How To Replace A Media Buyer With An AI Agent, Step By Step
You replace a media buyer with an AI agent in seven steps: secure full admin access to your own account, write the operating rules with real numbers, name the person on your team who reads the weekly report, stand up agents for the six daily jobs inside a hard spend cap, run them in parallel with the buyer, switch one job at a time as the agent's output passes the same review, and end the retainer only when the conditions are met. The order is the point. Owners who skip steps one through three and jump to "turn on the agent" are the ones who end up with a gap in results and a story about how AI did not work.
Step 1: Get your account back
Before anything else, confirm you have full admin access to the ad account, the pixel or dataset, the audiences and the page. If campaigns currently run in the agency's account, this is a migration conversation with the buyer, and it happens first. You cannot replace someone on an account you do not control. If the buyer resists, you have learned something about the arrangement, and it belongs in the decision itself.
Step 2: Write the rules, with numbers
This is the step that decides whether the rest works. Sit down with whoever knows the account best, which may be the buyer in their last weeks, and write one page per job.
| Job | What the page must contain |
|---|---|
| Creative | The offer, the angles, the audience, brand constraints, format list, how many variants per brief, who approves |
| Audiences | Sources: your list, your pixel, your buyers. Refresh cadence. Exclusion rules. Lookalike seed rules |
| Launching | Campaign template, naming convention, starting budget, who checks before live |
| Reading numbers | Which metrics, pulled when, compared against what thresholds |
| Trimming and scaling | Pause rule with a number. Scale rule with a number. Minimum spend before any rule fires. Daily cap |
| Reporting | Format, cadence, what changed, why, what was learned, what is next, who receives it |
If you cannot write a number for a threshold, you are not ready to automate that rule, and the honest truth is your buyer was probably running it on feel. Pick a conservative number, write it down, and plan to revise it after the first cycle.
A rule without a number is an opinion. Agents do not run on opinions.
Step 3: Name the human who reads the report
Not the owner. Someone on your team who will read the weekly report every week, act on it, and raise the things the rules did not anticipate. This is the single most common failure point: owners set up the agents, nobody is assigned to read the output, and three weeks later the account has drifted. The whole model on my account is that agents do the six jobs and a person reads a weekly report. Without the person, you have automation with no accountability. Pick them now, before the agents run.
Step 4: Set the cap at the platform level
Write the daily spend cap into the agent's instructions, and also set it in the ads platform as a hard limit. Two layers. The instructions can be wrong; the platform cap cannot be argued with. The right first cap is a number you would be annoyed but not harmed to lose in a day if every rule fired badly. Raise it later, after the agent has run clean inside it.
Step 5: Run in parallel
Keep paying the buyer. Stand up the agents on the six jobs alongside them. For the jobs that act on the account, start with the agents in a recommend-only mode: they produce the creative, the audience builds, the trim list and the report, and your named human compares them with what the buyer did that week.
You are looking for two things. Where the agent's output matches or beats the buyer's, that job is ready to switch. Where it falls short, find the gap, write it into the rules page, and run another week. If you cannot write down what the buyer did better, you have found the judgment work that is worth keeping a human for. That is useful information, not a failure.
Step 6: Switch one job at a time
Move jobs from the buyer to the agents in this order, and only move the next one when the previous one has held.
- Reporting. Lowest risk, highest learning. The agent writes it, your human reads it, the buyer still runs the account. You learn whether your team can act on an agent-written report.
- Reading the numbers. The agent does the daily read and flags thresholds. No actions yet. You confirm its reads match reality.
- Creative production. Agent produces variants to the brief, your human approves, buyer launches. Your creative volume goes up and the buyer stops being the bottleneck.
- Audiences. Agent builds and refreshes from your sources on schedule. Buyers become the exclusion and the seed without anyone remembering to do it.
- Launching. Agent launches from the template, your human checks before live.
- Trimming and scaling. Last, because it is live money. Agent applies the written thresholds inside the cap. By now you trust its reads and your rules have been through several revisions.
When all six are running on agents and the weekly numbers have held through a normal cycle on your account, including at least one creative refresh and one report your team acted on, the buyer is no longer running the account. They are advising, if they are doing anything.
Step 7: End the retainer, or change it
Now the decision is clean. If the buyer brought real strategy and taste during the parallel run, offer them an advisory arrangement at a fee that reflects strategy and taste rather than the daily loop. If what they brought was the loop, thank them and end it. Either way, do it after the conditions are met, not on a date you picked in advance. What each version actually does is a good reference for the renegotiation, because it separates the parts of the old fee that still have a job from the parts that do not.
What this gets you
Your account, with full access. Your rules, on a page your team owns. Your person, reading a weekly report. Agents doing the six jobs every day inside a cap. And a retainer line that disappears or shrinks to the part worth paying for. The expensive thing you have removed is not the fee. It is the dependency, and what a bad media buyer actually costs you is mostly about that.
This is the capability I help owners build. Not as a new agency, but as their own team running their own account. If you have read this far and want to do it with someone who runs the loop this way already, the application on the home page is the next step.
FAQ
How long should the overlap with my current media buyer last?
Long enough that the agent-run jobs have produced results you trust across a normal cycle on your account, including at least one creative refresh and one weekly report your team acted on. Do not set a date in advance; set the conditions and end the retainer when they are met.
What should the daily spend cap be?
Whatever you would be comfortable losing in a single day if every rule fired wrong. Start conservative, set it at the platform level as well as in the agent's instructions, and raise it only after the agent has run cleanly inside it for a while.
Who on my team should own the weekly report?
Someone who is not the owner and who will actually read it every week. A marketing lead, an operator, a chief of staff. The owner should be able to read it but should not be the single point of failure for whether it gets read.
What if the agent's results are worse than the buyer's during the overlap?
Fix the rules, not the agent. Compare what the buyer did that week with what the rules said, find the gap, and write it into the page. Then run another cycle. If you cannot write down what the buyer did better, that is the judgment work worth keeping them for.