Meta is removing explicit AI-usage measures from employee performance reviews while encouraging workers to test Hatch, a new agent capable of operating websites and other applications.
Three employees told Wired that internal guidance no longer ties evaluations to how extensively workers use AI tools. Engineers were also reportedly told that adoption dashboards and token counts would not be used to measure impact.
The change follows months in which some employees felt pressure to generate more AI activity even when it did not fit the task. At the same time, Hatch is consuming more computing resources than earlier chatbots and coding assistants, keeping questions about appropriate AI use active inside the company.
Evaluation shifts from usage to outcomes
Almost a year ago, Meta said workers would be graded on “AI-driven impact.” According to earlier reporting cited by Wired, that translated in practice into examining how chatbots and agents improved daily work. Usage was associated with internal labels including “AI Native,” “AI First” and “AI Enabled.”
The new review language removes references to AI usage and the AI Native designation. It says outcomes may be supported by AI or other means, restoring a broader focus on what an employee contributes rather than the tool used to achieve it.
Meta spokesperson Tracy Clayton said the update clarifies what had always been the policy: employees are assessed on contributions, and labels such as AI Native were never performance criteria. Some workers interviewed by Wired viewed the revision as subtle but welcome because it reduces pressure to use AI where it offers little value.
A lawsuit filed in July by roughly two dozen former employees gives that dispute a legal dimension. The plaintiffs allege Meta violated US antidiscrimination law during May layoffs because workers on health and family leave could not accumulate usage and were penalized. Meta denies the allegations, which remain unresolved.
Tokenmaxxing created activity without clear value
Some employees said colleagues repeatedly prompted AI systems to increase internal usage figures, including the number of tokens consumed. One worker built a leaderboard that assigned labels such as “Token Legend” based on usage. The dashboard was removed in April after details became public, and Meta later began rationing employee AI access.
The accounts suggest that a numerical adoption target can encourage visible activity without demonstrating better work. Wired does not provide company-wide data showing how common tokenmaxxing was or whether it materially affected output, so the examples should not be treated as a complete picture of Meta’s workforce.
Removing usage from reviews does not mean management has lost interest in AI capability. Employees believe leaders still want them to demonstrate effective use of the tools. The difference is that the formal guidance now emphasizes results rather than consumption itself.
That distinction matters as agentic systems require more tokens than conventional assistants. A worker can use more computing resources because a tool performs a longer autonomous task, not because the resulting work is more valuable.
Hatch tests autonomy, privacy and trust
Hatch can browse the web and operate other applications, making it similar to the OpenClaw agent. Meta employees have tested it on corporate devices for several weeks ahead of an expected public release. The company is encouraging participation but does not require it.
Two employees told Wired that some colleagues hesitate to connect Hatch to personal email, calendars and other private accounts. They worry both about privacy and about an autonomous agent making a consequential mistake. Trust was also weakened by an earlier, now-paused Meta project that tracked keystrokes and other work-device activity to gather AI training data.
Other employees have embraced Hatch for booking appointments and organizing their lives, describing it as effective. Some nevertheless question the environmental and economic cost of using large quantities of computing resources for routine tasks. Meta declined to comment on the testing.
Employees also fear that productivity gains could lead to more job cuts. Wired cites reporting that Meta considered another major layoff round after cutting 8,000 jobs in May, but abandoned it amid software bugs linked to AI reliance and slower agent progress. CEO Mark Zuckerberg has said further mass layoffs are not expected this year.
Meta’s policy change separates AI adoption from a simple volume contest. Hatch will now test whether the company can encourage meaningful agent use while giving employees enough choice, privacy and confidence to judge when automation actually helps.
