs, but it created video used in the campaign through manually constructed illustrations. To further strengthen its AI capabilities, R/GA in July acquired AI system design shop Addition to build custom AI-enabled creative work for clients.

“A lot of the controls, a lot of the logic, and a lot of the brand tonality happens outside of the generative AI process,” Coronges said.

To deploy agentic AI in a safe and effective manner, marketers also need to determine their agentic structure and strengthen their data set, according to a LiveRamp blog post. There are several types of agents that brands can deploy, such as creative development agents and analytics agents, among others. A set of agents is also usually governed by a “superagent” that oversees all agentic operations and activates the agents under its umbrella.

Agentic operations can also benefit from clean data; brands can take actions like assessing their data signals and employing data clean rooms, the post read.

“Marketers need the right environment to be able to test these kinds of innovations without putting their organizations at risk,” Frank told Marketing Brew. “One of the near-term barriers that people are going to have to overcome is, how [to] give marketers tools to experiment with these new technologies without taking on risk or getting into trouble with agents that may not behave the way you expect them to.”

Communication is key

As marketers explore various agentic AI use cases, communication throughout the process remains important, Coronges told Marketing Brew. That means talking through what models are being used, how AI is being used to create specific content, and “de-risking” the technology overall, he added.

“The more you see it as a black box and something that you just don’t understand, the more risk there is,” Coronges said.