Claude Can Now Touch Your Ads Manager: Building Client-Ready Meta Dashboards in Seconds

A homewares brand we'd just started chatting to had a founder who'd wired Claude into his Ads Manager over a weekend and asked it one question: "what should I change this week?" It came back confident. Pause Google, it's bleeding you, spend up and revenue down. He nearly did it. Google was his best channel that month, comfortably ahead of target. The tool had read a green cell as a red one and invented a crisis out of thin air.

That's the thing I want to walk through here, because the integration is genuinely good and that story is the trap sitting right next to it.

Meta has wired Claude (and the same kind of connector for other models) straight into the ads platform. You connect it once, give it permissions, and you can ask it questions about your account in plain English and have it pull live numbers, build tables, plot charts, and yes, actually edit campaigns. We've been running it across accounts for a bit now. Here's what it's brilliant at, where it lies to you, and the prompts we actually use.

What the reporting prompts look like

The first thing worth knowing is that the output is properly shareable. Ask it for a table and you get a clean HTML table, not a wall of text. Ask for a chart and you get an interactive one you can hand to a client.

The prompts we lean on are boring on purpose. Boring is what keeps it honest.

• "For this account, pull total spend, revenue and blended ROAS for the last 7 days. Numbers only, no recommendations." A plain pull. No interpretation, nothing for it to get clever about.
• "Over the last 365 days, in groups of 7 days, show total spend and purchases in a table." This is the one that genuinely saves time. A year of week-by-week data, laid out, in about a minute. Doing that by hand is half a morning.
• "Take that same data, plot spend against purchases and CPA over time as an interactive chart I can hand to a client." Now you've got a visual a founder can actually read.

Notice none of those ask it what to do. They ask it to fetch and arrange. That distinction is the whole game, and it's where most people come unstuck.

One honest caveat on speed: a full year of data with a few charts on top took a fair while to run, and it burns through tokens fast. For a once-a-week report that's fine. Just don't expect it to be instant when you ask for something hefty.

Where it confidently gets it wrong

Back to the homewares founder. The reason the tool told him to kill his best channel is the same reason I'd tell you to never action a recommendation it volunteers: it's reading the surface of a dashboard, not understanding your business.

It misread the colour coding. Green meant good, red meant bad, and it had them backwards, so a channel that was beating its target got flagged as the emergency. When the founder pushed back, it folded instantly. "You're right, I misread the colour coding, let me reread." Which is reassuring and alarming in equal measure. It will agree with you just as confidently as it disagreed.

Here's my take on what's actually happening. The model is excellent at one job and hopeless at another, and they look similar from the outside.

What it's genuinely great at is spotting anomalies. Something three standard deviations off the norm, a day where spend tripled, a sudden CPA spike. You want a tool flagging those. That's a real edge.

What it's hopeless at is strategy, because it can't see the things that actually decide strategy. It doesn't know your cost of goods. It doesn't know you're sitting on dead stock you need to clear, or that you're deliberately running a channel at a loss to feed retention. It makes calls on the numbers in front of it and nothing else. So when it says "spend is up 51% and revenue is down 45%, this channel is the problem," it sounds like analysis. In reality it's pattern-matching on a screenshot with half the picture missing.

To put the risk in perspective: one wrong "pause this campaign" actioned without a sanity check can cost you more than the tool saves you in a month of reporting. The reporting is the safe win. The unprompted advice is the part that'll bite you.

The context layer that fixes it

Here's where it gets interesting, because the failure isn't permanent. You can fix it, and the fix is the actual skill.

The same screenshot that produced the nonsense produced a genuinely sharp analysis once we gave the model a framework for how to read it. Not more data. A method.

For us that method is the order of operations: which number is the goal, and which numbers exist only to explain that goal. Contribution margin sits at the top because that's what we're actually steering. New customer acquisition sits under it, because a margin gap while you're winning new customers is a completely different problem to a margin gap while you're bleeding them. Channel metrics sit under that, as the diagnosis layer, not the headline.

Feed it that hierarchy and the same tool that wanted to kill Google comes back with: the margin gap is small, new customer acquisition is running hot which is healthy, the channels are performing, so this is a minor miss during growth, not a crisis. Same numbers. Opposite conclusion. The only thing that changed was the lens.

So the context we load before asking anything that matters covers a handful of things:

• Your definitions. Is revenue gross or net. What goes into contribution margin and what doesn't. Get this wrong and every number downstream is quietly off.
• The order of priority. Which metric is the scoreboard and which are diagnostic. Without this it treats every red cell as equally urgent.
• Your hard noes. The things you'd never do regardless of what a number says. Channels you protect, floors you won't go under.

Meta's own assistant won't carry this for you, and that's not a knock, it's by design. When the platform's stats tool got demoed, someone asked what methodology it was trained on to make media-buying calls. The answer was that there isn't one. It's built to execute what you ask, not to hold a point of view. It doesn't know your business objectives, so it can't sequence your metrics for you. That context has to come from you, every time, or it'll default to the most literal reading of the screen.

This is the bit people miss when they say the tool is "hallucinating." Often it isn't broken. It just doesn't know what you're trying to win, so it picks an objective and runs in a direction. Give it the objective and the framework, and it stops guessing.

The editing prompts, and the guardrails on them

Reporting is most of the value, but it can edit too, and this is where I'd be careful and specific.

It can change almost any setting by being asked in plain English. The simple end is renaming a campaign. The sharper end looks like: "find every ad under a 0.5x ROAS over the last 7 days in active campaigns, list them, and let me pick which to pause individually." It pulls the list, you decide, it actions only what you confirm.

Two guardrails I'd treat as non-negotiable.

First, permissions. When it asks whether to allow an action, it offers "always allow" and "allow once." Use allow once, every time, for anything that writes to the account. Always allow hands it free rein to change whatever it likes on a single careless "yes," and that's exactly the situation you don't want with a tool that occasionally reads a chart upside down.

Second, the launch safeguard. It can only create campaigns in the paused state, which is a genuinely good piece of design. So the sensible pattern is to have it build the shell, the objective, the budget, the structure, then go in yourself, check every setting, and switch it live by hand. It can't know every preference you have. Treat what it builds as a draft, never a finished launch.

One limitation to set expectations. Right now it can't see your creative. Not your copy, not your images, not your video. So the dream workflow where it reads your top ads and spins up iterations isn't here yet. Today it's reporting, editing, and creation of shells. Useful, but bounded.

A dashboard isn't the same as knowing what to do

So here's where I land on it. This integration collapses the time between "I have a question about my account" and "here's a clean chart answering it" from hours to about a minute. That's real, and if you're running your own account it's worth setting up this week.

But a fast dashboard and clarity are not the same thing. The tool will hand you a beautiful interactive chart with a confident recommendation underneath it, and the chart can be right while the recommendation is dead wrong, because it doesn't know your margins, your stock, or what you're actually trying to win this quarter.

That gap, between the data and the decision, is the entire reason we sit on a weekly cadence with clients rather than just shipping them a dashboard and waving goodbye. The numbers are easy now. Knowing which number to act on, and which to ignore, is the part that still needs a human who knows the business.

If you've got the integration running and you're not sure whether what it's telling you to do actually holds up against your unit economics, that's worth a proper look before you action anything. Send me the recommendation it's most confident about and the context it doesn't have, and I'll tell you whether I'd trust it. What's the last thing your AI told you to change, and did you check it against your margins before you did?