Practitioners of online controlled experiments describe novelty and primacy effects: early results differ from the settled effect because users are reacting to the change itself. Some engage more with something new; others are slower until they learn it. Either way, the first few weeks are a poor guide to the long run.
For AI pilots the risk is mostly upward. A new assistant gets attention, enthusiastic early users and close support from the delivery team, and the first month’s figures reflect all of that. A business case built on those weeks can overstate the benefit that remains once the tool becomes routine.
The practical response is to judge a pilot on the later weeks of its run, plot the effect over time to see whether it is fading, and schedule a benefit review six to twelve months after scaling, when the novelty has gone.