Microsoft just put its engineers on AI token budgets.
A leaked internal memo from EVP Jay Parikh says "tokenmaxxing is not what we are optimizing for": divisions get token budget targets, engineers can track their own spend, and a cheaper model became the internal default. Some engineers were reportedly burning hundreds to thousands of dollars a month.
I fully support this. Tokens are a resource like any other, and resources need budgets.
But I think there is a bigger story here. With AI agents doing the work, we are getting something that was virtually impossible with human engineers: the cost of a feature, calculated in tokens, which means in money.
You could never honestly answer "what did this feature cost?" with people. A salary is spread across meetings, reviews, rework, and thinking in the shower. Tokens are metered like electricity.
People have not left the process, of course. Someone still directs the agents and reviews their work, so tokens are not the full cost yet. But the metered share of development grows every month.
You cannot manage what you cannot measure. For the first time in software, we are starting to measure this.
The value side was always measurable: customers decide what is valuable by using a feature, or not using it. Now the cost side catches up.
So here is a question for engineering leaders: if you saw the real price of every feature next to its real usage, which ones would you have never built?
First published on LinkedIn. Discussion on LinkedIn