Read-only AI access for a week first, or full write access from day one?

I'd start read-only for a week or two. Pull search terms, ask why CPA moved, see which campaigns eat the budget. You find out fast whether the assistant reads the account the way you do. Once it does, give it write access for one narrow job, like adding a batch of negatives you've already reviewed.

The argument for going straight to changes is that the preview step already protects you, so waiting feels slow. I don't think that holds up. A preview only shows what the model proposes, not whether it understood the account's history, and that's the part that bites you. It can look perfectly reasonable and still be wrong because it didn't know about last quarter's restructure or why a campaign was paused.

How do you sequence it when you're running more than one platform?

Agreed on the read-only week. Before you give it any write access, ask it why a campaign was paused last quarter. If it guesses instead of admitting it doesn't know, the preview step won't catch that, because the preview only shows you what it's about to do, not whether it understood the history. And if you're running this across platforms, test one account at a time.

The "why was it paused" test only works if you already know the real answer, so pull the change history before you run it. In Google Ads, Change history shows who paused what and when. In e-commerce the pause often isn't a person's decision at all. A Shopping or PMax campaign can go quiet because the feed started failing in Merchant Center, and the campaign view won't show that. Ask the assistant with the real answer in hand, and you'll see whether it admits it doesn't know or invents a story that sounds plausible.

You can catch a made-up story without knowing the real answer, so you don't need it in hand before running this test. Ask why the campaign was paused, then look at what the assistant cites. A confident "probably weak CTR" with no date, no edit and no status reason is the giveaway. A real answer points to something you can click and check.

On Meta this matters more than people expect. An ad set can be paused by a person, or it can end up disapproved or limited after a policy review, and those look almost identical in a summary view. If the assistant doesn't pull the status fields, it will often just guess. So judge the answer on specificity: does it name a source it could actually have read, and does it admit when it can't tell? Pulling the change history first is still a good idea, but you don't need it to get a useful signal.

Judging by specificity doesn't hold up for me. A model can name a source that sounds right, like "disapproved for a policy issue" or "pulled after a Merchant Center error," without ever reading the field. Named sources are cheap to generate, so they reward exactly the confident guess you're trying to catch. A better test is to ask it to quote the exact field name and value it relied on, then open that same field yourself. If it can't give you something you can check in thirty seconds, treat the answer as a guess, however tidy it sounds.

Quoting the field makes sense, but only if it's actually visible in the platform UI. Ad review reasons often live in a separate panel instead of the campaign list, so if you can't find the same value quickly, the test stalls and you end up digging through the change log anyway.

Fair, but I'd check the change history before anyone digs through review panels, because it answers the cheaper half of the question: did a person or a script change the status? In Google Ads, if the pause shows up there with a user and a timestamp, you're done and the assistant doesn't need to justify anything. If there's no edit at all, the status came from the system, and that's where you look for the policy or feed reason. Splitting it that way means the quoted field only has to match one of two places instead of the whole platform UI.

The stall you describe is real, though. Some review reasons only show up in a detail view, so if the assistant quotes a value you can't find within a minute or two, count it as a miss and move on instead of waiting for it to resolve.

The split works, and I'd boil it down to one rule: no write that depends on a "why" gets approved until someone has checked that why in one of those two places. If the assistant says a campaign was paused for a feed error and you haven't seen the error yourself, the reactivation waits. The preview shows you the action, but not the premise behind it.

Take a dental clinic's Shopping campaign that looks paused. The assistant proposes raising the budget and restarting it. If someone paused it on purpose for a seasonal break, that restart undoes a decision the model never knew existed.