Meta just made one of the boldest moves in the AI agent race. The company is now installing software on US-based employees’ computers that records their mouse movements, clicks, keystrokes, and screenshots. All of it feeds directly into training Meta’s AI agents to use computers the way humans do.
If you run a service business, pay attention. This isn’t some research paper or product teaser. It’s a massive company turning its entire workforce into a training dataset for AI agents that automate real work tasks. The implications for every industry are hard to overstate.
What happened
- Meta launched an internal tool called Model Capability Initiative (MCI) that runs on US employees’ work computers.
- MCI records mouse movements, clicks, keystrokes, and occasional screenshots inside work-related apps and websites.
- The collected data trains Meta’s AI models to interact with computers the way humans do, including automating the kinds of tasks employees perform daily.
- Meta says the data won’t be used for performance assessments and has safeguards to protect sensitive content.
- Meta CTO Andrew Bosworth announced the initiative as part of Meta’s Agent Transformation Accelerator (ATA) in an internal memo on Monday.
- Bosworth’s stated vision: agents primarily do the work, and humans direct, review, and help them improve.
The numbers
- Meta has roughly 70,000+ employees, making this one of the largest known employee-activity-to-AI-training pipelines in the world.
- The tool is currently deployed to US-based employees only.
5 things every service business operator should take from this
- The training data bottleneck for AI agents is real. Meta, one of the richest tech companies on Earth, can’t just simulate how humans use computers. It needs actual human behavior. That tells you how far we are from agents that “just work” out of the box. If you want agents that handle your workflows, someone needs to collect examples of how those workflows actually happen.
- Employee consent and trust are the new infrastructure. Meta says MCI data won’t be used for performance reviews and that safeguards protect sensitive content. Whether employees believe that is another story. Any business exploring similar data collection needs to lead with transparency or risk a revolt.
- “Agents do the work, humans direct” is the operating model Meta is betting on. Bosworth’s memo lays it out plainly. The end state isn’t AI helping you do your job faster. It’s AI doing your job while you supervise. That’s a fundamentally different org chart.
- Your competitors will copy this playbook. If Meta proves that recording employee workflows produces better AI agents, every large enterprise will follow. The question is whether smaller service businesses can get ahead of this curve without the same resources.
- Compliance is going to get complicated fast. Recording keystrokes and screenshots in a workplace touches labor law, privacy regulations, and in some states, wiretapping statutes. The legal landscape around this kind of data collection is still being sorted out. If you move here, move carefully.
The hot take
Meta is right to do this, and most companies criticizing the move will quietly do the same thing within 18 months. Here’s the reality: you cannot build AI agents that automate computer-based work without watching how humans actually do that work. Synthetic data and simulations only get you so far. The companies that figure out ethical, transparent ways to collect real workflow data from their teams will build dramatically better AI agents than those that don’t. The squeamishness around employee monitoring is understandable, but the competitive pressure is going to win. The only real question is whether companies do it openly (like Meta is doing) or quietly in the background. Open is better.
The Agency OS play
You don’t need to be Meta to start building this muscle. This week, pick one repetitive computer-based workflow in your business. Something like processing invoices, updating a CRM after client calls, or formatting reports. Ask the person who does it to record their screen (with Loom or any free screen recorder) for a full day of doing that task. You now have training data. Not Meta-scale training data, but enough to start prototyping an AI agent or automation that handles pieces of that workflow.
More importantly, start the conversation with your team about data collection and consent now, before you need to. Write a simple one-page policy that says what you’ll record, why, and what you won’t use it for. Make it clear that workflow recordings aren’t performance reviews. If you run a law firm, a clinic, or a financial advisory practice, you’ll also want to check with your compliance counsel on state-specific rules around employee monitoring and any industry-specific data handling requirements.
The bigger move: start cataloging every task in your business that involves a human clicking through software in a repeatable pattern. Those are your AI agent candidates. Rank them by hours spent per week and error rate. The top three on that list are where you should focus first. Meta is spending billions to solve this problem at scale. You can solve it for your specific workflows with a fraction of the effort, because you only need agents that work for your business, not everyone’s.