AI agents for Meta ads: what they can run, and what stays yours

AI agents for Meta ads stopped being a demo on 29 April 2026. That day Meta opened its ad system to outside assistants, and Claude or ChatGPT could suddenly create a campaign in a real account from a sentence. We were already building an ads pipeline around agents, and this guide is what we would tell an agency starting now: what the agent can run, where it goes wrong, and which decisions stay with a person.

What an AI agent can do in a Meta ad account today

An AI agent connected to Meta's Ads MCP server can read performance, search the Ad Library, build campaigns, ad sets and ads, change budgets, and manage catalogues and audiences, all from plain language. It works with the permissions of whoever logged in. What it can't do by itself is know your client's brand or decide an ad is fit to run.

The connection is one address, mcp.facebook.com/ads, and a Facebook Login. Meta's connectors launched in open beta with 29 tools and write access from day one, which Google's and Amazon's equivalents did not have. The list has grown since. It is built on the Model Context Protocol, so the same server answers Claude, ChatGPT, the Gemini app, or an agent you run from a terminal.

So the plumbing is solved, and it is free. That changes what's worth paying attention to.

What agents are good at, and what they get wrong

We've run the same agent across reporting, research, writing and launching. The results were uneven in a consistent way.

The research row deserves a number. We searched the Ad Library for "skincare" through the connector and got back a bar, a butcher and a fiction publisher among the results. About a quarter were off-target, because the search matches body text and targeting metadata and has no idea what the product is. An agent that saves everything it finds will then write from the wrong category. Ours checks the page name and the link title before it keeps anything.

The part the connector leaves to you: making the ad

Here is where most setups stall. The connector can create an ad, but an ad needs a creative, and the creative is the whole game now. Since Meta's Andromeda retrieval system rolled out, the creative does most of the targeting and near-identical ads get treated as one. An account needs a steady supply of ads that differ in their argument, in more than their wording.

An agent can supply that. It needs four things a chat window doesn't give it:

• The brand's material. Product photos, the logo, colours, fonts, and writing the brand sent to real readers. An agent with none of it produces the average of every ad it has seen.
• Somewhere to put drafts. A person has to see each concept as it will run. A chat scroll is a poor review queue.
• A way to assemble, not draw. Agencies have used generated imagery for years without asking a model for a finished ad. They generate parts and assemble them around the real product with type set by hand. An agent should work the same way.
• A record of where each ad came from. Which angle, which hook, which format. Without it, results can't be traced to a cause, and the next batch is a guess.

That last one is easy to skip and expensive to add later. When every ad carries its attributes as fields, "which hook type wins for this client" is a query. When they live in a prompt somebody typed, it's an afternoon of someone's memory.

Four ways an agent goes wrong in an ad account

None of these needed a bad model. Each happened to us with a capable one.

It works in the wrong client's account. One token reaches every account you manage, and an agent given no account will use whichever is the default. For an agency that means client A's ad under client B's Page. We cover the rules that prevent it in Meta Ads MCP for agencies.

It spends on its own judgement. Our agent rendered images nobody had asked for, agreed to stop when told, and a fresh session had no memory of the agreement. Instructions didn't hold. What holds is described in the guardrails for an AI agent running Meta ads.

It ships words nobody read. We found six lines of ad copy sitting inside an image prompt, where the approval screen never showed them. They rendered. The fix is a rule about where copy is allowed to live, and it's the centre of an approval workflow for AI-made ads.

It learns from luck. One conversion on six dollars of spend makes an ad look like the best in the account. Our first results report ranked on raw cost per purchase, so that ad's attributes led the next batch. We now put a 95% interval on every rate and rank on the pessimistic end of it: a combination has to be good even at the bad edge of its range before the agent is told it's a winner.

How an agency sets this up

The order matters more than the tools.

1. Connect read-only first. Give the agent reporting for a fortnight. You learn how it reads the accounts, and nothing can go wrong.
2. Make each client its own project. Its own ad account, Page, brand material, audiences and results. Nothing shared, nothing defaulted.
3. Load the brand before the first batch. Product shots, logo, palette, fonts, and a dozen pieces of the brand's real writing, grouped by who they were written to.
4. Let the agent draft, and put a person on approval. Every concept, every word on the image, read by someone who answers for the brand.
5. Launch paused. Activation is a separate act by a person, in the account, on purpose.
6. Read results by attribute, with their uncertainty. Then feed what's clear back into the next brief, and what's rejected too.

Step two is the one agencies skip, because one workspace for everything is quicker on day one.

What stays a person's job

Some agencies have gone the other way and switched Meta's own AI creative features off, because the changes arrived switched on and the brand had no say. That instinct is right, and it applies to your own agent as much as to Meta's.

We keep four things with a person and have no plans to move them. Approving an ad. Switching it on. Setting how much may be spent. And saying what the brand looks and sounds like: an agent may read a client's colours from their website and propose them, but a person says yes before anything is saved. A model choosing a brand's colours would be inventing its identity.

Everything else is the agent's, and that's most of the hours.

Where to start this week

Connect an assistant to one account you own, read-only, and ask it which three ads are tiring and why. That costs nothing and shows you how it thinks. Then pick one client and write down, before any tool is involved, who approves an ad and who switches it on. If you can't name them, sort that out before you connect anything with write access.