What changed · Ad tech
Agentic media buying: what to hand over, and what to keep
The question is no longer whether agents will run media. It is which decisions you are willing to not be in the room for.
20 August 2026 · 6 minute read
Agentic media buying is arriving in three places: platform-native agents that act with human approval, supply-side agents that troubleshoot and optimise yield, and an emerging protocol layer for agents to negotiate with each other. Adoption is cautious, and the reason advertisers give is consistent: accuracy and transparency. The useful way to approach it is not to ask whether to adopt, but to sort decisions by how reversible they are. Delegate the reversible, high-frequency, low-consequence ones first, keep the ones that are hard to undo, and insist on being able to see what was done and why.
What is actually shipping?
Three distinct things, often discussed as one.
- Native platform agents inside the ad platforms, executing changes with a human approval step. This is the most mature and the least alarming.
- Supply-side agents that troubleshoot delivery and optimise yield. Mostly invisible to advertisers, and already influencing what you pay.
- A protocol layer for buyer and seller agents to negotiate directly. Early, and the one with the largest implications if it matures.
Why is adoption slower than the announcements suggest?
Because a majority of advertising professionals name accuracy and transparency as the barrier, and both objections are reasonable. An agent that makes a good decision you cannot inspect is indistinguishable from one that made a lucky decision, and you will not know which until it stops working.
There is also an accountability problem that nobody has solved. When an agent spends badly, the responsibility still sits with a person who was not consulted, which is a poor arrangement to volunteer for.
How should you decide what to delegate?
Sort by reversibility, not by importance. The question is not how much the decision matters; it is how cheaply you can undo it.
| Decision | Reversible? | Delegate? |
|---|---|---|
| Budget shifts between existing ad sets | Yes, immediately | Early candidate |
| Bid adjustments | Yes | Early candidate |
| Pausing an underperforming ad | Yes | Early candidate |
| Audience expansion | Mostly, after spend | With a cap |
| New creative going live | No, it was seen | Keep approval |
| Total budget increases | No, money is spent | Keep |
| Anything touching brand positioning | No | Keep |
The rule worth adopting
Delegate decisions you could undo before lunch. Keep the ones a customer would see, or that spend money you cannot get back. This sorts most cases without needing a policy document.
What should you insist on before delegating anything?
- A log you can read. What changed, when, and on what basis. If the agent cannot explain itself, you cannot review it.
- Hard limits. A spend ceiling the agent cannot cross regardless of what it concludes.
- A confirmation on anything expensive. The right pattern is the agent proposing and a person confirming, not the agent acting and a person auditing.
- Reversibility by default. If undoing requires engineering help, it is not reversible.
The uncomfortable part
Supply-side agents are already shaping what you pay, and you did not opt in. The negotiation is increasingly machine to machine on one side, whether or not your side is automated, which is a genuine argument for adopting rather than waiting.
That argument is real. It is also the argument that gets used to justify delegating things that should not be delegated, so hold both thoughts at once.
The pattern that works is the agent proposing and a person confirming. The pattern that fails is the agent acting and a person auditing afterwards, because nobody audits afterwards.
If you are an agency
Clients are going to ask about this, and the answer that builds trust is a delegation policy rather than a position for or against. Bring a reversibility table to the conversation, agree what is delegated, and put spend ceilings in writing. Being the agency that made the boundaries explicit is a considerably better position than being the one that either resisted or over-promised.
How we do this ourselves
Our own automation follows the same rule. The autonomous agent in the AEO/GEO Monitor has exactly one launch path and it sits behind a confirmation, and anything that spends money tells you when it last ran and asks before spending again. We would rather be asked twice than explain an unexpected bill.
How Ads Manager works