Meta is installing tracking software on US-based employees' computers that will record their mouse movements, clicks, keystrokes, and occasional screenshots, then feed that activity into its AI training pipeline. The program, disclosed this week through internal memos obtained by Reuters, is called the Model Capability Initiative (MCI), and it represents one of the most direct attempts yet by a major tech company to turn its own workforce into training data for AI agents meant to replace white-collar tasks.
The memos describing MCI were circulated inside Meta Superintelligence Labs. According to the reporting, the software will only run on a designated list of work apps and websites rather than everything on an employee's machine, and Meta has told staff the data will not be used for performance reviews or any purpose beyond model training. A company spokesperson justified the collection in blunt terms: "If we're building agents to help people complete everyday tasks using computers, our models need real examples of how people actually use them."
From 'AI for Work' to 'Agent Transformation Accelerator'
MCI is being folded into a broader internal effort that Meta CTO Andrew Bosworth has rebranded from "AI for Work" to the Agent Transformation Accelerator, or ATA. In a separate memo to employees, Bosworth laid out the end state he is aiming for: "The vision we are building towards is one where our agents primarily do the work and our role is to direct, review and help them improve."
That framing matters. It confirms what AI leaders have mostly hinted at in public — that the point of agent research is not copilot-style assistance but substitution of routine knowledge work. Meta is explicit that the data it needs to close the gap is the kind models cannot currently learn well from the open web: the small, fiddly motions of real workers navigating dropdowns, keyboard shortcuts, form fields, and internal tools.
Why this story is different from old-school workplace surveillance
Employee monitoring software has existed for years, and gig-economy platforms have long tracked keystrokes and mouse movement for productivity scoring. What is new here is the purpose. MCI is not framed as a productivity-measurement tool; it is a data-harvesting pipeline whose explicit goal is to train models that can eventually do the monitored employees' jobs.
That reframing sits on top of an already fraught US regulatory picture. There is no federal cap on how much employers can surveil workers on company equipment, and most state protections focus on consent and notice rather than limiting use of the data. Meta can plausibly argue MCI is lawful under current rules — but the optics of training a replacement on the person you're recording are sharper than standard productivity analytics.
What to watch next
Three things are worth tracking. First, whether other hyperscalers follow Meta's lead; capturing knowledge-work trajectories is arguably the single most valuable untapped training corpus in enterprise AI. Second, how Meta's own employees respond, particularly in Superintelligence Labs where the program lives. And third, whether regulators in Europe, where employee-monitoring rules are significantly stricter, move to block any extension of MCI-style collection to non-US staff.
For now, MCI is a quiet but important pivot in how frontier AI companies source data for agents — one that trades scraped internet text for the recorded labor of their own workforce.
