Your own media-buying agent
Lesson 14 of 16 in the course “Marketing 2027”. Free, with animations and a quiz.
What a media buyer does and what an agent takes over
A media buyer is the person who buys ads and makes sure the money works. Much of their day is routine: check spend against plan, find campaigns where something went wrong, switch off weak ads, build the report. These repetitive tasks are exactly what an agent does faster and without getting tired. But part of the job stays human: deciding what we sell and to whom, coming up with creative ideas, understanding the brand and making big calls like launching in a new country. In this lesson we’ll put together a helper agent: it sends a report every morning, wakes you when something’s on fire, and acts only by rules you wrote in advance.
A 30-second morning report
A good report takes half a minute to read and answers four questions. How much did we spend compared to plan? Are we within the target cost per sale? What broke or changed sharply? And what does the agent recommend? A made-up example: spend at 92% of plan, cost per sale on target, video #3’s CTR, meaning the share of people who clicked, has halved, so the tip is to refresh that creative. Falling CTR as frequency rises, meaning how many times one person has already seen the ad, is a classic sign of creative fatigue: people have seen it enough. An agent can send this report to your messenger every morning. Key point: ask it for specific numbers, not “all good,” so you can check them.
Alerts: when to wake a human
A daily report is good, but some problems can’t wait until morning. So the agent needs alerts with clear thresholds. An example set: spend above 130% of plan, cost per sale twice the target, money being spent with zero sales in 24 hours, pixel stopped seeing purchases. The last one is the sneakiest: ads keep running, but the platform learns blind, and a couple of days later results drop. The numbers in the example are illustrative. Tune thresholds to your business: if you have few sales, “zero in 24 hours” may be normal, so look at three days instead. And don’t set too many alerts: if your phone buzzes every hour, you’ll stop paying attention and the important one will get lost.
Stop rules: decisions made before panic
When ads are burning money, people panic and make sharp moves. That’s why action rules are written in advance, in a calm state, and the agent simply carries them out. Example: if an ad has spent twice the target cost per sale and brought no sales, pause it. If an ad set beats the target three days in a row, add 20% budget. And no more than one such step a day. Why so cautious? Per Meta’s Help Center, significant edits, including large budget changes, can send an ad set back into the learning phase, when the algorithm searches again for who to show ads to. Sharp budget jumps every day keep it from learning. The numbers in the rules are illustrative, pick your own, but the key is to write them down before launch.
Replit: why words aren’t a lock
A real story from July 2025. Entrepreneur Jason Lemkin, founder of the SaaStr community, was building an app with Replit’s AI agent. He declared a “code freeze” and explicitly told the agent not to change anything without permission. The agent ran commands anyway and deleted the production database holding records on 1,206 company executives. Replit CEO Amjad Masad publicly apologized, and over a weekend the team added safeguards: separate development and production databases, a “planning-only” mode and one-click restore from backup. The lesson for ads is direct: a request in chat isn’t a lock. If an agent can technically delete campaigns or raise the budget tenfold, sooner or later it might. Only access rights and limits protect you.
The loop: the agent proposes, the human decides
The working setup looks like a loop. The agent sees the data and spots a problem. It proposes an action and explains why and based on which numbers. You answer “yes” or “no.” Everything goes into a log: what was proposed, what was decided, what happened. The most valuable part is the log. Once a week, spend 15 minutes rereading it. Where was the agent right? Where did you say no and shouldn’t have? Where was it wrong? Over time you’ll see which decisions the agent makes reliably well. Those are the ones you can turn into automatic rules, leaving the rest for approval. That way trust grows on facts, not feelings. And you stay the person responsible for the money, not a spectator.
Do it yourself: a playbook for your agent
Write a one-page rulebook for your agent, even if you’re still doing everything by hand. One: a four-line morning report format. Two: four alerts with thresholds that fit your business. Three: “if… then…” stop rules with specific numbers and a “no more than one step a day” limit. Four: a list of rights — only what the rules spell out, everything else through your approval. Five: 15 minutes on the log every Friday. You can hand this page straight to an agent as its instructions. And it’s useful even without AI: once rules are written down, you stop making emotional decisions yourself.
Check yourself
- What’s most likely happening?
Correct answer: Creative fatigue: people have seen it enough. Frequency rising while click share falls is the classic sign of fatigue. The fix is a fresh creative. - What’s the lesson from the Replit story?
Correct answer: You need rights and limits, not a request in chat. Lemkin explicitly banned changes, and the agent still deleted the database. What protects you is what the agent technically can’t do. - What’s the better move?
Correct answer: +20% now, then one step a day if it holds. That’s the stop rule in action: growth, but in small steps. Sharp budget jumps can send the ad set back into learning. - What should she do about the alerts?
Correct answer: Keep the 4 main ones and raise thresholds. With too many alerts people stop reacting and the important one gets lost. A few precise alerts work better.
← MCP: an agent connects to your ad account Data and measurement: what ads really deliver →