Do you still need a media buyer with AI?

You open the account on Monday. Advantage+ picked the audience. The algorithm chose the placements, set the bids, and rotated the creative overnight. The dashboard is green. And somewhere in the back of your head a quieter question is forming: if the machine did all of that, what exactly am I being paid for?

Short answer: You still need a media buyer because AI optimizes toward the goal you hand it, but it cannot choose that goal, judge whether a cheap result is a real one, or decide when a losing test deserves more time. Those three calls set profit, and they stay human.

The takeaways

• AI has taken the execution, not the judgment. Bidding, targeting, and budget allocation inside a campaign are largely automated now; the decisions that sit above them are not.
• The most expensive mistake is optimizing to the wrong event. Tell the machine to find cheap leads and it will, whether or not those leads ever buy.
• A good result and a real result are different things. A cheap CPA on day two is usually luck reverting to the mean, and only a human decides how long to wait before believing it.

What has AI actually taken over?

The tactical layer, and most of it. Inside a campaign, machine bidding sets the price of each auction, the targeting expands past the seed audience on its own, and budget shifts toward whatever is converting this hour. Meta reports the majority of its advertisers now lean on its Advantage+ products, and Google's Performance Max works the same way: you feed it assets and a goal, it handles the rest.

This is good, and overdue. Manual bid tweaks and hand-built lookalikes were never where the edge lived, and pretending otherwise wasted a lot of afternoons. But notice what all of that automation shares: it optimizes within the boundaries you set. It makes the how faster. It never touches the what or the whether.

What can't AI decide for you?

The goal. That is the whole game, and it is the one input the algorithm cannot supply for itself. You tell Advantage+ to optimize for form fills, and it will get you form fills at a beautiful cost, sourced from the cheapest, least-qualified corner of the internet it can find. The machine did its job perfectly. You just pointed it at the wrong finish line.

Choosing the right event to optimize toward means holding context the algorithm has no access to: your margins, which lead actually turns into revenue, what a customer is worth over a year, whether this month is about growth or cash. I have watched a campaign hit its cheapest-ever cost per lead in the same week the business lost money on it. Picking a goal that maps to profit is a human judgment, made before the automation ever runs.

How do you know if a cheap result is a real one?

You wait, and you know what you are waiting for. An ad that posts a stunning cost per acquisition on day two is usually not a winner; it is a lucky sample that will drift back toward its format's normal as spend accumulates. Regression to the mean is not a metaphor here, it is the default outcome, and acting on the early number is how buyers scale a fluke into a loss.

This is judgment work that resists automation, because the machine that spends the money is the same one reporting the result, and it has no incentive to tell you the sample is thin. A buyer who understands variance gives a promising ad enough budget to prove itself and refuses to crown it before the evidence is there. That patience, sitting between a green dashboard and a scaling decision, is most of the job now.

When should you let a losing test keep running?

Longer than your nerves want, and here is the honest part. I built an automation layer for exactly this, and I still made it default to recommendations rather than full-auto, because a kill decision made on thin data is how you murder an ad that was two conversions away from working. Rules that fire on eight metrics are useful, but rules with no judgment behind them mis-kill.

So the guardrails are deliberately conservative: a learning-phase lockout that leaves an ad alone under five days or €200 in spend, a ROAS floor that pauses a marginal ad instead of killing it, a check that never strips a campaign below three active ads, and a 24-hour window to undo any kill. Every one of those exists to protect a human's call from the machine's impatience. That is the shape of the new job: you own the judgment, the automation owns the execution, and the safeguards keep the two from colliding. If you want to see how that split works in practice, the automation approach is built around recommendations a person approves before anything acts.

So what does the job become?

Smaller in busywork, heavier in the parts that were always hard. You stop building audiences and start deciding what to optimize for. You stop rotating creative by hand and start reading what the algorithm did with the creative it had, then briefing the next round. And the reports you used to pull? They matter now for one question the machine cannot answer on its own: what do these numbers mean for the business behind them?

None of that is going away, because none of it is busywork. It is a set of decisions, and decisions were the job the whole time. The machine just cleared the desk so you could finally get to them.