The copilot seat audit your vendor already paid for
The usage data that settles the renewal argument is sitting in an API you own. How to read it, what counts as use, and the number that decides seats.
A client asked me last month whether their coding assistant was worth the renewal. Eleven hundred seats, list price, second year. I asked what the usage data said, and the room did the thing rooms do, so this piece is the homework I assigned them, written up for yours.
Here is the part people miss: the audit is already built. The major tools ship usage reporting because enterprise buyers demanded it years ago. GitHub's Copilot metrics API publishes per-team activity: active users, suggestions shown, acceptance behavior, by language and editor. Your license admin can pull it this afternoon. Nobody does, because the data arrives after the purchase, and after the purchase everyone's incentive is to believe.
What the pull reveals, at every company I have watched do it, is a curve, not an average. A third of seats use the tool like a limb: daily, deeply, measurably faster (your own baselines tell you how much). A middle third touches it weekly, mostly for boilerplate, fine, that is worth something. And a bottom third has not triggered a suggestion in ninety days. They are not resisting. They tried it in onboarding week, it fumbled their codebase or their editor, and they quietly went back to work. Every seat in that third is pure renewal padding.
Three honest numbers to compute before the renewal meeting. Seats with zero activity in 90 days: cut them, and let anyone who objects re-request a license, which takes a day and filters sincerity. Acceptance rate by language: where it craters (usually the legacy stack the vendor demo never showed), that is a training gap or a tooling gap, and it is fixable, which is cheaper than it is ignorable. And the spread between your top and median users: if your best people get triple the value of the middle, the cheapest productivity program you can run this year is having them show their workflow at a lunch session. The tool is the same. The habit is the product.
One warning about what not to do with the data, because I have watched this go wrong too: do not rank individuals by acceptance rate and call it performance. The metrics measure fit between tool and task, not engineer quality, and the fastest way to poison an AI rollout is to turn its telemetry into surveillance. Audit seats, not people. Aggregate by team, decide about licenses, and say out loud that is all the data is for.
The client, for the record: cut 290 seats, funded the training gap from the savings, and renewed the rest without the usual hostage negotiation. The vendor took it fine. They have the same dashboard, and they had been waiting to see if anyone would look.