Most AI Agents Can't Do Their Job—WPP Explains Why

Everyone in marketing has heard the pitch. AI agents will run campaigns, negotiate media buys, and adapt in real time. Boards are asking about it. Vendors are selling it. CMOs are putting it in their slides.

But at the IAB UK AI Growth Summit last week, WPP's chief AI officer said the quiet part out loud.

"The reality is that companies will deploy an army of agents across the organization, and forgive the technical term, but it's going to be a shit show," said Dr. Daniel Hulme. "Most of those agents are not going to be capable of doing their job."

The Gap Between the Press Release and the Reality

Hulme's framing is blunt, and he knows it. He describes the current state of agentic AI as the "teenage sex phase" of development. Everyone thinks everyone else is doing it. When you actually look, they're not.

This pattern isn't new. Programmatic advertising was supposed to transform media buying. It eventually did, roughly a decade after the press releases said it would. Data clean rooms were going to solve the loss of third-party tracking data. They're still mostly unlocking that promise. CTV measurement was going to bring digital-style accountability to television. Anyone reconciling a cross-platform report can tell you how that's going.

What's different with agentic AI, Hulme argues, isn't the hype cycle itself. It's that almost nobody is being honest about what deploying these systems at scale actually requires.

Why Testing Is the Real Work

Here's the part most companies aren't hearing. Hulme says at least 80% of the work required to build a functional AI agent is testing. Not building the agent. Not integrating it with your data. Testing.

That's an inversion of how most organizations budget and plan for technology projects. The assumption is that once you've built the thing, the hard part is done. With AI agents, Hulme suggests the opposite is true. The build is the easy part. Making sure the agent can reliably do its job is where most of the time and money goes.

And the testing problem in marketing is more complicated than it sounds.

The Second-Order Problem Marketers Are Ignoring

Hulme identifies what he calls the second-order problem, and it's the reason marketing-specific agent deployments are particularly difficult.

Here's how it works. You train an agent on historical campaign data. You deploy it to make decisions autonomously. The moment it starts acting, it changes the environment it was trained to predict. Consumers respond differently. Competitors adjust their bids. Media prices shift. The data the agent was built on no longer describes the world the agent is operating in.

"If I build a magic oracle and predict human behavior, I haven't changed that behavior," Hulme said. "Now I predict the models are out. You cannot predict the future based on the past."

This is not a software bug you can patch. It's a structural feature of agentic AI in any environment where the agent's actions affect the market it's trying to model. Marketing is almost entirely that kind of environment.

Marketers testing AI agents aren't just checking whether the agent follows instructions. They're checking whether the instructions still make sense after the agent has been running for a week, against a market that has already moved in response.

What Hulme Actually Recommends

Hulme's critique isn't that agentic AI is a bad idea. He's WPP's chief AI officer. WPP has deployed more than 28,000 AI agents across its operations. His argument is that the industry is currently running the weakest possible version of the technology and calling it a revolution.

Right now, he says, the industry is doing "very fast, very sophisticated rule following." The genuinely disruptive version of AI, in his framing, is systems that make decisions, observe outcomes, and adapt. The advertising industry is not there yet.

His practical advice is pointed. "The reality is that quick wins and low hanging fruit can be solved by a third party at a fraction of the cost," Hulme said. "You need to be focused on the problems that differentiate your business."

That's a different strategic question than "how do we deploy agents." It's "where in our business does agent failure actually hurt us, and are we prepared to do the 80% testing work required to prevent it." Most organizations haven't asked the second question yet.

The teenage sex phase of agentic AI will end eventually. Every hype cycle does. The companies that emerge with something real will be the ones that treated the testing as the product, not the afterthought.