Why Meta Stopped Grading Employees on AI Use After Tokenmaxxing

Meta’s reported reversal on AI-based performance scoring reflects a simple management problem: using more AI is not the same as doing better work.

Earlier in 2026, AI use became a visible workplace metric at several technology companies. At Meta, employees reportedly faced language around AI use and “AI Native” status, while internal token-usage rankings helped turn model consumption into a competition. The resulting behavior acquired a name—tokenmaxxing—or maximizing AI-token use to improve one’s standing. 237

Meta has now reportedly removed AI usage and “AI Native” status from performance-review criteria. The new approach is meant to judge contribution and outcomes, whether work is supported by AI or achieved by other means. 79

Why token counts became a bad performance metric

A token is a unit of text that an AI model processes. It can be useful for measuring system demand and cost, but it is a weak measure of employee performance.

A worker can generate more prompts, run longer tasks, or use a model on low-value work without improving the quality, reliability, speed, or business value of a result. Once usage is linked to recognition or reviews, the metric can become the goal. Reporting across the industry described employees inflating AI use to game rankings, while companies faced rapidly rising inference costs. 11116

That is the core failure of tokenmaxxing: it optimizes a proxy for productivity rather than productivity itself.

Meta reportedly told employees it would limit AI use after an “exponential increase” in costs, and reporting says Meta and Amazon removed token-use leaderboards. 12 The retreat does not show that AI has no workplace value. It shows that raw consumption cannot establish that value on its own.

What contribution-focused reviews are designed to change

Reported Meta guidance removes AI usage from evaluations and tells managers not to use token totals or AI-adoption dashboards to judge impact. 79

In practical terms, that should move the review question from “How much AI did this person use?” to “What did this person deliver?”

A contribution-focused assessment can center on the outcome of the work, such as:

• quality and correctness of the deliverable;
• reliability and maintainability of systems or processes;
• reach and usefulness to customers or internal teams;
• business or operational impact; and
• sound judgment in choosing when AI helps and when it does not.

The reported wording is important because it allows the same outcome to be credited whether it was supported by AI or accomplished through other methods. 79 AI remains available as a tool, but it is no longer supposed to be the scorecard.

Why Meta is still testing Hatch

The review-policy change is not necessarily a retreat from AI adoption. Meta is also reportedly encouraging employees to test Hatch, a more autonomous AI-agent project. 7

That distinction makes strategic sense. An agent that can take actions across websites and applications is a substantially different product proposition from a chatbot that only produces text. Internal testing can help identify whether an agent completes tasks accurately, handles permissions safely, and produces a user experience worth trusting before any broader release.

The key difference is the metric. For a product such as Hatch, the useful evidence is not simply how many tokens testers consume. It is whether the agent reliably completes a valuable task with appropriate user control and without creating errors, security problems, or costly supervision.

The broader lesson for workplace AI

Meta’s reported change fits a wider industry shift. Companies that initially urged employees to use as much AI as possible have encountered a harder question: does that spending create measurable gains? Reporting has described rising AI costs without a matching increase in productivity, pushing some employers toward limits, budgets, or more selective deployment. 11011

For employers, the better approach is to treat AI adoption as an experiment tied to outcomes:

1. Define the work that should improve.
2. Measure quality, time saved, reliability, and cost—not just tool activity.
3. Keep human accountability for decisions and output.
4. Compare AI-assisted work with an appropriate non-AI baseline.
5. Stop incentives that reward needless model use.

The central takeaway is that AI usage data can still be operationally useful, especially for capacity planning and cost control. But it is a poor substitute for evaluating employee contribution. Meta’s reported policy change recognizes that distinction: reward the work’s impact, not the amount of AI consumed to produce it. 79