Tokenmaxxing Is Not the Goal

Microsoft set division-level AI token budgets this week. Switched the default internal model from frontier to GPT-5.6. Built a dashboard so managers can see exactly what each engineer spends. Jay Parikh’s email to staff said it plainly: “Tokenmaxxing is not what we are optimizing for.”

This from the company that shipped Copilot into Office, Teams, and GitHub, then pitched it as the future of work to every CIO with a budget. Their own internal data showed the problem. Engineers burning hundreds to thousands per month in tokens. Some productive. Some habit. Nobody could tell which was which.

Most teams I work with measure AI success by adoption. Seats activated, prompts sent, workflows touched. Microsoft just proved those metrics tell you whether AI is being used, not whether it is working. Different questions. The fix was not pulling AI back. It was building an operations layer: budgets per division, visibility per engineer, frontier models only when the task demands it.

Seventy-three percent of orgs say AI is used regularly. Ten percent say it is core to operations. That gap is where the money disappears. Microsoft found it inside their own building and started closing it. Not with a better model. With better discipline.