When the Ad Buyer Is a Bot: What Agentic Media Buying Means for Optimization
This Summer, Meta, Google, Amazon, and OpenAI all shipped agentic advertising tooling within days of each other. The pitch was naturally “let the AI manage your campaigns.” Set a goal, set constraints, and the agent handles allocation, bidding, and pacing across platforms without a human touching it between planning and reporting.
Most of the industry coverage has focused on whether the tools work. That's a reasonable question for month one. But there's a harder question underneath it that almost nobody is asking: what are these agents optimizing toward, and what happens when they get it wrong at machine speed?
If the signal is a problem, agents can make it worse.
Right now, a media buyer optimizing to ROAS on a retail platform is making a human-speed mistake. They're chasing a number that measures credit, not cause, and their budget is drifting toward campaigns that are good at claiming sales rather than creating them. That drift happens over weeks and quarters, and there's usually enough friction, a QBR, a budget review, a mediocre quarter, to catch it before it goes too far.
An agent removes the friction. It can reallocate budgets across campaigns, channels, and retailers thousands of times a day, each time feeding on the same distorted signal. If the input is ROAS, the agent will optimize ROAS with extraordinary efficiency, and it will do so by starving campaigns that drive real incremental sales in favor of campaigns that intercept demand that already exists. So the goal gets achieved but the business gets hurt. And because the agent moves faster than any human reporting cycle, the damage compounds before anyone notices.
We should note, this isn’t gaming the metric. Agentic buying won't need to game anything. It will just optimize, legitimately, toward whatever objective function it's given, at speed and scale..
What agents actually need to do this right
An agentic media buyer is only as good as its objective. Change the objective and you change everything about how it behaves.
Give it iROI, the causal signal that measures what media actually drove versus what would have sold anyway, and the same machine that was destroying value becomes the most powerful optimization engine in commerce media. It can identify diminishing returns per campaign in real time, shift budget toward incremental headroom while it still exists, and do it continuously rather than waiting for a human to pull the analysis.
This is exactly why the Wavemaker, Incremental, and Skai work for Church & Dwight won The Drum's Grand Prix. It was a closed loop: daily causal signals from Incremental fed via API into Skai's automated bidding rules, so the system optimizing campaigns was doing so against actual incremental return, not last-touch credit. The result was huge incremental sales improvement across the Church & Dwight portfolio on a fixed budget. FWIW, the automation wasn't the achievement; the signal was. The automation just multiplied it.
That architecture, measurement to activation to continuous learning, is what agentic buying needs to not be a disaster. And notably, it also happens to be the architecture that already exists for brands running causal measurement today. They're not waiting for agents to arrive. They're the ones who will benefit most when they do.
