AI assistant and budgets: confirm every change, or approve a standing rule?

I keep going back and forth on this. With an AI assistant moving budget between campaigns, the safe default is approving each change as it comes up. With 15 campaigns that gets tedious fast, so people start saying "just shift anything under target to the winners every Monday." It sounds efficient, but the rule has no idea why a campaign is under target. On a weekly view, a new campaign still in its learning phase looks exactly like a dead one.

My view is that a standing rule is fine for small capped moves, like no more than 10% a week, and anything bigger should still get a human yes. Where do you draw the line, especially if a rule has ever moved money during a seasonal dip?

I'd put a time filter in front of any percentage cap. Conversions often land days after the click, so a campaign's last seven days look worse than they really are. A dental clinic whose bookings close a week later would get cut every Monday for no real reason.

A time filter doesn't fix this, at least not for e-commerce. Conversion lag matters, but most product purchases close within a few days, so a seven-day window already covers most of it. The real problem is what the rule does to the campaigns it feeds. A budget increase on a Smart Bidding campaign, especially Performance Max, can push it back into learning, so the winner that got extra money on Monday often looks worse a week later. The rule reads that dip as a reason to move money again, which means it ends up causing the very signal it reacts to.

So limit how often money moves, not just how much. One shift every two or three weeks, with a hard cap on each move, keeps campaigns out of constant relearning. A time filter alone just makes the loop slower.

Agreed on the lock-out, but I'd run it per campaign instead of per account. If a campaign got a budget increase in the last two or three weeks, the rule should leave it alone whatever its numbers say, otherwise it keeps reacting to the relearning it caused itself. What I can't pin down is how long that window should be. It probably depends on conversion volume, since a campaign with a handful of conversions a week takes much longer to settle than one with hundreds. How many conversions does a campaign need before its weekly number means anything?

It depends on how many conversions each bidding unit gets, not on the campaign total. Meta tracks learning per ad set, and the usual rule of thumb is about 50 optimization events a week to exit it. So a campaign with 300 conversions spread across six ad sets can have every ad set starved, while a single-ad-set campaign with 60 purchases settles fine. Google Smart Bidding works at the campaign level, so the same account-wide number means something different on each platform.

That's why I'd tie the lock-out to the same count. If a budget raise only hit ad sets still under that threshold, lock them for longer. Ad sets already past it settle faster, so a fixed two or three weeks is too blunt for both.

Agreed on counting per ad set, and for local services the real number is often smaller than it looks. A plumber's or clinic's "conversion" is usually a phone call, and it only counts if it clears the call-duration filter. Say a law firm gets 40 calls a week but only 9 last past a 60-second threshold. The bidding system sees 9 events, so an ad set that looks busy in the call log can still be starved. Count what the platform actually receives, not what the phone shows.

Agreed, and on Meta the gap can run the other way. Say the law firm optimizes its click-to-call ad set on the call button tap. Meta counts every tap, including calls that hang up after 20 seconds, so the ad set can look healthy in Ads Manager while the events it learns from are mostly junk. Then your lock-out rule gets two contradictory signals: Meta says the ad set is fine, the call log says it's starved.

I'd pick the optimization event based on what counts as a real lead, even if that means fewer events. If the call tracker can pass the qualified call back, sending it through the Conversions API as a lead event is the cleanest option, with some lag. That way the number in Events Manager for that ad set is the one to trust, because it's the same count the learning phase uses.

That's the case for making the rule check agreement before it acts at all. A good rule of thumb: if the platform's conversion count for an ad set and your own count (CRM bookings, qualified calls, whatever you actually care about) differ by more than about 20% over two weeks, the rule should leave that ad set alone. A gap that size usually means the optimization event is measuring something other than a lead, so any budget decision built on it won't hold up. Fix the event first, then let the rule back in.

The 20% threshold looks shaky to me, or at least the way the gap gets measured. Platform counts and CRM counts rarely line up on the same day. Google and Meta attribute by click or interaction date, while a CRM usually logs when the booking was made, and offline imports can lag a day or more. So a two-week gap of 20% can come from timing alone, and a rule that trips on it will freeze good ad sets for no reason.

Compare the same date window on both sides and watch the ratio over several weeks. A steady 60% match is usually something you can live with. A drop from 90% to 60% is the real alarm, because it means something changed in the event, the form or the call filter.