AI agents are not replacing media buyers; they are replacing the parts of media buying that were always mechanical, while raising the value of the judgement that was never mechanical at all. The evidence from how agentic systems are actually being deployed points the same way: agents run the continuous, repetitive optimisation, and humans set the strategy, draw the boundaries, and approve anything that matters. The role changes shape. It moves from operating the machine to directing and governing it. That is a real change, and it is not disappearance.
TL;DR. The honest answer is that agents replace tasks, not the role. The tasks most exposed are the continuous, rules-adjacent ones: bid adjustment, pacing, routine reallocation, the work a buyer did in daily optimisation passes. The work that becomes more valuable is strategy, boundary-setting, judgement about trade-offs, and accountability for outcomes, none of which an agent can own. Every serious deployment model keeps a human setting goals and approving material decisions, because an agent is a fast, literal worker that will optimise toward a flawed signal without the sense to stop. The useful mental model is an agent as a tireless junior buyer whose autonomy widens only as its record earns trust, not as a replacement for senior judgement.
What does the evidence actually show?
It shows augmentation with a shifting boundary, not replacement. The systems being deployed do not remove the human; they change what the human does. Across the platforms and operators building agentic buying, the consistent pattern is that agents handle the moment-to-moment optimisation and humans handle the strategy and the sign-off. The industry's own framing of the mature model is telling: agents treated as tireless, fast, auditable junior media buyers and analysts, explicitly not as a replacement for senior judgement.
The adoption data supports the same reading. Industry research shows a large share of digital video buyers are live with, testing, or planning agentic campaigns, and holding companies have executed live agent-to-agent buys for clients. What that live volume actually looks like is smaller than the announcements imply: in one on-record account, the most forthcoming operator in the category put daily agent-bought, agent-sold spend at a few thousand dollars, not the trillion-dollar shift the keynotes describe. But those deployments are structured as delegation within boundaries, not as removal of the buyer. The agents buy; the humans decide what the agents may buy, and remain accountable for the result. This is what adoption of agentic buying looks like in practice: not empty desks, but buyers doing a different job.
It is also worth being clear about what the evidence does not show. There is no credible account of agentic systems running major campaigns end to end with no human strategy or approval, and the platforms that have tried to push autonomy fastest are precisely the ones the governance literature warns about. The direction of travel is more autonomy over time, but from a base of human control, widened deliberately, not a wholesale handover.
Which parts of media buying are agents genuinely taking over?
The continuous, repetitive, rules-adjacent parts, the work that was always closer to operation than to judgement.
Bid management is the clearest. Adjusting bids in real time against an outcome target was already semi-automated under programmatic, and agents complete the automation, doing it continuously and in connection with the rest of the campaign rather than as an isolated rule. Pacing is similar: keeping delivery on track against a flight is exactly the kind of constant, mechanical monitoring an agent does better than a human checking twice a day.
Routine optimisation is the broader category. The daily passes a buyer made, shifting budget toward what is working, narrowing an underperforming audience, adjusting a placement, are continuous, data-driven adjustments an agent can make as the data arrives rather than when someone next opens the platform. This is genuinely the buyer's old day-to-day, and agents are genuinely absorbing it.
Reporting and monitoring shift too. An agent that logs its own decisions and their outcomes produces a running account of the campaign that a human used to assemble by hand. The assembling was labour; the interpreting is judgement. Agents take the assembling.
What these have in common is that they were never the hard, distinctively human part of the job. They were the part that filled the hours. Removing them is why the role changes rather than why it ends.
Which parts can an agent not take over?
The parts that require judgement about what should happen rather than optimisation toward a set target.
Strategy is the first and largest. Deciding what a campaign is trying to achieve, how it fits a brand's wider goals, what trade-offs are acceptable, these are choices about ends, and an agent optimises toward ends it is given rather than choosing them. An agent handed the wrong objective will pursue it flawlessly. Someone has to be right about the objective, and that someone is human.
Boundary-setting is the second, and it is becoming the core of the buyer's craft. What may the agent do on its own? Where must it stop and ask? What is it forbidden from doing regardless of what its optimisation suggests? These decisions determine whether an agent is safe to deploy, and they require exactly the judgement about consequences that an agent lacks. The buyer's expertise moves here, from making each call to deciding which calls the agent may make.
Accountability is the third, and it cannot be delegated at all. When an autonomous decision goes wrong, someone has to answer for it to leadership, to a client, to a board, and that someone is not the agent. An agent has no career, no relationship, no stake. The human carries the outcome, which means the human has an irreducible reason to stay in control of the decisions that create it. You cannot outsource the blame, so you cannot fully outsource the decision.
Judgement about the non-obvious is the fourth. An agent optimises against the signals it can read. A human notices the thing the signals miss: that a placement performing well on the metric is damaging the brand, that a supply path looks efficient but feels wrong, that the data itself might be untrustworthy. This is the skepticism an agent structurally lacks, because it acts on its inputs and cannot doubt them the way a person can.
Why does an agent specifically need a human, not just prefer one?
Because of a failure mode that is a property of autonomous systems, not a temporary flaw to be engineered away. An agent does exactly what it is told against the data it is given, at speed. When the signal is flawed, the agent does not hesitate; it makes the same wrong decision repeatedly before anyone notices, because hesitation is a human trait and the agent has none. Analysts call this silent failure at scale, and the human approval gate exists precisely to be the interrupt the agent cannot provide for itself.
The governance literature is blunt about this. Enterprises adopting agentic systems widely still have only a minority running safely in meaningful production, with the rest closer to autonomous-looking chatbots than to trustworthy production systems. The recommended operating models all keep a human as a named authority over each agentic process, with intervention triggers and the ability to roll back, and they run bounded pilots under oversight before anything scales. This is not caution for its own sake. It is the recognition that an agent's speed, its greatest asset, is also what makes an unsupervised error catastrophic, and that a human is the only thing that reliably stops it.
So the human is not in the loop as a courtesy or a transitional measure. The human is in the loop because the specific way agents fail requires an outside interrupt, and because the specific decisions that carry accountability cannot be held by something that bears no consequences.
What does the media buyer's job become?
It becomes a job of direction and governance rather than operation. The buyer of the near future spends less time making individual optimisation calls and more time on the things that determine whether the agent's calls are any good: setting the right strategy, drawing the right boundaries, reviewing the agent's decisions and the reasoning behind them, and deciding when the agent has earned more autonomy and when it has not.
This is a more senior job, not a lesser one. Operating a platform was skilled labour, but it was labour. Directing an autonomous system, deciding what it may do, and owning its outcomes is closer to management than to trading. The buyers who thrive will be the ones who move up that curve, from executing the buy to governing the thing that executes it, and who develop the new craft of configuring and supervising agents well. The buyers most at risk are not those replaced by agents but those who define their value entirely by the tasks agents are absorbing, and do not move up to the judgement that agents cannot.
The evidence-based answer, then, is not the reassuring "agents will only ever assist" nor the alarming "agents will replace you." It is that the mechanical core of the job is moving to agents, the judgement core is becoming more valuable, and the role is relocating from the keyboard to the decisions about what the agent at the keyboard is allowed to do.
This article references industry analysis of agentic media-buying deployment and governance, published adoption research, on-record reporting on live agent-to-agent buying, and enterprise studies of agentic AI in production. Where it describes the human's role in autonomy and approval, it reflects the graded-autonomy and oversight models documented across those sources.