The future of agentic AI will not be prompt-based

Every agentic tool being pitched to marketers right now works the same way. You describe what you want and the agent attempts to deliver it. Better briefs in, better campaigns out. It feels like magic the first time. But this is the 1.0 version of what’s possible, and the industry has mistaken it for the destination.

Almost all of the AI tools built for the open web use the same generally available large language models that power ChatGPT and Claude under different logos. LLMs do one thing brilliantly: They learned to read and write by studying everything published on the internet. With that, they’ve rightly earned their place in the stack to generate ad creative, draft copy, speed up productivity and orchestrate complex workflows.

But think about what actually drives your results as a media buyer. It isn’t language. It’s knowing who is likely to buy, when they’re in-market, what they’ll pay and which touchpoint moves them. That is a prediction problem about human behavior; a model that learned everything from text has never actually seen any of it.

But it’s not merely an audience problem. That familiar AI chat interface was designed for queries, not marketing briefs. Every marketer knows the pain of a brief gone wrong. Despite knowing what you wanted, the words didn’t quite capture it, and six weeks later the results came back sideways. Prompting can further complicate this process. Whatever you type is what the agent optimizes, so the gap between what you meant and what you wrote becomes the gap in your results. Your outcomes now hinge on your ability to phrase things well, which is a strange skill to bet a media budget on.

The next generation of AI fixes both by eliminating the prompt entirely. Rather than writing paragraphs describing what you want, you’ll point at the KPI that matters most: a purchase, a new account, an upgrade, and tag it with the pixel already on your site. The intelligence that event trains lives with your brand, not behind a platform’s walls; and it follows your spend across the open web and CTV.

Behind that pixel sits a different kind of AI, a foundation model trained not on behavior instead of text — real transactions, search activity and ad engagement at population scale. The same way language models learned to write by reading, behavioral models learn to predict what consumers will do by studying what consumers have actually done. When you designate a purchase event as your goal, this kind of model already understands the patterns that lead to a purchase like yours. The agent can autonomously plan, buy, measure and optimize toward your configured goals rather than its best interpretation of your paragraph.

This completely changes how a brand or agency can successfully compete in the AI era. Working with a leading performance agency, my company, Yobi, deployed this AI strategy against a configured purchase event collected by The Trade Desk. The result: two times the Google-measured revenue on The Trade Desk’s CTV inventory versus the brand’s YouTube campaigns — with half the budget.

The decisions that actually determine performance, whom to reach and how, belong to a model that was built to make them.

So here are the questions to ask the next vendor who demos an AI agent: Which AI model is being leveraged, and what data was used to train it? 

This op-ed represents the views and opinions of the author and not of The Current, a division of The Trade Desk, or The Trade Desk. The appearance of the op-ed on The Current does not constitute an endorsement by The Current or The Trade Desk.