Meta Quietly Retires Its AI Usage Scoreboard, Then Rolls Out a Bigger AI Habit

Meta has quietly ended a policy it spent much of the year building. Engineers were told this week that "AI adoption dashboards" and "token counts" will no longer factor into performance evaluations, according to The Information. Wired, citing anonymous sources inside the company, reports that the new review guidance drops "usage of AI" as a criterion entirely. The replacement language reportedly asks for outcomes that "can be supported by AI or other means." Several of Wired's sources described the change as a relief.

That relief says something about how the previous system felt from the inside.

The backstory runs through a Kevin Roose column in the New York Times back in March. Roose described AI-enthusiastic employers, Meta and OpenAI among them, pushing staff to run up their AI usage. The behavior earned a name: tokenmaxxing. Roose pointed to OpenClaw, then a popular token-hungry agentic platform, as a likely spark. His reporting also described leaderboards at Meta, essentially gamifying AI consumption with a scoreboard attached. Workers who used AI heavily were rewarded. Those who didn't were chastened, per his sourcing.

The incentive structure ran into a supply problem in June. The Financial Times reported that Google was throttling Meta's AI tokens, a constraint that forced some internal rethinking of the tokenmaxxing push. The tokens, it turned out, were not unlimited. A company cannot indefinitely reward employees for consuming a resource that a supplier is actively restricting.

The replacement tool consumes more, not less

While the leaderboard incentives were unwinding, Meta was building something to take their place. In 2024, Mark Zuckerberg predicted that "a highly intelligent and personalized AI assistant" would reach more than a billion people that year, and that Meta AI would be it. Whether the company's ChatGPT-style app ever approached that scale is unclear. One data point suggests it hasn't landed well with the audience Meta actually has. TechCrunch's Amanda Siberling tried the Meta AI web app for a story and found that Instagram alerted her friends she was using it. They mocked her for it.

Meta appears to have drawn a lesson from that experience. Rather than fight for space in the consumer chatbot market, the company is betting on a productivity tool. The Information reported in May that Meta was building Hatch, its own answer to OpenClaw, a consumer-facing version of the same token-hungry agentic category that produced tokenmaxxing in the first place.

Wired reports that Meta "has been encouraging, but not requiring, employees to use its most advanced AI experiment." That experiment is Hatch. There is still no public release. Staff have reportedly had internal access "for the past several weeks." The tool is said to consume more tokens than conventional chatbots and AI coding assistants, not fewer.

So the scoreboard came down. The equipment got bigger.

Reaction among staff is split. Some are apparently enthusiastic about Hatch. Others, per Wired's sourcing, are keeping their distance, citing an April news cycle in which it surfaced that Meta intended to train AI on employee keystrokes. That program was later paused.

Agentic AI remains embedded in Meta's operating culture regardless of what the review rubric now says. AI chief Alexandr Wang circulated a memo this week explaining that the company's migration from Google Chat to Slack was designed partly to make agent use easier. Wang called Slack "the strongest platform available today for agents." He acknowledged that "not all of us are building agents today," but added that "everyone at the company will benefit from a robust and useful agent ecosystem."

The message is consistent even if the incentive mechanism changed. Individual usage quotas are gone. The institutional push toward agents is not.

What actually worries staff, according to Wired, is not Hatch itself but what a successful rollout might trigger. Reuters reported last week that Meta had drawn up plans to cut as much as 60% of its workforce in certain areas. Those plans stalled, per the report, because the AI software meant to absorb departing employees' work was too buggy to deploy.

That detail deserves more attention than it has gotten. The cuts were not shelved on principle. They were shelved because the replacement technology wasn't ready yet. That is a timing problem, not a policy reversal, and timing problems get solved.

Meta did not respond to questions about the Reuters report or about internal testing of Hatch.

The Bottom Line

Removing token counts from performance reviews reads less like a retreat from AI and more like a maturing of the metric. Individual usage quotas were always a blunt instrument, easy to game and vulnerable to supply constraints, as the Google throttling episode demonstrated. Hatch represents a shift from measuring adoption to embedding it structurally, through infrastructure choices like the Slack migration and through a tool designed to consume more resources than what came before it. For investors watching Meta's AI capital allocation, the relevant signal isn't the review policy change itself. It's the Reuters detail about workforce cuts stalling on buggy software. That suggests Meta's AI substitution timeline is running behind its ambition, and the gap between the two is where both cost pressure and execution risk will show up first.