The Upper-Funnel Boundary Agents Can’t Cross

By Savannah Westbrock, VP of Commercial Strategy & Growth, Inuvo, as seen in The Drum

It’s been a whirlwind year for those of us trying to make sense of the promise of agentic media buying. If you attended CES in January, you likely walked away with the impression that the tech was not ready for primetime. Now, agencies are running live agent-to-agent buys. Campaigns have been executed end to end from a natural-language brief with no human touching a line item. Holding company executives mention buying agents on earnings calls without hedging.

Adoption is arriving the way most real transformations do: with the architects celebrating technical milestones, and most end users still scratching their heads, completely unaware of how this technical advancement fits into an overall strategy. For now, they’re doing what they’ve done before and delegating the bottom-of-the-funnel work that automation has been claiming for years. It isn’t a coincidence that Google PMAX and Meta Advantage+ adoption numbers sit in the high 80s and 90s, while less than 10% of buyers trust the AI recommendations built into omnichannel campaigns. Handing over clicks and conversions has been easy, but marketers continue to rely on their own strategic expertise for the decisions that shape long-term business growth.

There are two reasons for that, and they are more connected than they look.

The first is that short-term performance buys confidence. Few buyers question the black-box algorithms promising better performance because they’re easy to measure: CPC is either in range or not; CPA either goes up or down. Whether you hand decisioning over to an agent, an algorithm, or an undefined “AI” doesn’t matter, because you’re chasing short-term goals and pivoting is easy if it fails. For many buyers, the scope of their roles ends here, so it’s an easy win. Which brings us to the second reason, and the one the industry keeps talking around. Upper-funnel prospecting is genuinely difficult for buyers and structurally hard for an agent, and not because either is immature. It’s because, as an industry, we still struggle to define clear, measurable awareness goals. Sure, we have simple test-and-control methods and proprietary MMM solutions, and we have piles of thought leadership debating whether MTA is useful or dead on arrival. You can’t build a protocol for something we’re still fighting about at conference hotel bars.

An agent improves by climbing a gradient. Upper-funnel feedback is sparse, delayed, and noisy, so there is very little gradient to climb. Worse, nearly every “new” audience input available today is still derived from conversion history when you look behind the curtain. Second-party files, lookalike expansions, retargeting pools and clean room overlaps all describe people who aren’t actually new. An agent working from those inputs runs out of road precisely at the edge of the known audience, by construction rather than by limitation. Hand it a brief saying “grow net-new buyers in this category” and it has nothing to decompose that task into, because the audience definition it would need does not exist in the data it can find. And even if it acted, there is no agreed measure of whether it acted well, which means most buyers will still hesitate to approve the delegation in the first place. Strategic growth is still their job, not the robot’s.

None of that is a flaw you fix with a better agent. A more capable model reading the same inputs is no better than a weaker one when tasked with prospecting. It just arrives at the same boundary faster. Agentic adoption of strategic pursuits sits below 10%, while PMAX takes more budget.

(It’s also worth saying plainly that for many advertisers this is not a problem at all. If the job this quarter is squeezing juice from demand that already exists—branded search, retargeting, retail media—then the agents work and the boundary is somewhere you were never going anyway.) Real growth comes from an audience input that isn’t downstream of prior conversions. It comes from interest forming outside the known base rather than behavio r already recorded inside it. The industry progress on protocol and standards heading toward portable audience signals is genuinely useful here, but it’s still focused on data transport rather than origin. A portable signal derived from the same conversion history is still a description of the past. It just runs out of sidewalk faster.

Agents crossed the easy half of the funnel quickly, and they crossed it because that half offered them both a gradient to climb and a simple number to report. The growth half of the funnel is not waiting on better models, and it is not waiting on the protocol argument to resolve. It is waiting on something truly new for the agents to read.