Per-Seat AI Pricing Was Always Temporary

If your AI tooling budget looks like a SaaS line, you're about to find out it isn't one. Every serious vendor in this category is moving flat per-seat pricing onto a consumption meter, and the enterprises that planned around predictable seat math are going to open a quarterly bill that looks more like AWS than Office.

The press called the GitHub Copilot token-billing change a developer-relations fumble. It isn't. Copilot ran flat-rate during land-and-expand, the install base is mature, and Microsoft is now metering usage, which is exactly what you do once the seats are sticky. It's a deliberate monetization shift, and treating it as a PR problem lets every other vendor doing the same thing slip past unnoticed. The parallel "enterprises are rethinking AI" storyline is the same shift in a different costume. Sentiment didn't change. The meter did.

Here's the part buyers should sit with. If your justification for rolling out an AI assistant was adoption rate and seat count, you have no defense now that the per-seat number is a variable. This is exactly the failure mode of treating adoption as the goal. Counting hammers used per week was never the business case. It just looked like one while the bill was flat.

The teams in decent shape are the ones who tied the tool to a specific before-and-after, the three-days-to-fifteen-minutes kind of number, because they can still tell whether the new bill is worth paying. The teams who can't are about to discover their finance org cares how often they swing.

I don't know where the meters settle. I do know surprise variable cost is a worse problem than annoying fixed cost, and most procurement contracts were written for the second one.