When should autopilot pause a campaign?

The most expensive mistake an autonomous ad manager can make is pausing a campaign that was working. The second most expensive is pausing a bad one too late. The line between them is drawn not by a single metric but by when the decision is taken.

No decisions during the learning phase

Platforms run an ad set in exploration mode until enough conversion signal accumulates. The CPA you see during that window is almost always higher than the real one.

Hence the first rule: nothing in the learning phase is paused and no budget is changed. Intervening early throws away the signal collected so far.

Thresholds are measured in spend, not days

"It has been bad for three days" is not a reason. Nothing is known about an ad set that spent ten dollars over three days.

• Pause — twice the target CPA has been spent with no conversion.
• Scale — CPA is below target and stayed below across the last three readings.
• Refresh creative — frequency is rising while CTR falls; the set works, the creative is tired.

The third is the one most often missed. Campaigns don't fatigue — creatives do.

Budget moves in steps

When scaling, autopilot does not double a budget; the step is clamped to a ceiling. Large jumps re-trigger the learning phase, and a winning set looks like a loser for several days.

Hands-off campaigns are left alone

For campaign types that run their own optimisation, such as Performance Max, manual intervention hurts. Those sets stay outside the rule engine's scope; autopilot only manages the budget ceiling.

Every decision leaves a record

Each decision is written down with its reason: which threshold fired, which data was read, what happened next. Without that record an autonomous system quietly turns into a black box nobody can audit.