There’s a particular kind of AI moat that cannot be bought, rented or scraped from the open web: It is the granular record of how humans actually do their jobs. The hesitation before a dropdown menu, the specific backspace pattern when someone abandons a sentence, the order in which tabs get opened. This week, Meta became the first company at scale to start manufacturing it.
The Dropdown Menu Problem
Meta disclosed the ‘Model Capability Initiative’, or MCI program, in a memo to a staff channel belonging to Meta Superintelligence Labs — not Human Resources, which is the first signal about what the program actually is.
On a designated list of work apps and websites, MCI records mouse movements, click locations, keystrokes and intermittent screenshots from U.S. employees’ work computers. The memo’s stated goal is not performance management; it is to help Meta’s AI models learn computer-use behaviors they currently struggle with — specifically, navigating dropdown menus and using keyboard shortcuts. Meta says employee activity on company machines has been monitored for years and that safeguards protect sensitive content.
Meta is spending between $115 billion and $135 billion on AI capital expenditure in 2026. But the piece it says it needs — the input compute alone cannot produce — is a corpus of humans using dropdowns and Ctrl-F.
The Only Hyperscaler Who Can Pay The Price
Every company racing to ship AI agents needs workflow data — the granular record of how humans actually operate inside enterprise software. The open web doesn’t contain it. The obvious workaround is synthetic generation, but a model can’t generate a convincing hesitation before a dropdown menu. Enterprise customers generate it constantly, but using it for model training risks breaching contracts and eroding trust.
Microsoft, Salesforce, Oracle and ServiceNow sit on top of an ocean of this data inside their customers’ environments. They won’t touch it — the same trust that makes enterprise software a multi-hundred-billion-dollar business also constrains how that data can be used. It won’t take more than one leaked memo suggesting otherwise to reprice every contract in the industry.
Google and Apple face a different constraint: ongoing antitrust pressure and a user base already sensitive to how personal data is handled. OpenAI and Anthropic don’t have workforces large enough to generate a usable corpus.
That leaves Meta. Tens of thousands of U.S. knowledge workers on company-controlled devices. A business with no enterprise customers to lose, and a willingness to absorb the reputation cost. The MCI is the workforce moat in practice.
A Program That Exists Because Of American Law
The MCI works the way it works because of a specific accident of American labor law. U.S. federal statutes place no limit on monitoring employees on company devices. European employees fall under a different regime — the General Data Protection Regulation would likely prohibit equivalent monitoring across the EU. Meta’s program is, by design, restricted to the United States.
The companies that will produce the training data for the next generation of enterprise AI agents are companies that can legally instrument their own workers at scale. Which means the center of gravity for workflow-data production is forming, quietly, in exactly one jurisdiction. Every hyperscaler with U.S. headcount is now doing the same math Meta did — and the ones without U.S. headcount are doing a different, more expensive one.
The Boring Moat Forms Again
The companies training AI agents for real-world work need the record of real-world work, and that record exists in exactly two places: inside enterprise customers’ systems, where nobody can touch it, and inside employees, where Meta just showed up with recording software.
Every hyperscaler racing to ship agents will need this data. Most of them can’t generate it themselves, which means a market will form around it — through acquisition, licensing, synthetic-data partnership or the less-examined path of copying the MCI playbook in their own U.S. offices. Not every company will be able or willing to absorb the legal, reputational and workforce cost of doing so, and those that can’t will end up buying from those that can.
The reputation cost is not the barrier it looks like. It is the moat. Compute was the 2022 moat. Models were the 2024 moat. Workflow data is shaping up as the 2026 moat — and at the moment there’s exactly one company at this scale politically willing to manufacture it in-house.
