Meta is learning a hard lesson about AI agents: the most valuable training data may also be the most sensitive data a company has.
Meta’s plan to train AI agents on the daily computer activity of its own employees has moved from an internal productivity experiment into a privacy fight. The company’s Model Capability Initiative, known as MCI, was first described as a way to capture mouse movements, clicks, keystrokes and occasional screenshots from U.S. employee computers. The newer concern is that the system appears to reach deeper into the working day than that simple description suggested.
According to Reuters reporting based on internal Meta materials, MCI can collect interaction data across more than 200 apps and websites, including email, chat and browsing activity when U.S.-based employees use monitored work systems. That matters because the line between learning how a person uses software and collecting the substance of their work is not a technical detail. It is the difference between observing workflow and absorbing workplace communications.
Meta says the point is to give its AI agents real examples of how people complete everyday tasks on computers. That is a reasonable engineering problem. AI models can write emails, summarize documents and generate code, but agents still struggle with the messy sequence of ordinary work: moving between tools, choosing the right menu, pasting information into the right place, and correcting small mistakes without human hand-holding.
The problem is how Meta is trying to solve it. A model that needs real office behavior cannot get everything it needs from public web pages or synthetic data. It needs the context of actual work. Yet actual work includes confidential discussions, HR questions, legal topics, customer information, strategy documents and personal messages that happen to pass through company systems. Once that data becomes training material, employees are right to ask who can inspect it, how long it is stored, and whether it can be separated cleanly from the private substance of their jobs.
Keystroke logging already carries a heavy surveillance history. Companies have long used monitoring tools to check productivity, detect misconduct or secure sensitive systems. Meta’s argument is different: this is not for employee performance reviews, but for model training. That distinction may be important inside the company, but it will not calm everyone outside it.
The European angle is especially sensitive. Internal documents cited by Reuters indicate that messages involving non-U.S. employees may be captured when they communicate with U.S. colleagues whose machines have the tool enabled. Meta has said it notified non-U.S. employees that this could happen in the normal course of communicating with U.S. colleagues. That may be disclosure, but disclosure is not the same as comfort, particularly in a region shaped by GDPR and years of fights over how Meta handles personal data.
There is also a labor issue hiding inside the AI issue. Workers are being asked, directly or indirectly, to help train systems that could automate portions of their own jobs. Even if the near-term use is productivity support, the direction of travel is clear across the AI industry. Companies want agents that can operate software, complete business processes and reduce the need for humans to carry out repetitive digital tasks.
Meta is already adjusting the rollout
The backlash has had an effect. Reuters reported on June 2 that Meta is scaling back parts of the program after weeks of employee pushback. An internal memo from Stephane Kasriel, a vice president in Meta’s AI model-building Superintelligence Labs unit, said workers would get new controls, including the ability to pause collection for up to 30 minutes and request exemptions.
That is not a full retreat. It is a sign that even Meta, a company used to defending aggressive data practices, understands that workplace AI training can create morale and legal problems if employees feel they have no practical choice. A pause button is useful, but it also proves the underlying point: people want moments at work when the machine is not learning from them.
For other companies building AI agents, Meta’s experience is an early warning. The race is no longer just about buying GPUs, licensing models or launching chatbots. The next contest is over proprietary behavioral data, the kind of information that shows how real work actually gets done. Large employers have that data sitting inside their own systems, but turning it into training material will be harder than simply writing a new internal policy.
Enterprise buyers should watch this closely. If AI agents are trained on workplace activity, customers will want to know whether their own communications, support tickets, documents or employee actions could become part of a vendor’s improvement loop. The answer will shape procurement, compliance reviews and trust in agentic software.
Meta may still prove that this kind of training makes AI agents more useful. But usefulness alone will not settle the matter. The companies that win the next phase of AI will need better models, and they will also need a convincing answer to a simple question: when work becomes training data, who controls what the model gets to learn?
Also read: Meta is learning how employees work by watching their screens, Colorado keeps AI pricing rules on hold after Polis vetoes HB26-1210 and Meta scales back its employee tracking plan after AI data backlash
