Meta is paying its own workers to be the training set

Reuters reported on Sunday that Meta has begun installing telemetry software, the Model Capability Initiative, or MCI, on US employees' laptops, capturing mouse movements, clicks, keystrokes, and periodic screenshots, with the explicit purpose of generating training data for internal AI agents.¹ The MCI sits inside a larger programme Meta calls the Agent Transformation Accelerator. EU staff are excluded, which Reuters attributes to workplace-monitoring law in several member states. The internal reaction was, per the reporting, immediate and unhappy.

I want to take this seriously as a piece of market structure rather than as a privacy story, because the privacy story will be covered extensively elsewhere and the market-structure story is, I think, more interesting and less obvious.

Start with what MCI actually is. It is not productivity monitoring in the usual sense, Meta is not measuring whether its employees are working hard enough. It is instrumentation to capture the trajectory of knowledge work at keystroke-and-pixel resolution: what application is open, what is on the screen, where the cursor goes, what is typed, what is clicked next. This is, precisely, the data format that agent training wants. Large language models trained on text learn to produce text. Agents trained to operate computers need trajectories, sequences of (screen state, action) pairs, and those trajectories are, empirically, the bottleneck. Anthropic's computer-use model, OpenAI's Operator, Google's Project Mariner: each has shipped against a wall of limited real-world agent trajectory data. Synthetic data helps. Public video helps a little. What you actually want is millions of hours of competent knowledge workers doing their actual jobs in their actual software, annotated with intent.

It is not productivity monitoring in the usual sense, Meta is not measuring whether its employees are working hard enough.

Meta employs roughly 75,000 people. If meaningful fractions of the US cohort run MCI through a working year, that is on the order of 10⁸ hours of annotated, in-domain, English-language, knowledge-work trajectory data. I am not aware of any other party in the market with a plausible path to a comparable corpus. Scale AI, which Meta paid $14bn for a controlling stake in last year, can source human labellers to produce trajectories, but those are synthetic tasks performed by contractors, not real work performed by the workers you are trying to replace. The MCI corpus is, in training terms, distributionally matched to the target deployment. That is the thing that is hard to buy.

So this is, first, an inference-economics move dressed as an HR policy. The binding constraint on agent deployment is not frontier model capability; it is the cost and reliability of getting an agent to complete a multi-step workflow without a human intervening. Reliability comes from training on competent trajectories. Meta has concluded, correctly, I think, that the cheapest route to those trajectories runs through its own payroll.

Second, it is an agent-economics move with a very specific theory of the firm attached. The standard agent-deployment pitch to enterprise is: buy our agent, point it at your workflows, save on headcount. The problem with this pitch is that the agent doesn't know your workflows, and the customer won't give the vendor screen-recording access to learn them, because that would be insane. Meta's solution is to train the agent on Meta workflows first, ad ops, recruiting, engineering review, legal intake, and then, presumably, deploy it against Meta headcount. The Agent Transformation Accelerator name is doing some work here: the programme is explicitly about transforming agents and, the unstated half, the headcount the agents are trained against.

This is where the FDE market-structure frame becomes useful. The prevailing go-to-market story for enterprise agents has been embedded-engineer-led, labs send forward-deployed engineers into customer environments to instrument workflows and fine-tune models. It is expensive, slow, and doesn't scale. What Meta is doing is the inverse: instrument the workflows first, at its own expense, inside its own firm, and use the resulting models either internally or as products. It is vertically integrating the FDE function by making the customer and the training lab the same entity. If the models that come out are good, Meta has a route to agent products no external vendor can match, because no external vendor will ever get that data. If the models are not good, Meta has at minimum cut its own internal labour cost, which is, and this is the part that is slightly amusing, being financed by the labour of the workers being cut.²

Three things to note about the filing, or rather the lack of one. There is no 8-K, no formal disclosure, no mention in the most recent 10-Q of MCI as a specific initiative. The Agent Transformation Accelerator appears in internal documents Reuters has seen but not in investor materials. This is consistent with Meta's practice of discussing AI capex and headcount reductions at high aggregation levels during earnings calls, and not breaking out the mechanism. I would watch the next 10-Q for any new language around "productivity initiatives" or "internal tooling" in the MD&A, and any adjustment to the stock-based-compensation footnote, because a programme that quietly reduces headcount against a fixed RSU pool changes per-employee dilution meaningfully.

The EU carve-out is the part of this that I find most structurally revealing. Meta's position is that MCI cannot run on EU employees because of workplace monitoring law, which is true as far as it goes, though the relevant regimes (German co-determination, French CNIL guidance, the GDPR Article 88 framework) permit monitoring with works-council agreement and purpose limitation. Meta is not seeking those agreements. The read-through is that the company has decided its agent-training corpus will be a US-workforce corpus, which means its agents will be trained to do knowledge work the way Americans at Meta do knowledge work. For a company whose advertising revenue is globally distributed and whose agent products will presumably be sold globally, this is a choice with downstream consequences that I don't think have been fully thought through. Agents trained on Menlo Park workflow patterns deployed into Munich ad agencies is a specific sort of bet.

What to watch. First, whether other large-cap tech employers follow, Amazon, Google, and Microsoft all have the same training-data problem and the same US-EU split. If one of them announces a comparable programme in the next two quarters, MCI becomes a category rather than a Meta idiosyncrasy. Second, whether Scale AI's revenue mix shifts; if Meta-internal trajectory data partially substitutes for contracted labelling, Scale's largest customer becomes a smaller one, and the $14bn investment starts to look less like a data-supply deal and more like a talent acquisition. Third, the first wrongful-termination or NLRB filing, because the legal theory under which knowledge-worker telemetry is used to train the agent that replaces the worker is genuinely novel and will need to be litigated. Fourth, whether MCI extends to contractors, because if it does, the data corpus roughly triples.

I'd also watch for the internal memo. There is always a memo.

Footnotes

¶