Letting AI run your ads: the brief nobody wrote

Short answer: an AI left to run an ad account does what it was told and fills every gap with a default, usually the platform’s. In an agency pilot I run, one assistant for 13 client accounts and more than 11 media buyers, every failure I found in August 2026 was a missing brief or a missing rule. Not the model.

Why do people hand their ad accounts to AI?

Because they were told to.

By a client who read about it. By a boss who saw a demo. By LinkedIn, where every week someone says manual ads management is dead. I’ve written some of those headlines myself, so I’m not pointing fingers.

And by the platforms. Google and Meta push the same direction every quarter: Performance Max, Advantage+, recommendations you can auto-apply. That’s where they make more money. I don’t think that makes them villains. It makes them a party with an interest, and it’s worth remembering whose default you’re accepting.

The sentence I hear most when I talk to agencies is some version of this: we want AI in the team, we can’t lose quality, and we don’t trust it.

Nobody hands them the missing piece, which is how. So one group gives the AI the keys and hopes. The other group checks everything it says until the AI is extra work. This piece is about the first group. The second group, the one that checks every number, has its own.

What actually goes wrong when AI runs the account?

An AI running an ad account doesn’t break loudly. That’s the problem. It does exactly what it was told, and nobody told it the things that matter:

• what a good lead is for this client, not just a lead
• which conversion counts, and which one only looks like it counts
• what was tried last spring and why it was switched off
• the target, and what happens to the rest of the budget when one campaign gets more

Every gap gets filled with something. Usually the platform’s default, or the most common answer on the internet. The account keeps running. The reports keep coming. Nothing looks broken.

Then there’s the part people don’t expect. The model tends to agree with you. Ask it should I scale this campaign? and it will usually find a reason to say yes. That isn’t a bug in one tool. Anthropic’s researchers measured it across the leading assistants:

“Sycophancy is a general behavior of state-of-the-art AI assistants, likely driven in part by human preference judgments favoring sycophantic responses.”

Sharma et al., Towards Understanding Sycophancy in Language Models, 2023.

Put autonomy on top of that and you get your own bias, executed faster, with nobody in the room to say wait.

What did a real agency pilot teach me?

At the agency where I work, I run a pilot: one AI assistant for 13 client accounts, with more than 11 media buyers working across them. A shared drive stands in for the company wiki. When I mapped the whole system in August 2026, the rules were good. That’s the uncomfortable part.

The contract the assistant works under is better than most I’ve seen. It ranks its sources, and what you tell it in the chat comes last, below every approved fact. It labels every statement: confirmed fact, client statement, assumption, recommendation, open question. It has to show the current value next to the proposed one before it writes anything. And when it isn’t sure, it has to say conflict, missing or unconfirmed instead of guessing.

Where did the pilot break?

The pilot still broke, in five places. None of them was the model.

• The wall between clients was a sentence. Of the 16 tests we ran, the one on keeping clients apart only passed on a condition: it depended on the instructions, not on anything technical. Whether approval could be enforced by the tool, or only by people following the process, was still an open question.
• Eight of the 13 accounts were invisible, and nothing said so. The route for adding a project updated the registry the assistant reads. The route for adding an account didn’t. Following its own rules, the assistant stopped on those eight and answered, politely, that the project wasn’t registered. The busiest account in the pilot was one of them. An index that stops being updated doesn’t fail loudly. It refuses quietly.
• A document lied with authority. One account’s summary said no audit had been done yet. It sat next to twenty audit documents in the same folder. Anything reading that summary would repeat it, with total confidence.
• Dates passed silently. Two documents flagged a deadline as action required. The date came and went, and nothing in the system noticed.
• I was wrong too, and just as sure. I reported that one account was leaking budget into the wrong cities. It wasn’t. The targeting was deliberate, and the reason lived in someone’s head, not in any document. It’s recorded as a retraction. An AI told to find waste would have found the same thing, and been just as confident.

Every one of those five is the system around the model. And that assistant could only read. Now picture the same gaps with permission to move budgets.

Isn’t this the same mistake as hiring the wrong agency?

Pretty much, yes. I know because I made it.

Eight or nine years ago I was doing dropshipping and I hired an agency to run my ads, because I didn’t know how to judge the work. €2,000 in fees, about €2,000 in ad spend. Zero sales. Not few. Zero.

I didn’t lose that money because agencies are bad. I lost it because I handed over something I couldn’t evaluate, and I had no way to know it was going wrong until the money was gone.

Handing your ad account to an AI you can’t brief or check is the same mistake with a faster invoice. What I took from it became the rule I still work by: everything I try, I try on my own money first. Then it goes to a client.

What does a model fill in by itself, and what should a system give it?

The difference between an AI you can leave near an account and one you can’t is rarely the model. It’s what surrounds it: who it is answering about, what it compares against, what it does with a gap, and who signs before anything changes.

So when does autonomy make sense?

Sometimes. I’d rather say that than pretend the answer is never.

Autonomy depends on the money first. With a client spending €300 a month, it barely matters who pushes the buttons. With a client at €5,000 or €10,000 a month, you have to know very well what you’re pushing and why, and so does anything acting on your behalf.

And it depends on the order. Brief first. Then watch it be right, for long enough to mean something. Then loosen. Most people I see do it backwards: they switch it on, and the brief is whatever the platform assumed. Autonomous vs approved AI ad tools goes through that choice properly.

What I still don’t know

I haven’t found the account yet where I’d leave an AI running unsupervised. I’m honestly not sure the reason is the AI. It might be that none of us has ever written the brief well enough to deserve it.

And the brief ages. Google and Meta change what a setting does with a two-line announcement, and a brief written in September can be quietly wrong by November. I don’t have a clean answer for that yet, beyond reading it again every time a platform moves.

Questions people actually ask

Related reading

• Using AI on your ads, then checking every number. The other half of the room: the people who don’t trust it, use it anyway, and end up with more hours than before.
• Autonomous vs approved AI ad tools. Both can write. Only one waits for you, and that decides what a wrong answer costs.
• What not to automate in your ad accounts. Which half of the job is safe to hand over, and which half is where money gets lost.

Last updated: September 2026