The agentic AI payoff is knowing when not to use it

IAB Tech Lab CEO Anthony Katsur argues the winners in agentic AI will deploy it with a hybrid architecture, routing predictable, rules-based work to deterministic software and reserving AI for ambiguity and unstructured data. He cites McKinsey, BCG and Gartner agreement that the single biggest lever CIOs have on AI spend is cutting token usage, saving 20 to 30 percent on AI costs, with layered architectures slashing blended token costs by as much as 87 percent versus a single frontier model. The sharpest line is his warning on the Jevons paradox: the more efficient AI gets, the more of it companies will use, so spend only scales if the return is 10x or 100x, not incremental. "There's no faster way to lose client trust than to get the numbers wrong," he writes, directing that agents be used for insight and recommendation, never for financial calculation, booking or billing. He also flags OpenProposal, a feature in IAB Tech Lab's updated AAMP 3.0 protocol, which aligns how buyer and seller agents evaluate proposals against existing media buying standards.

The crux

Agentic AI only pays off when it is pointed at high-frequency, ambiguous tasks and kept away from the deterministic transaction plumbing, so the discipline of knowing when not to use AI is what separates a return from a token bill.

Sell-side read

For publishers this is a yield and ops message: put agents on the report synthesis and cross-platform optimisation where pattern matching earns its keep, and keep deterministic software on pricing, booking and billing so a wrong number never costs you client trust.

Watch

OpenProposal's adoption inside AAMP 3.0, which would standardise how publisher-side and buyer-side agents negotiate, and whether it ships before the end of the year.