Last week, I wrote about where AI should and shouldn’t be making decisions in media. Agentic buying takes that conversation a step further.
Once we decide a machine can make a buying decision, we need to ask a different question: how do we know it’s making that decision in the advertiser’s best interest?
I’m not particularly worried about whether the technology will be capable of buying media. It will. I’m more interested in what happens when we give an agent an objective and the authority to pursue it. Who defines what it optimizes for? What information does it act on? What happens when improving one metric comes at the expense of the wider business? And crucially, can the advertiser see why the agent made the decisions it did?
These aren’t entirely new problems. Programmatic made media buying more efficient, but also created an ecosystem where supply paths, fees and decision-making became harder to see. More recently, products like Performance Max and Advantage+ have asked advertisers to hand over more control without necessarily getting more transparency in return. Agentic buying has the potential to push that trade-off much further.
Getting the objective right is harder than it sounds
Made-for-advertising sites are a useful lesson in what happens when automated systems get very good at achieving an objective that doesn’t fully represent what the advertiser wants.
For years, programmatic buying rewarded efficiency: reach more people, deliver more impressions, lower the cost. Automated systems found increasingly effective ways to do exactly that, including directing spend toward inventory designed primarily to generate large volumes of cheap ad impressions. From the system’s perspective, it was doing its job. However from most advertiser’s perspective, half their budget was being wasted. The quality and value of what they were buying didn’t line up with their expectations.
The challenge is that fixing this isn’t as simple as adding another KPI. Media buyers are constantly balancing outcomes that compete with each other: cost against quality, efficiency against incrementality, et al. We know the most efficient impression isn’t always the most valuable one, and improving one number can come at the expense of something that matters more. An agent needs these trade-offs reflected in the objective and constraints it is given.
That’s a specification failure. The system hasn’t necessarily failed to optimize; the objective has failed to capture the full outcome we actually care about. If we tell a machine to maximize impressions at the lowest possible cost, we shouldn’t be surprised when it finds the cheapest possible impressions.
Agentic buying raises the stakes because the system has more freedom to decide how that objective should be achieved. As agents become better at planning, buying and adapting without intervention, a poorly specified objective becomes more consequential. More intelligence doesn’t correct an incomplete definition of success. It can make the system much more efficient at pursuing it.
So before handing over meaningful budget, advertisers need to be clear about more than what an agent should optimize toward. They also need to define the boundaries around that optimization: what it shouldn’t sacrifice in pursuit of the target, and when a decision has moved beyond the remit it was given.
Even the right objective can’t capture everything
But better specification only gets us so far. Even if we could perfectly describe the outcome we want, a buying system can only optimize against the information available to it. Some of the things that matter most in determining whether media actually worked are difficult, or impossible, for the agent to observe directly.
The counterfactual is a good example. An agent can see that someone was exposed to an ad and later converted. What it can’t observe is whether that person would have converted anyway. If the agent is rewarded for attributed conversions, it has no inherent reason to distinguish between creating an incremental conversion and finding someone who was already highly likely to buy. Both look like success in the feedback it receives.
That means an agent can get better at its objective without necessarily creating more value for the advertiser. More training data doesn’t make the counterfactual appear because, by definition, the alternative outcome is unobserved. You need a measurement framework capable of separating correlation from causation, and you need to make sure the signals feeding the agent reward the outcome you actually care about.
The same limitation becomes more obvious when you move beyond a single channel. A retail media agent could improve its own return while cannibalizing sales attributed to affiliate. A performance agent could maximize short-term efficiency while making choices that work against longer-term brand growth. Neither agent necessarily sees a problem because the consequence sits outside the objective it has been asked to optimize.
This is why I’m wary of treating channel-level optimization as synonymous with business optimization. Media channels interact, and the most efficient decision within one system isn’t automatically the best decision for the wider plan. An agent can win its metric while losing the account.
That creates two different problems for agentic buying. There are outcomes an agent cannot directly observe, and there are consequences it may never consider because they fall outside its objective or remit. The more buying decisions we delegate, the more important it becomes to understand both. Otherwise, we risk building incredibly sophisticated systems that optimize individual parts of the media plan while nobody is responsible for whether those parts add up to the right business outcome.
More autonomy should come with more transparency
None of this is an argument for keeping AI out of media buying. There are already areas where AI can operate at a scale human teams simply can’t. Supply path analysis across log-level data is an obvious example, alongside anomaly detection, pacing, QA and increasingly the first stages of planning and campaign builds. At Brainlabs, we have a saying that anything repeated three times should be automated, and there is still a lot of media execution that meets that bar.
The distinction is how much authority sits behind that automation. As an agent moves from helping execute a defined task to deciding how advertiser money should be spent, the standard for transparency and accountability has to rise with it. We’ve already seen what happens when greater platform automation comes with less visibility into how decisions are being made. Agentic buying shouldn’t simply extend that bargain.
For me, there are some fairly basic requirements before an agent should have meaningful freedom over media budget. Advertisers need to know what it will and won’t optimize toward, which inventory it can access and what data is informing its decisions. Fees should be disclosed, log-level data should be available, and important decisions should be independently auditable rather than assessed solely by the system that made them.
There also needs to be clear human ownership of the data layer. Agents will make decisions based on the information available to them, which means someone has to be accountable for where that information came from, when it was last verified and whether it still reflects the reality of the business. A sophisticated agent acting on outdated or incorrect inputs can still make a very bad decision, just with a lot more confidence and speed.
That doesn’t mean putting a human back into every decision. If every pacing adjustment, inventory choice or optimization requires approval, we’ve removed much of the value agentic systems could create. The point is to make sure greater autonomy is matched by the ability to understand and audit what the system is doing.
I have very little doubt that agentic buying will work technically. The more important question is who it will work for. If the next generation of media buying gives machines more control while advertisers once again get less visibility into where their money goes, how decisions are made and whose interests those decisions ultimately serve, then we haven’t fixed programmatic’s trust problem. We’ve automated it.
