Meta will mine employee computer activity to train its AI agents

Meta is installing an internal tool called the Model Capability Initiative, or MCI, on U.S.-based employees’ work computers. According to Reuters’ reporting echoed by The Verge, Fortune and the BBC, it runs inside selected work apps and websites and captures mouse movement, clicks, keystrokes and occasional screenshots ^[1][2][7]. Meta says the data is meant to teach AI agents how humans actually navigate software, and the company says it has safeguards to protect sensitive content ^[1][7].

The important detail is not just that Meta is observing work activity, but that it is doing so with a specific training purpose. Andrew Bosworth’s internal memo framed the broader Agent Transformation Accelerator effort as a future in which “our agents primarily do the work and our role is to direct, review and help them improve” ^[1]. That makes MCI less like ordinary workplace analytics and more like a live data pipeline for agent training ^[1][2].

The rationale for MCI is a gap that text-heavy AI training cannot fill. Computer-use models need to learn procedural behavior—how a person searches a menu, uses a shortcut, drags a file, or recovers from a mistake—not just what words mean ^[8]. Periodic screenshots add the visual context that lets an input event be paired with the exact interface state it occurred in, turning isolated clicks into reusable demonstration trajectories ^[8].

That is why the industry is moving toward behavioral traces instead of only documents and code. Fortune and other outlets noted that Meta is pursuing this data while racing OpenAI and Anthropic to ship more capable work agents, and TechTimes described the approach as part of a broader shift toward real-world user behavior as training fuel ^[2][5]. In that sense, MCI is a workaround for the data bottleneck that has limited computer-use AI since its earliest releases ^[8][9].

The employee reaction reported across outlets was unusually blunt. One worker asked, “How do we opt out?” and Bosworth’s answer, according to The Verge’s reporting, was that there is no opt-out on a work-provided laptop ^[1]. The BBC likewise quoted employees calling the practice “very dystopian” and describing it as the latest sign that AI has become all-consuming inside the company ^[7].

Meta is also asking employees to accept this change at a moment when the company is already cutting jobs and shrinking its hiring pipeline. Reuters-based coverage said Meta has discussed possible layoffs of up to 20%, while the BBC reported roughly 2,000 cuts already this year and a dramatic drop in job postings, from about 800 listings in March to just seven later on ^[2][7]. Even if Meta says MCI data will not be used for performance reviews, the timing makes the tool feel less like neutral research and more like part of a broader restructuring ^[1][7].

Meta is not alone in hunting for better training signals. Fortune reported that OpenAI has used third-party contractors to gather real work products like PowerPoints and spreadsheets, while Meta itself has poured billions into Scale AI and brought Alexandr Wang into the Meta Superintelligence Labs orbit ^[2]. The message across these efforts is consistent: frontier labs are running out of easy public data and are now looking for proprietary workplace behavior to improve agent quality ^[2][8].

The spending context matters because MCI is not a side project. The BBC reported that Mark Zuckerberg is planning roughly $140 billion in AI spending in 2026, almost double the prior year, while Reuters-based reporting tied MCI to Meta’s Agent Transformation Accelerator and the company’s effort to make agents proficient enough to replace routine computer work ^[2][7]. In other words, Meta is not merely studying how employees work; it is trying to compress that behavior into a product category it can sell and deploy at scale ^[1][2].

This story lands at the intersection of AI product development and workplace governance. If the best training data for agents comes from ordinary office behavior, then every enterprise experimenting with computer-use tools will need to decide how much telemetry it is willing to collect, how long to keep it and whether employees can refuse ^[1][7]. The shift also blurs the line between operational monitoring and model training, which creates new risks even when a company says performance management is off-limits.

For AI practitioners, the takeaway is that agent quality may increasingly depend on data policies, not just model architecture. Organizations that want to build or deploy similar systems will need sharper controls around screenshots, credentials, confidential information and downstream reuse, because the raw inputs are often indistinguishable from sensitive work product ^[5][8]. As agents move from demo to workflow, the winners may be the firms that can collect useful behavior without destroying trust in the process ^[1][9].