Can AI replace a marketing agency? A few months ago one of our clients decided it could, and I am going to tell you what happened, because the answer is more useful than a yes or a no.
We lost him to Claude. Not to another agency. He had discovered it about a month earlier, then he found the Meta connector, and that was it. He was gone. If you are asking whether AI can replace a marketing agency, his account is the closest thing to a controlled test I have seen, so here is the whole story, including the part where he came back.
What he built, and why it was genuinely good
He built a daily automation that watched the account every morning. It pulled his metrics, checked them against his margin, then made the changes itself. Budgets up, budgets down, ad sets off.
Shortly after, he wrote his own direct response skill so the model knew how to think about copy. Then he built a small dashboard so he never had to open Ads Manager again.
Pretty cool, right? Genuinely. I am not being sarcastic. Most agencies do not have that much wired together.
Here is the part he did not know
We already had all of it. Everything is connected on our side as well. An agent watches every account around the clock. It writes into a client log every day. We run a watchlist that flags any account trending down before a human notices, and it gives us suggestions on what changes to make.
Emphasis on suggestions. We decide.
That is not us being precious about our jobs. It is us knowing what the model cannot see. Autopilot flies the plane, but there is still a pilot in the seat. Nobody boards a plane because it has great autopilot and no pilot.
What the model cannot see
An ad account is not just numbers. It is an election week killing your CPMs. It is a trend that dies on a Tuesday. It is a competitor dropping 40% off and quietly eating your auction. None of that is in the data yet, and by the time it is, you have already paid for it.
A model reacts. A marketer anticipates. The automation he built was reading yesterday's numbers and making today's decisions, which is exactly backwards for a channel where the cost of a mistake lands before the mistake shows up in a report.
Why month one always looks fine
So he ran it himself. The first month was fine. Genuinely decent ROAS. That is the part that gets people, because month one always looks fine.
You are harvesting work that was already done. The creative is still fresh. The audiences are still warm. The account structure is still the one somebody built on purpose. A well-run account has momentum in it, and momentum is the easiest thing in the world to mistake for skill.
Then month two. Then month three. And the account went where accounts go when nobody is actually steering. No new creative angles going in, no read on why the winners were winning, and a set of rules doing the same thing a little harder every day until the thing stopped working.
He came back
Long story short, he reached out a few weeks ago. Fair play to him, he did not dress it up. He said it had not worked and asked if we would take him back. That takes more than most people have. Plenty of founders would be too proud and would sit in a dying account for another six months rather than send that message.
We have taken the account back. Full funnel rebuild, same as any account we pick up: offer and margin first, then creative, then structure, then the automation goes back on top as a set of eyes rather than a set of hands. That order matters, and it is the same order we run in every audit.
Having the tools does not make you the operator. Having a football does not make you a professional footballer.
How to use AI on your ad account without handing it the keys
The lesson is not that AI does not work. We use it every single day and it is genuinely brilliant. The lesson is that a tool does not hand you the years that tell you when to ignore it. So if you are going to put a model on your account, here is the shape that holds up.
- Let it watch, not act. Daily monitoring, anomaly flags and a written log are where a model earns its keep. Budget and structure changes are a human decision, made with a reason attached.
- Judge it over ninety days, not thirty. Month one is inherited momentum. If you cannot see creative fatigue, CPM drift and a new angle landing inside the window, you have not tested the system yet.
- Keep the creative pipeline human. Rules can scale a winner. They cannot find the next one. The account died the day the new angles stopped going in.
- Give it the numbers that matter. Contribution margin, breakeven ROAS and max CPA, so it is comparing against the right line. A model optimising to platform ROAS is optimising to a number that does not know your costs.
Do not confuse access to Claude with expertise. There is still no model that outperforms someone who has actually done it, and if you would like a second pair of human eyes on an account before you decide how much of it to automate, that is exactly what our audit is for.
Frequently asked questions
Can AI replace a marketing agency?
Not the part that decides. AI can replace the execution layer of an agency: monitoring, reporting, first drafts of copy and a lot of the daily admin. What it cannot replace is judgement about things that are not in the data yet, like a competitor's sale, a dying trend or a news week moving CPMs. A client of ours tried, held up for a month, then declined for two, and came back.
Can Claude run my Meta ads on its own?
It can read your account through the Meta connector, pull your metrics and make changes on a schedule, and one of our clients built exactly that. It works as long as the account is coasting on creative and structure a person already built. Once that runs out, a model reacting to yesterday's numbers has nothing new to feed the account with.
Why do AI-managed ad accounts look fine for the first month?
Because they inherit momentum. The creative is still fresh, the audiences are still warm and the structure was built on purpose. A model can keep a well-built account ticking over for a few weeks without adding anything. The decline starts when the winners fatigue and nobody is finding the next angle.
What should AI actually do in a Meta ads account?
Watch, log and suggest. Daily anomaly checks, a written account log and a watchlist of accounts trending down are where a model genuinely saves time. Budget, structure and creative direction stay with a person, with a reason recorded for each change, and the model measures against contribution margin rather than platform ROAS.