Meta Platforms has started deploying internal tracking software on devices used by its U.S.-based employees, collecting data such as mouse movements, keystrokes, and periodic screenshots. The company is using that information to train artificial intelligence models aimed at understanding how people actually work on computers.
The internal system is called Model Ability Initiative, or MCI. It runs across work-related applications and websites, according to internal materials cited in the report. Company documents say the data is meant to improve how AI systems interpret everyday computer use, including keyboard shortcuts, menu navigation, and movement between different application states. One internal note described the process in simple terms: employees help improve the models by doing their regular jobs.
Data collection expands under Agent Transformation Accelerator
The rollout sits inside a broader shift in Meta’s internal structure as the company pushes AI deeper into routine workplace functions. Employees are being encouraged to use AI tools more heavily for tasks such as coding, even where that may reduce productivity in the short term. Meta has also introduced a broader “AI builder” role and restructured parts of its engineering organization around AI-led development.
To support that effort, the company has created a dedicated Applied AI engineering group focused on building systems that can write and deploy software. Some engineers have already been reassigned. In a separate internal message, Meta Chief Technology Officer Andrew Bosworth said the company is increasing data collection under its “AI for Work” initiative, which has now been renamed the Agent Transformation Accelerator.
Bosworth said the company is working toward a model in which agents handle most of the work, while employees direct them, review their output, and help improve them. He added that the systems are being designed to detect moments where human input is needed so they can get better over time. He did not provide technical detail on the training pipeline, but said workplace interaction data remains central to the effort.
Meta says MCI data is for training, not performance reviews
Meta spokesperson Andy Stone said information collected through MCI is used only for model training and not for evaluating employee performance. He said that if the company wants to build agents that can help people complete everyday computer tasks, the models need real examples of how humans click buttons, move through menus, and operate software in practice.
Meta also said safeguards are in place to avoid collecting sensitive data. It did not specify what categories of information are excluded. The company outlined the purpose of the tool, but left the boundaries of the filtering process largely undescribed.
AI Zuckerberg project and social commerce push move in parallel
Alongside the workplace automation drive, Meta is also testing an AI-powered internal communication system built around a photorealistic digital version of Chief Executive Officer Mark Zuckerberg. The system is being trained on Zuckerberg’s voice, tone, and communication style so it can simulate natural conversations and deliver responses aligned with his views.
The project is still in an early stage, yet it points to Meta’s continued investment in virtual human systems that can speak and respond in real time across different settings. At the same time, the company is tightening links between creators, AI tools, advertising, and shopping activity on platforms including Instagram and Reels. According to the report, the update is intended to make it easier for users to move from discovering content to completing purchases within the same experience.