Regression to the mean is a statistical effect, not a behavioural one. Any measure that moves from week to week will have extreme weeks, and an extreme week is more likely to be followed by an ordinary one than by another extreme. Nothing has to change for the number to improve.
It matters for AI pilots because pilots are usually placed where performance was worst, at the moment the problem was most visible. When the next quarter looks better, the improvement is credited to the AI, even though part or all of it would have happened anyway. The same error appears in safety programmes, school interventions and turnaround plans.
The protection is a baseline taken over a full business cycle and, where possible, a comparison group measured over the same weeks. If the pilot team was chosen because it was struggling, the comparison group should be chosen the same way.