We Connected Claude to Our Clients' Facebook Ad Accounts - Here's What It Can and Can't Do

"So does this mean you don't need a media buyer anymore?" a client asked me on a call a few weeks back, half-joking, after he'd seen a screenshot of Claude pulling his ad account spend into a tidy table. Fair question. It's the question everyone's circling right now.

So we did the obvious thing. We connected Claude to a set of real client Facebook ad accounts through the official Meta integration, and we ran it for a stretch across live spend. Reporting, pacing, edits, the lot. This is the honest field report: where it genuinely saved us hours, where it's quietly dangerous, and the discipline that keeps it from torching a live account.

No hype. I run Meta for ecommerce brands for a living, so I care about what this actually does on a Monday, not what it might do in a keynote.

What we actually plugged in

The setup is a connector. You add the Facebook Ads integration inside Claude, authorise the business you're working on, and it can now read and act on the accounts you've granted. Once it's connected, you can talk to your ad account in plain English: pull last 7 days' spend, plot it against purchases, change a budget, rename a campaign, launch a shell.

One thing worth saying upfront: it got better the more we used it. Early on it fumbled which tool to reach for. A few days in, it knew. That's the texture of a V1 tool, and it's worth setting expectations with your team accordingly.

Now the scorecard.

Where it saves real hours

Reporting and ad-hoc questions. This is the standout. "Pull the last 365 days, group by week, show spend, revenue, ROAS and purchases in a table" used to be a few minutes of exports and spreadsheet wrangling per account. Now it's a sentence, and you get a clean table or chart back. Across a roster, that's hours a week you get back. For a quick read on a Monday, or a number a founder wants right now on a call, it's genuinely faster than opening Ads Manager.

Anomaly-spotting. Large language models are good at this, and so is this one. Point it at the numbers and it'll flag the things that are a few standard deviations off the norm, the spend spike, the campaign that fell off a cliff. That's the right job for it. You want to be told when something's abnormal, and it surfaces that well.

Pacing checks. "Are we on track against this month's budget, and which campaigns are over or under?" is a question it answers quickly and repeatably. One of the better ways we've used it is as a fast pacing read across several accounts in a row, instead of clicking through each one.

The mechanical edits. Renaming campaigns, nudging budgets, building a paused campaign shell so a human can finish the setup. It does the boring button-pushing fine. If you've ever lost twenty minutes glitching your way through Ads Manager to make a small change, having it done before you finish the sentence is a real saving.

To put a number on the time side with an invented but realistic figure: on an account where weekly reporting and pacing used to eat maybe three to four hours, we've clawed back the bulk of that. Not because the tool is clever, but because it removes the manual export-and-format tax. That's the honest size of the win.

Where it's dangerous

Here's the part the tutorials skip, and the part that matters more.

It recommends on numbers, not on your business. This is the big one. The tool can see what's in the ad account. It cannot see your cost of goods, your inventory, your margin, your back-end LTV, or what you're trying to do this quarter. So when it tells you an ad is "underperforming", it's reading the surface. I can't count the times we've decided to scale an ad that looked bad on Meta because everything off-platform said it was actually winning, or held one back that looked fine on ROAS because it was quietly hurting the account longer term. The agent doesn't have that context, and acting on its broad recommendations is how you make confident, well-formatted mistakes.

It can't see your creative. In this version it can't pull your copy, your images, or your video out of Ads Manager. So any "why is this working" question it answers is blind to the single biggest driver of Meta performance: the creative itself. It's reading the scoreboard with the players off the pitch.

"Pause everything under 0.5x ROAS" is a trap. It's the move people will gravitate to, because it's easy to ask for. And on the surface it'll happily find those ads and pause them. But a blanket ROAS rule with no view of learning phase, attribution windows, off-platform contribution, or what that ad does for the customer journey is exactly the kind of decision that looks tidy and costs you. The agent will execute it cleanly. That's the danger, not a safeguard.

It optimises toward the platform's view. Best practice on Meta is not the same as the best decision for your brand. Good media buying is largely the work of pulling the platform toward what your business actually needs, which is the opposite of just doing what the account's surface numbers suggest. An agent pointed only at in-account metrics drifts toward the platform's definition of good, not yours.

The permission discipline that protects a live account

This is the bit I'd make non-negotiable for any team turning this on, because a live account is real money.

By default, every action sits in "needs approval". Good. Leave it there. The temptation is to hit "always allow" the first time it asks, because the prompts feel repetitive. Don't. "Always allow" hands it free rein to edit whatever it wants off the back of a casual "yes" in a tiny exchange, and that's how a careless instruction becomes a live change you didn't sanction.

The discipline we settled on is simple:

• Read-only things run free. Pulling insights, building reports, answering questions. No risk, let it go.
• Anything that changes the account is approve-once, every time. Budgets, statuses, launches, renames. You approve each action as it happens, not as a standing permission.
• Launches stay paused. The tool can only create campaigns in a paused state, which is a genuinely good safeguard. Use it. A human eyeballs every setting before anything spends a cent.

That setup gives you the speed on the safe stuff and a human gate on anything that can hurt you. It is the whole difference between a useful assistant and a loaded foot-gun.

Our actual position: it lowers the floor, not the ceiling

Here's where I've landed after running it.

The honest read across the industry is that average or bad media buying was already on its way out, and tools like this finish the job. If your "media buying" was logging in, pausing the red ads, scaling the green ones, and doing what Meta suggested, an agent does that now, and probably does it faster. That floor has come up. A local business that never wanted to touch Ads Manager can describe its product and be spending sensibly tomorrow. That's a real and good thing.

But the ceiling hasn't moved. The work that actually grows an account, knowing which off-platform data point to look at next, reading a creative, weighing inventory and margin and the longer game, deciding to back the ad that looks wrong because you understand why it's right, is exactly the work the agent can't do. It has none of that context. If anything, that work gets more valuable, because the mechanical layer underneath it just got commoditised.

So the way I'd frame it for any founder weighing this up: an AI agent makes a competent operator faster and makes a weak one look passable. It does not make a strategist out of a tool. The edge still sits with the person who has the context, the off-platform picture, and the judgement to overrule the obvious-looking call. The agent just clears the busywork out of their way so they spend more time on the part that matters.

We're using it. Daily, on the reporting and the pacing and the mechanical edits, with the permissions locked down hard. We are not letting it make strategy calls, and we won't until it can see everything the account can't show it, which is most of what actually matters.

If you're experimenting with the integration on your own accounts, I'd genuinely love to hear where it's saved you time and where it's burned you. That ledger is being written in real time right now, and the more honest field notes we trade, the faster everyone works out where the real line sits.