Your AI adoption numbers come from the vendor who sold you the tool.
Seats activated, weekly active users, prompts per head. Those numbers exist because a software company needed to know whether it had captured a customer. They arrive free, already formatted for your steering committee.
The advice going around right now goes one level deeper: ignore the seat count, watch an ordinary Tuesday: does the AI assistant come up on its own in people’s work, without anyone nudging them to use it?
That is a better question than counting activations, but notice whose question it still is. It’s still the vendor’s question, because habitual, unprompted usage predicts renewal. If people instinctively reach for the tool, the contract gets renewed. So the “smarter” metric measures adoption more honestly, but it still measures adoption, not impact.
It tells you the habit stuck. It says nothing about what came back. Whether lead time moved, whether defects dropped, whether anything actually changed in the work.
A team can be genuinely hooked on a tool while lead time, defects, and cost sit exactly where they were. That is the normal result of accelerating one step in a system nobody redesigned.
Two questions I ask now. Most rooms cannot answer either.
Did you have a baseline from before the rollout? A number, written down, dated.
Who owns the outcome number today, and where does it get reported? Usage belongs to whoever bought the licenses. Delivery belongs to engineering. The number that connects them belongs to nobody.
The silence after the second one is the diagnostic.