Scaling an agent is what invalidates the pilot that justified it
WPP's Chief AI Officer put a name on the dirty secret of agentic AI: the teenage phase. Everyone claims to be doing it. Almost no one actually is.
The interesting part of Daniel Hulme's argument isn't the metaphor. It's the second-order gap underneath it.
Train an agent on last year's campaign data. Deploy it. The agent's own decisions move prices, change consumer response, provoke competitor reactions. Within a week, the market it was trained for no longer exists — and the agent is now making bets against a world it can't see.
This breaks the traditional pilot-measure-scale playbook for CMOs. Every ROI case built on a six-week pilot expires the moment you scale, because scaling is what invalidates the training distribution.
The CMOs winning the next 18 months aren't the ones with the most agents in production. They're the ones who treat continuous evaluation as the product — not the agent.
If your AI roadmap measures deployment count instead of post-deployment drift, you're tracking the wrong number.