What happens when AI agents run media buying
AI agents can read ad accounts and push buttons. A real two-arm test shows where they help buyers, where they break, and which guardrails hold.
Every media buyer has had the same thought at 2 a.m.: a machine could do this. Watch the numbers, kill the losers, scale the winners, never get tired. Now that AI agents can read ad accounts and press the buttons themselves, that thought has turned into a real engineering question.
The honest answer so far is that agents make excellent assistants and unreliable bosses. That isn't a hunch. It is roughly what the platforms are signalling, and what the more candid field reports say.
The test that settles the argument, for now
One of the clearest public accounts comes from a performance agency whose CEO described, at an industry conference this year, building about 15 agents for his buyers and keeping five. AffHouse wrote it up as a case study of an AI media-buying test run with two $5,000 arms, and is careful to say it could not independently confirm the agency or its figures.
The design was simple. Two arms, $5,000 each, the same pre-tested creatives and the same stop rule: pause when the cost per install or target action ran too high. In one arm the bot recommended and a human buyer decided. In the other, the bot launched, paused and managed campaigns on its own.
The autonomous arm spent about $1,000, judged every creative a loser and switched them all off. Its return on that spend came out just under 61%, a number the CEO himself called meaningless because the data was too thin. The agent didn't fail by overspending. It failed by deciding too early, on too little evidence, with complete confidence.
A second incident deserves a spot above every buyer's desk. A bot duplicated 70 campaigns, then tried to delete 100 and removed 30 that were working. The agency's response was a blanket rule: bots make no decisions. The one exception it kept was letting a bot pause unprofitable campaigns overnight, when waiting for a human costs real money.
What agents already do well
The five agents that survived share a pattern: they collect, reconcile and report. One recorded affiliate-program terms from chats, such as caps, KPIs and budgets, and checked them against the tracker. Another launched campaigns and uploaded creatives on command and produced daily reports. A third read the team's records and flagged buyers whose output was slipping.
None of that is glamorous, and all of it is where buyers lose hours. It is also low risk. A wrong report gets caught. A wrong deletion does not get undone.
The platforms are approaching the same line from the other side. Google's official MCP server for the Google Ads API ships read-only, and its documentation says it cannot modify bids, pause campaigns or create assets. Meta went further when it opened its ads AI connectors in beta on April 29, 2026, letting outside agents create and edit campaigns. Write access is a capability, though, not a recommendation.
Where autonomy breaks
Media buying looks like a numbers problem, which is why it tempts automation. In practice the hard calls sit in the gaps between the numbers.
- Small samples. Early results are noisy, and an agent with a crisp stop rule will act on noise as readily as on signal. That is exactly what happened to the autonomous arm.
- Context outside the account. A buyer knows an offer's cap is close, a creative is seasonal or a partner is late with postbacks. The agent only sees what it is fed.
- Irreversible actions. Pausing is cheap to undo. Deleting campaigns, resetting learning or burning through a creative batch is not.
- Platform rules. Policy status, account integrity and restricted categories are not things an agent should learn by trial and error.
There is also the platforms' own automation to account for. In mid-2025 The Wall Street Journal reported that Meta aims to let brands hand both ad creation and targeting to its AI by the end of 2026. When the platform is already making allocation decisions, an outside agent stacking its own decisions on top can end up fighting the algorithm instead of steering it.
Guardrails that hold up in practice
The useful question isn't whether the agent should be allowed to act. It is which actions, within what limits, and who signs off.
Split actions into tiers. Reading, reporting and drafting are free. Pausing is allowed only for pre-defined, urgent and reversible cases, like the overnight exception above. Anything that raises spend, deletes objects or changes targeting needs a named person to approve it.
Cap the blast radius in money, not only in permissions. Give the agent its own budget line with a written stop-loss, and run it against a buyer-managed arm on the same creatives. That turns "is the agent any good?" into a measured comparison rather than a feeling.
Limit volume as well. The duplicate-then-delete incident was a rate problem as much as a judgement problem. A hard ceiling on how many objects an agent may create or remove per hour would have stopped it after the first dozen.
Log every action with its reason: what the agent saw, which rule it applied and what it changed. When something goes wrong overnight, a human should be able to reconstruct it in minutes, not by scrolling through change history line by line.
Budgeting for the agent itself
Agents cost money to run, separately from the media they manage. The agency in the case reported peak AI spend of about $3,500 a month, falling since, and routed repetitive data collection to a cheaper model while keeping a stronger one for reasoning. It also found that automating a process worth about 30 minutes a week took longer than it ever saved.
That points to a simple screen before building anything. How much money does the task touch? How often does it run? What does a delay cost? How stable is the process? How bad, and how reversible, is a mistake? Frequent, stable and reversible tasks are the best candidates. Rare, high-stakes calls are where human buyers still earn their pay.
If you hire agents rather than build them, the same logic belongs in the contract: pay for defined outputs, hold funds until they are delivered and agree on stop conditions in advance. That is the thinking behind escrow-backed service levels on marketplaces such as Agent·Bazaar.
For a running record of how performance teams actually use these tools, AffHouse covers affiliate and media-buying practice closely. The short version today: let the agent be a pair of hands, and keep the wallet and the delete key with a person.
