Sign in Subscribe
By Remco Livain in AI —

I pointed the Meta MCP at a real ad account. Here's what it found in ten minutes.

A few weeks after wiring Claude into Meta Ads Manager, I ran a real ROAS audit on a live account. One zombie campaign, a hidden optimisation trap, and a lesson about where the human judgement still lives.

A few weeks ago I wrote about Meta opening Ads Manager to Claude through their official MCP server, and how to wire it up. That post was about the plumbing. This one is about what happens when you actually point it at a real account with real money on the line.

This week I ran my first proper ROAS audit through it — a consumer brand I look after, spending just under €20,000 a month on Meta across eight campaigns. I wanted to know two things. Could the AI surface the problems faster than I could clicking through Ads Manager myself? And could it execute the fixes without me touching the interface at all?

The short answer to both is yes. But the interesting part is where the value actually showed up, because it wasn't where I expected.

What it found in ten minutes

I asked for the last 30 days, broken down by campaign, with opportunity score and auction-ranking diagnostics. The table came back clean. And it was ugly.

Of eight campaigns, exactly one was ROAS-positive — a prospecting campaign returning 3.2x on a small budget. Everything else was either flat or underwater. The worst offender was a top-of-funnel geo test that had been left running since last summer, quietly burning around €6,000 a month at a ROAS of basically zero. Nobody killed it. It just kept spending, month after month, because no one was looking at it the right way.

That alone justified the exercise. But the AI also caught something subtler that I'd have taken much longer to spot manually: most of the "Sales" campaigns were optimised for add-to-cart, not purchase. That's the kind of configuration mistake that hides in plain sight, and it's a well-documented way to silently kill your ROAS. The campaigns look like they're doing their job — the dashboards are green, carts are filling — but you're paying Meta's algorithm to go and find people who like clicking "add to cart" and never check out. The rule of thumb the good media buyers follow is simple: optimise for the lowest-funnel event you have enough volume to feed, and for an account this size that's purchase, not cart. Switching that single setting is almost certainly worth more than any new creative round.

The expensive problems in an ad account are rarely the creative. They're the settings nobody revisits and the tests nobody remembers to turn off.

What it did, not just said

Here's the part that still feels new. I didn't export a report and go fix things by hand. I had it execute, live, from the chat:

It scaled the winning prospecting campaign by 50%. It paused ten ads sitting in the bottom of the auction rankings for conversion rate — several of them the click-baity type that earn cheap engagement and no sales. And it did the maths on the trade so the net daily spend barely moved: roughly €70 a day freed from dead ads, almost exactly offsetting the extra I put behind the one campaign actually making money. Same budget, pointed at things that work.

Every action was reversible — status changes only, no deletions — and Meta's connector has a nice safety reflex where budget changes auto-pause the campaign until you confirm you meant it. I had to reactivate the one I'd just scaled. Mildly annoying, entirely correct.

The judgement was still mine

I keep coming back to this because it's the whole point. The AI was extraordinary at the gathering: pulling thirty data points, ranking them, flagging the leaks, doing the reallocation arithmetic. What it could not do was decide. It surfaced the dead geo test, but the call to actually kill €6k of monthly spend belongs to a human who knows the account, the brand, and what that budget was supposed to be testing. It flagged the add-to-cart problem, but understanding why that's the real ROAS drag — and not just chasing the surface-level "low ROAS, make new ads" reflex — comes from having done this for twenty years.

This is the part I think a lot of people are getting wrong about AI in marketing right now. Hand the same connector to someone without the reps, and they'll get a beautiful table of numbers and no idea which one to act on — or worse, they'll act on the wrong one. They'll make new creative for a campaign whose problem is a tracking setting. They'll pause the prospecting campaign because its raw ROAS looks scary, not realising it's the only thing feeding the funnel. The tool doesn't supply the judgement. It just removes the hours of clicking between the judgement and the result.

This maps onto something I've written about before — that generic AI produces generic outputs, and the leverage now sits with the person who can direct it with real domain context. It's not just my opinion. McKinsey's 2025 State of AI work found that while 88% of organisations now use AI somewhere, only around 6% are capturing meaningful enterprise-level value from it — the gap is execution and operating model, not access to the tool. Iansiti and Lakhani made the structural version of the same argument in Competing in the Age of AI: the advantage accrues to those who rebuild how decisions get made around the data, not those who simply buy the software.

What this means for how marketing teams are built

Three years ago, an account in this state would have needed a junior account manager living inside Ads Manager to even notice the geo test, then a few back-and-forth meetings to agree to pause it. This week it took one operator, one conversation, about twenty minutes, and the fixes were live before lunch.

I don't think that means fewer marketers. I think it means a different shape: fewer hands copying numbers between tabs, more senior judgement per dollar of media spend. For me personally it means I can do work that used to require waiting on a network of specialists to pull reports and feed me numbers — I get the insight directly now, and I act on it the same hour. A smaller team, a tighter loop, more done. The skill that's appreciating in value isn't operating the tool; the AI does that now. It's knowing which number actually matters, and having the nerve to turn the expensive, comfortable, broken thing off.

The dead campaign was the easy win. The settings nobody had revisited were the real one. The AI found both in the time it takes to make a coffee. Deciding what to do about them is still the job — and after two decades of doing it by hand, that's the part I'm not worried about.

Sources & further reading

External

Related posts

Subscribe to Remco Livain

Don’t miss out on the latest issues. Sign up now to get access to the library of members-only issues.
jamie@example.com
Subscribe