Meta Drops AI Usage Quotas but Rolls Out Hatch, a New Agentic Tool

Meta told employees this week that it will stop using AI adoption dashboards and token counts to evaluate job performance, ending a system that had employees racing to maximize AI usage to score well on reviews. At the same time, the company is rolling out an internal agentic tool called Hatch, which can autonomously browse the web and control other applications, to workers on their corporate devices ahead of a planned public release. Three Meta employees shared details of both announcements with WIRED.

What happened

Detail	Fact
Policy change announced	This week (internal announcement)
Old criteria	“AI Native,” “AI First,” “AI Enabled” designations tied to token usage
New criteria	Impact-based; outcomes “can be supported by AI or other means”
New internal tool	Hatch: agentic AI, browses web, operates apps on corporate devices
May layoffs	8,000 jobs cut; about two dozen affected workers filed an antidiscrimination lawsuit in July
Further layoffs	Plans reportedly called off after AI-linked software bugs and slow agent progress

For roughly a year, Meta had graded workers in part on their “AI-driven impact,” which in practice meant tracking how many tokens (the units of text processed by AI models) employees consumed. The system produced labels like “AI Native” for heavy users. One employee built an internal leaderboard, calling top users “Token Legends.” When that dashboard leaked internally, it was taken down in April. A couple of months after that, Meta began rationing AI usage.

The new guidance, confirmed to WIRED by Meta spokesperson Tracy Clayton, removes those labels from the review rubric and tells engineers the company “will not use AI adoption dashboards or token counts to evaluate impact,” according to The Information. Clayton added that labels like “AI Native” were never formally used in performance evaluations, a claim that sits in tension with the lawsuit filed in July by about two dozen employees who say the usage metrics contributed to unfair treatment during the May layoffs.

Why it matters

Tying compensation reviews to a raw usage metric created a predictable problem: people optimized for the metric rather than the outcome. Employees were prompting AI tools repeatedly, sometimes in ways colleagues described as frivolous, purely to run up their token counts. That is a textbook example of Goodhart’s Law, where a measure becomes a target and ceases to be a good measure.

The lawsuit adds legal weight to the policy reversal. Workers on health and family leave alleged they could not accumulate usage and were penalized as a result. Meta has denied those allegations and the case is ongoing.

Meanwhile, the rollout of Hatch raises its own questions. The tool is described as similar to OpenClaw, the viral computer-use agent that attracted attention earlier this year. Employees can already use it on corporate devices. Some have used it to book personal appointments and organize their calendars outside work. But trust is shaky: Meta previously ran a project that logged employee keystrokes and device activity to gather AI training data. That project has since been paused, but employees say it damaged their confidence in company AI tools. Several are reluctant to connect Hatch to personal email or calendar accounts.

There is also a bigger workforce anxiety underneath all of this. Reuters reported last week that Meta was planning another significant round of layoffs on top of May’s 8,000 cuts. Those plans were reportedly shelved after AI-linked software bugs and slower-than-expected agent development gave leadership pause. CEO Mark Zuckerberg has said he does not expect further mass layoffs this year. Even so, employees worry that if Hatch does drive productivity gains, headcount reductions could follow.

For businesses watching how large companies deploy agentic AI internally, this is a useful case study. Agentic tools that can take actions autonomously, rather than just answer questions, consume far more compute than a standard chatbot. Token volumes are already rising at Meta even without the old incentive to inflate them artificially. That has real cost and environmental implications, especially as these tools move toward public release. If you are evaluating similar tools for your own team, our notes on AI integration for businesses cover what to assess before connecting an autonomous agent to live accounts.

Our take

Meta’s token-counting regime was always going to end badly. You do not measure productivity by how many words someone throws at a language model. The fact that it took a lawsuit and a viral internal leaderboard to prompt a course correction is not a great look, but the correction itself is the right call.

Hatch is the more interesting story. Agents that can take actions on a computer, browse the web, and interact with third-party apps are genuinely different from a chat interface. The privacy concern employees raised is not paranoia. Connecting an employer-controlled AI agent to your personal calendar or email is a meaningful data decision, and Meta’s keystroke-tracking history gives employees reasonable grounds for caution.

For business operators considering agentic tools, the Meta situation illustrates a pattern worth watching across the wider AI news landscape: early adoption driven by internal pressure, followed by trust erosion when the tool does something unexpected or the data practices become visible. Rolling out slowly, with clear data boundaries and voluntary participation, would have been cheaper than the trust damage Meta is now trying to repair.

What to do about it

1. Audit any internal AI metrics you are using now. If you are measuring adoption by volume (prompts sent, tokens consumed), replace that with outcome-based measures before it distorts behavior.
2. Before deploying an agentic tool to your team, document exactly which accounts and data sources it can access, and make that list visible to everyone who will use it.
3. Run a voluntary pilot first. Employees who opt in generate better feedback and more trust than those who feel pressured.
4. Watch Hatch’s public release. If Meta ships it externally, the feature set and data permissions will be public, making it easier to evaluate against other computer-use agents on the market.

The takeaway: measure what AI actually produces, not how much of it employees consume.

Source: WIRED · AI