I never imagined a day may come in my career where performance is assessed based on how many tokens I consume. Last week, during a call, one of our directors mentioned that a developer consumed all his tokens in a single day. “This is a sign he doesn't know what he is doing”, he said.

Now, everybody is busy preparing those .md files to be shared across projects, assuming this will reduce the tokens consumed. Copilot is moving to per token usage billing this June. This reminds me of the cloud billing story as well.

Goldilocks problem

The Goldilocks problem comes from the old story of Goldilocks and the Three Bears. One bowl was too hot. Another was too cold. The last one was “just right.”

For example, starting with a modular monolith before moving to microservices is a Goldilocks approach. Not “everything is a service.” Not “one giant mess.” The goal is to find the level of decomposition that is “just right” for the current system.

The same thing is now happening with AI agents and token consumption. You want to be consuming enough tokens but not consuming too many tokens. You want to get it just right.

Hottest metric

One guy told me that his CTO mentioned “$150 a day.” It’s performance review season, and suddenly directors are being asked to report token usage numbers for their teams to feed the new productivity metrics.

A few months ago, managers were judging developers for not using AI enough. Now that the bill has arrived, everybody is watching token consumption.

More tokens means more cost. Management believes this creates a transparent picture of who is doing what.

Not enough enthusiasm

One developer wrote me that he was recently laid off for “not enough enthusiasm toward AI” which sounded like a corporate euphemism for: “you burned the least tokens.”

Another said his CEO recently used the word “enthusiasm” during an all-hands meeting. Only later did he realize what that enthusiasm was being measured by.

When the bill arrived

In a sense, the AI bubble popped a while back, and companies are only now starting to realize it as they begin looking seriously at AI costs. The heavily subsidized “AI everything everywhere for everything” phase is starting to slow down, especially as investors begin asking harder questions about returns.

The comparison to the early cloud days keeps coming up. At first, every engineer had unlimited access to spin up cloud services everywhere. Companies pushed teams to use the cloud for everything. Then the investor cash slowed down, the real bills arrived, and the conversation changed.

We are now noticing the same shift with AI. A few months ago, teams were judged for not using AI enough. Now companies are watching token consumption, AI costs, and usage dashboards much more closely.

AI is not going away. What is fading is the idea that every product needs an AI chatbot and unlimited token consumption with no business justification.

Goodheart's law

When a metric becomes a target, people start optimizing the metric instead of the real goal.

This is like measuring success by how many lines of code you write.

We used to laugh about managers calling developers at the end of the day asking how many lines of code they wrote. The metric was broken because developers could write more code instead of better code.

If teams are measured by tokens, people may start burning tokens to appear productive. The developer consuming more tokens may actually be trying multiple approaches and validating ideas. The developer consuming fewer tokens may simply be copy-pasting the first thing the model produced.

The metric slowly stops reflecting reality because people adapt their behavior to game it.