AI can now do most of the mechanical work on a Meta account, from reporting to building campaigns to moving budgets, and Meta’s own systems already choose who sees each ad; what it cannot supply on its own is knowing which leads were actually good for this client, telling a real change from a lucky week on small numbers, and being answerable to the client, so let it read, draft and build, and keep a person on decisions that spend money.
Two things happened this year. Meta opened its ad accounts to AI assistants through an official connector, so an assistant can now read a campaign and change it. And Meta has kept pushing its own automation: Advantage+ campaigns already choose audiences, placements and budget splits, and Meta has said publicly it is working toward advertisers giving a product, a goal and a budget and letting its AI do the rest.
So the question is no longer whether AI can touch your ads. It is which parts to hand over.
What it does well already
- Reporting. Pulling numbers across campaigns and explaining what moved is fast and mostly accurate. Check the first few against Ads Manager.
- Building. Setting up a campaign, ad sets and ads from a brief is tedious work it handles well, as long as everything is created paused for a person to check.
- Delivery inside Meta. Who sees each ad, when, and on which placement is already decided by Meta’s systems for most advertisers. That part was handed to AI years ago.
- Spotting the obvious. A campaign spending with no results, an ad shown to the same people too often, a rejected ad. These are rules, and machines apply rules tirelessly.
Where it still needs a person
Not because the AI is not clever enough, but because the information is not in the ad account.
- Which leads were good. Meta knows who filled the form, not who answered the phone or bought. Only your sales team knows, and until those outcomes are sent back to Meta, any AI optimising the account is optimising for form fills.
- Whether a change is real. A cost per lead that fell from ₹400 to ₹300 over eight leads is well within chance. Acting on noise is the most common way automated accounts drift into worse setups.
- What this client counts as expensive. A ₹300 lead is cheap for a flat and ruinous for a gym trial. The right number comes from each client’s own history, not a benchmark.
- Being answerable. When a client asks why their budget doubled on Tuesday, “the AI did it” is not an answer an agency can give twice.
How to split the work
- Let AI read everything, all the time.
- Let it draft changes and build things paused.
- Keep a person on anything that moves live spend, and make sure that person can see the reasoning and the numbers behind the suggestion.
- Feed it the one thing it cannot see: which leads were good.
- Write down what each change was expected to do, and check later whether it did.
The useful question is not “can AI run the account?” It is “what does it need to know that only we know?”
For one account, a person checking an assistant’s suggestions in a chat covers most of this. Across thirty client accounts it does not, because nobody asks thirty chats a question every morning.