What to never let Claude do in your ad account

Most advice here is a list of scary things phrased as vibes: be careful, use guardrails, keep a human in the loop. None of it tells you what to do on a Tuesday when the model offers to shift budget from a losing ad set to a winning one and it sounds completely reasonable.

The four classes do, because they sort by how expensive it is to be wrong rather than by how risky the action sounds.

The four classes

The useful property of this split is that it survives contact with a persuasive argument. When something sounds sensible at 11pm, you are not deciding whether it is a good idea. You are checking which box it lands in.

The one that catches people

Financial is obvious once written down. The class people get wrong is Status, because it feels administrative and it is the one an assistant will offer to batch.

"Shall I pause the five ad sets that haven't converted this week?" is a Status action wearing Read's clothing. It sounds like tidying. It is five delivery decisions made on one sentence, and if the underlying window was wrong — a partial week, a shifted attribution setting — you have just paused five things for a reason that does not exist.

The rule that holds: an action on more than one entity needs the entities named. Not "the underperformers." The names.

These four classes are the spine of the whole weekly routine — they decide what Monday, Wednesday and Friday are each allowed to do. See how the week is split →

Check what you have already agreed to

You probably have at least one standing arrangement you would fail to justify if asked — a saved instruction, a project note, or simply a habit of approving things quickly. This is checkable, and it takes a minute:

List every change made to this ad account in the last 30 days,
grouped by who or what made it. For each, say whether it was a
budget or bid change, a status change, or a new draft.

Every change an agent makes lands in the activity log like any other change. Read the list. If there is a budget or bid row you did not personally decide, that is your answer and it is worth knowing tonight rather than at the end of the month.

Where the line moves with spend

• At the smaller end of the range. Status is the class that matters, because with few enough ad sets a single wrongly paused one is a meaningful share of your delivery. Financial barely comes up — you are not moving budget in a hurry.
• In the middle. All four classes are live at once, and Draft is where the leverage sits — enough creative throughput to matter, few enough people that you are still doing it yourself.
• At the top of the range, with a team. The classes stop being personal discipline and become access control — portfolio-level rules, one project per account, and the activity log as an audit trail rather than a curiosity.

What this is not

It is not a claim that the model is dangerous. Read and Draft cover most of what is useful, and both are free of consequence. The restriction is narrow and it is aimed at exactly one thing: an irreversible action taken on a number nobody checked.

Which is worth saying plainly, because the failure mode here is almost never dramatic. It is a confident wrong figure that nobody questioned, turned into a budget change by a system that was allowed to make one.

Try this tonight Run the activity-log prompt above for the last 30 days. If every row is Read or Draft, you are already operating the way this post describes and you can stop reading. If there is a Financial row you did not decide, that is the one thing to change this week — and the change is a sentence in your project instructions, not a new tool.

If you have not connected an account yet, the setup covers what the connector actually exposes, and the same four classes are written into the free skill.