The honest answer to “will media buying be replaced by AI” is that most of the job as it was defined five years ago is already gone. Audience building, bid management, budget allocation within campaigns: the platforms absorbed that layer, and the firm has said so for two years. So the interesting argument is not whether AI runs execution now. It is what the people announcing the replacement leave out of the frame.
Two pieces worth taking seriously
The replacement case is being made well right now, by people with data, and it deserves engagement rather than dismissal. Luke Jonas at Nest Commerce writes that “AI has already replaced large parts of what media buyers used to do,” naming audience building, bid management, and budget allocation within campaigns 1. His sharper claim follows: “Performance is no longer dictated by how clever you are at platform optimisation. It’s dictated by creative supply” 1. And Nest’s client data supports the creative half of that: brands running roughly 3x creative volume through Q4 2025 saw ROAS up 18% and revenue up 38% year over year 1.
Ben Pflugpeil at Superscale goes further. His June piece argues an agentic media buying platform “can replace most of the day-to-day execution your media buying team does on Meta and TikTok,” beating the human team on throughput (200+ creative variants a month against 8 to 30), on speed, on coverage, and on cost, roughly $500 to $5K a month against a $66K to $97K base salary 2. His summary line is the honest version of the pitch: “It replaces the team’s hands, not its head” 2. He also discloses, to his credit: “We build one of these agents, so treat the rest of this with appropriate suspicion” 2.
Here is where the firm sits, and it will read strangely coming from an agency: we agree with more of this than most of our peers will say out loud. The disagreement is narrower than “AI can’t do this,” and it matters more.
The definition is doing the work
Both pieces define media buying as the execution layer, then announce the execution layer is automated. True, and circular. If the job is audience building, bid management, and budget allocation, the job has been dissolving since Advantage+ shipped, and nobody at this firm will argue otherwise. But the job was never only the hands.
What remains is not “creative supply” as a volume metric either. It is judgment: which offer to run, whether the unit economics survive the CAC the auction will hand back, how to design measurement so you know when the dashboard is lying, and when to overrule a reported number with a holdout test. Judgment looks like this in practice: the platform says the retargeting is carrying the account, the holdout says most of that revenue was coming anyway, and someone has to decide which of those two numbers gets to allocate next month’s budget. No throughput metric answers that question.
Nest’s volume data is real, and it matches what the firm found when it documented creative fatigue hitting in five days and rebuilt its testing cadence around that window. Creative volume moves the number; on that point there is no dispute between us. But volume is an amplifier, not a strategy. Feed an agent the wrong offer and it will ship its 200 monthly variants 2 of the same bad guess, faster and more cheerfully than any human team could.
The math prices the hands and skips the outcome
The cost comparison in the vendor pitch is a subscription against a salary, and it quietly assumes the outcomes are equal. The $500 to $5K a month against $66K to $97K in base pay 2 reads as an obvious trade until you ask what the salary was buying. Part of it bought execution, and that part is gone; no argument here. The other part bought a person who owns the result.
When an AI media buyer optimizes into a corner, and optimizers optimize into corners because they maximize exactly the number they are given, someone still has to notice, diagnose, and answer for the quarter. Accountability is the line item the subscription does not include. Nobody in the tool-versus-salary comparison owns the outcome, which is a strange omission for a comparison whose whole point is outcomes. Ask any operator who has inherited an account that an automation ran unattended for two quarters: the spend was efficient the whole way down, against a conversion event nobody had audited, and unwinding it cost more than the salary the tool saved.
None of this defends the old agency model. The firm has argued that traditional PPC agencies bill hours against the layer that has been automated, and the current crop of AI media buying tools makes that billing model harder to defend every month. If an agency’s invoice is mostly button-pushing, these tools are coming for the invoice, and they should.
Cheap execution raises the price of judgment
AI compressed the cost of operating an account, and that raises the return on the judgment layer rather than eliminating it. When Google shipped Ask Advisor, the firm wrote that it absorbs 40 to 60 percent of an account manager’s operational week, and that the surviving hours move to offer design, creative concept, landing page work, and measurement. Agentic media buying brings the same shift to Meta and TikTok. Same pattern, same conclusion.
The economics cut both ways, and this is the part the replacement essays skip. Cheap execution makes good strategy cheaper to express: an operator with the right offer and a real read on incrementality can now run it at a volume that used to require a team. Cheap execution makes bad strategy cheaper to scale, too: the wrong offer also ships at ten times the volume, with confident dashboards the whole way down. The tools do not close the gap between operators. They widen it.
Where the line moves from here
The empirical dispute is live, and pretending otherwise would be its own kind of vendor pitch. If agentic tools keep improving at their current rate, the line between hands and head will keep moving, and some of what this piece calls judgment will get absorbed the way bid management was. The firm has updated its positions before when the data moved, and it will update this one the same way: in public, naming what changed.
For now the answer holds. The execution layer is gone and it is not coming back, and operators who mourn it are mourning the wrong thing. What got scarce is not hands. It is knowing what to build with them.