Can employers use AI usage metrics? Meta backs off

Meta told engineers this week it will stop using AI‑adoption dashboards and raw token counts as part of performance evaluations while the company continues internal testing of a new agentic tool called Hatch — a move that raises the question HR leaders have been asking: can employers use AI usage metrics?

On June 12, 2024, The Information reported that Meta instructed engineering managers to remove AI‑usage and token metrics from performance reviews and promotion discussions. The company remains encouraging staff to try Hatch, an internal assistant designed to automate coding and research tasks, but says those experiments will not be measured back into formal review scores, at least for now.

Meta framed the decision as a response to employee feedback and the need to balance productivity measurement with privacy and legal exposure. Company spokespeople told Wired that while internal pilots for Hatch and other AI tools will continue, the organisation no longer wants engineers to be evaluated by dashboards that track the number of API tokens consumed or similar unitised measures of AI activity.

Engineers and privacy advocates have criticised token‑counting and adoption dashboards as crude proxies for value: they reward ‘‘token‑maxxing’’ and can create perverse incentives to prompt more rather than better output. The backlash has been amplified by legal scrutiny. Regulators and lawyers are increasingly scrutinising employer monitoring and automated metrics for potential privacy violations and disparate impact on protected groups, and internal records showing employees’ AI interactions could become fodder in discrimination or data‑access disputes.

Meta’s move is the latest sign that large tech employers are recalibrating how they measure AI‑driven productivity. Other companies have already limited use of monitoring data in HR processes or dialled back plans to bake tool‑usage into promotions after staff pushback. For HR leaders, the episode underlines a practical tension: the desire to quantify gains from generative tools collides with measurement risk — both from a human‑resources fairness perspective and from a legal and privacy standpoint.

What Meta did not disclose is how it will evaluate AI experimentation in non‑formal ways, or whether the pause on token metrics is temporary. The company has not published any independent bias or privacy assessments of Hatch, nor has it indicated what internal governance will oversee the agent’s rollout. Details are scant on whether metrics will still inform team‑level resourcing decisions, informal manager judgments, or performance improvement plans. Meta also hasn’t shared whether employees can opt out of using internal agents or how data generated by Hatch will be retained, accessed or used in investigations.

For HR and employment lawyers, the change is both a warning and a lesson. Companies that rush to measure tool adoption with simple quantitative proxies risk creating incentives that distort work and expose them to litigation or regulatory complaints. At the same time, employers need workable methods to understand whether AI tools actually improve outcomes, not just increase token volumes.

How Meta proceeds with Hatch will matter beyond its own walls. If the company builds robust governance — clear opt‑outs, documented retention policies, and independent audits of the agent’s decisions — it could set a template for firms wanting to experiment without weaponising usage telemetry in reviews. If it does not, the episode may accelerate broader moves among employers to separate product experimentation from personnel evaluation, reshaping how HR teams define and measure productivity in an AI‑assisted workplace.