Let’s be honest about what happened at Meta this week.
On April 21, a Reuters exclusive landed: Meta had installed surveillance software on employee computers. Every mouse movement. Every click. Every keystroke. Random screenshots throughout the day. Hundreds of apps and websites being tracked.
On April 23, Meta announced it was laying off 8,000 people. Ten percent of the workforce. Gone by May 20.
Two days apart.
That’s the story. Not individually — together.
What the Surveillance Is Actually For
The tool is called MCI. Model Capability Initiative. It was deployed under a broader program originally called “AI for Work,” which was then rebranded to something more clinical: the Agent Transformation Accelerator, or ATA. The rebrand happened quietly. The data collection didn’t stop.
Meta’s internal memo described the purpose with unusual candor: the tool will “help our models get better simply by doing their daily work.”
Read that again slowly.
The models are getting better. By watching the employees do their daily work. The employees who are, as of April 23, 8,000 fewer in number.
Here’s what you need to understand about what MCI is actually capturing and why. Meta’s AI agents — the ones being built in Meta Superintelligence Labs, the ones Zuckerberg says will eventually do the work of an entire workforce — currently can’t perform basic software tasks. They can’t navigate dropdown menus reliably. They don’t know keyboard shortcuts. They can’t move through a Salesforce workflow or a GitHub repository the way a trained human can.
That’s the gap. And MCI is designed to close it.
The platform list CNBC obtained is worth sitting with for a moment. MCI tracked activity across: Google, LinkedIn, Wikipedia, GitHub, Slack, Salesforce, Atlassian, Threads — and originally ChatGPT and Claude, both of which were removed from the list after internal backlash. Think about what that list represents. It’s not random. It’s every tool category that knowledge workers use to do their actual jobs: communication, code, documents, search, project management, customer relationship management. The agents can’t navigate any of these natively yet. MCI is the curriculum.
Meta CTO Andrew Bosworth expanded the data collection scope when the program was rebranded to ATA. Bosworth’s memo made clear this isn’t a limited experiment. It’s a company-wide systematic extraction of how people do their work — the tacit knowledge that experienced employees accumulate over years, the kind that doesn’t appear in job descriptions or training manuals, the kind that usually walks out the door when people leave.
Employees are demonstrating the work. The agents are watching. Learning. Getting better. The humans are, quite literally, teaching their successors by doing their jobs.
“Every Employee Gets an AI Agent” — What Zuckerberg Actually Said
Zuckerberg has a vision for where this goes. He’s spoken about it publicly: every employee at Meta will eventually have their own AI agent. That’s the direction of travel. That’s where the $115–135 billion in capital investment this year is pointed.
But here’s the critical thing to understand: that’s a vision, not a current reality.
Right now, today, Meta employees do not have their own assigned AI agents. What they do have is MCI running in the background — capturing how they work, how they navigate software, what workflows they follow. They are being observed and logged, with no opt-out, graded on their AI adoption rates, but not yet equipped with the agents they’re supposedly being prepared for.
The sequence matters. First comes the data extraction. Then comes the agent deployment. The surveillance is Stage One. The agents are Stage Two. The layoffs are happening in parallel with Stage One, before Stage Two is fully realized.
If you want a timeline:
– Q4 2025: Meta sets a companywide goal — 55% of engineer code changes must be “agent-assisted” – Early 2026: Meta Reality Labs cuts 1,500 employees, shifts resources to Superintelligence Labs – March 23, 2026: Reports emerge that Zuckerberg is building his own “CEO agent” – ~April 13, 2026: A Zuckerberg AI chatbot is deployed for all employees – April 21, 2026: Reuters reveals MCI. Employee panic erupts in internal forums. – April 23, 2026: Meta announces 8,000 layoffs, effective May 20. – May 20, 2026: Layoffs take effect. MCI data collection continues.
The employees leaving in May will have spent weeks after the announcement still doing their jobs — still clicking, still navigating, still demonstrating the workflows — while MCI captures everything. The data collection doesn’t stop when the pink slips go out. Why would it?
What It Feels Like Inside the Building
When Reuters published the MCI story on April 21, Meta’s internal Slack boards erupted, per CNBC. Employees drew the connection themselves: their workflows had been under surveillance for weeks, and the layoff announcement came 48 hours later with no explanation from leadership of the link between the two.
This Is Taylorism With Better Hardware
Frederick Winslow Taylor had a clipboard. He stood on factory floors in the late 19th century and watched workers do their jobs. How long did each motion take? Which movements were inefficient? What could be standardized, optimized, eliminated?
The genius of Taylor — or the horror, depending on where you stood — was that he made work legible. Once you could measure it, you could optimize it. Once you could optimize it, you could systematize it. Once you could systematize it, you didn’t need the original craftsman anymore. You needed someone who could follow the system.
Meta has Taylor’s clipboard. It’s called MCI. And it doesn’t need to stand on a factory floor — it runs in the background while employees go about their day, generating a data portrait of exactly how human knowledge work gets done. Every shortcut. Every workflow sequence. Every decision point.
The legibility is the first step. The agents are the endpoint.
What’s different this time is speed and scale. Taylor needed years to study a factory and redesign its workflows. Meta’s MCI is capturing behavioral data across hundreds of platforms simultaneously, from thousands of employees, and feeding it directly into model training. The optimization isn’t happening in slow motion on paper — it’s being encoded into systems that will keep running after the people being studied are gone.
The Taylorist insight was that if you could fully describe how skilled work gets done, you could detach the skill from the person. Meta is doing that. Just with machine learning instead of stopwatches.
The other thing Taylor understood was that the workers being studied often helped him do it. They didn’t have much choice. Your job was to show up and do the work. The fact that someone with a clipboard was documenting your methods wasn’t framed as exploitation — it was framed as progress, efficiency, improvement. The workers who cooperated were the modern ones. The ones who resisted were obstacles.
Sound familiar? Employees at Meta are graded on AI adoption. The metric is adoption — how much of your workflow you’re already handing to AI systems. To resist is to score poorly. To cooperate is to feed the system data about how your job gets done. The clipboard is now a performance review.
What’s new is the velocity. Taylor’s optimization projects took years. The Bethlehem Steel study he’s most famous for ran across multiple years of observation, analysis, and redesign. MCI is compressing that timeline dramatically. It’s not studying one workflow at a time — it’s capturing hundreds of workflows across thousands of employees simultaneously, feeding into model training that updates continuously. The industrial-era playbook is the same. The execution is orders of magnitude faster.
And unlike Taylor’s clipboards, MCI doesn’t require the observer to understand what they’re watching. The data goes to the model. The model finds the patterns. You don’t need a Frederick Taylor standing there with a theory — you just need enough behavioral data and enough compute. Meta has both.
The Company Is Not Struggling. That’s the Point.
Let’s kill one misconception before it takes hold: this is not a story about a company in trouble cutting costs to survive.
Meta’s 2025 revenue was $200.9 billion. Up 22% year-over-year. Q4 2025 alone: $59.89 billion, beating estimates, up 24%. Operating profit well above $60 billion.
The company is not bleeding. The company is profitable at a scale that most humans can’t visualize.
And yet: $115–135 billion in capital expenditure in 2026 — nearly double the previous year. A new division, Meta Compute, explicitly building “tens of gigawatts” of AI data center capacity. Alexandr Wang, 29 years old, former CEO of Scale AI, brought in as Chief AI Officer. His internal memo closed with: “Superintelligence is coming.”
What’s happening here is capital reallocation. Human capital is being liquidated to fund machine capital. The 8,000 employees being cut aren’t the result of a struggling business — they’re a choice made by an extraordinarily profitable business that believes AI infrastructure generates better returns than the people it’s replacing.
That’s a different conversation than the standard layoff story. This isn’t “we have to cut to survive.” This is “we choose to cut because we’re betting the future on something other than you.”
Bloomberg Opinion didn’t soften it: “Meta Is Making Workers Train Their AI Replacements.”
That’s not a critique. That’s a description.
What Meta Did Not Say
The layoff announcement came from Meta executive Dana Gale. The framing: “running the company more efficiently” to “offset other investments.”
AI was not mentioned. Automation was not mentioned. The connection between MCI and the layoffs was not stated.
It didn’t have to be. The Reuters reporting established the timeline. The internal memos described the purpose. The Bloomberg Opinion headline said what the company wouldn’t. The employee reaction — panic in internal Slack boards, per CNBC — confirmed that the people inside the building drew the same conclusion.
Meta has not stated, in any internal document or public communication, that the MCI data is being used specifically to replace the roles of the employees being laid off. They don’t have to. The program design, the timeline, and the $135 billion capital reallocation do the work.
What they did do: they did not tell employees that their workflows were being captured for AI training purposes before the surveillance started. There is no opt-out. Employees are graded on their AI adoption rates — meaning the performance metrics include how well they cooperate with the very systems being built to replace them. And nobody announced any of this. It came out through a Reuters exclusive, after which the internal forums caught fire.
Gizmodo’s read — sourced from the Reuters reporting — was blunt: “Workers are essentially being told they are training the systems that will replace them.” That’s not editorializing. That’s reading the internal memos.
The timeline raises legal and ethical questions too. Companies have disclosure obligations when they conduct surveillance on employees, but those obligations vary significantly by jurisdiction, by what exactly is being captured, and by whether employees were notified in advance in some buried policy document. The question of whether Meta violated any of those obligations is for regulators and courts. The question of whether employees knew what they were agreeing to is easier to answer: they didn’t. They found out from Reuters.
The Dual Purpose
Here’s the frame that makes this story different from every other AI-and-layoffs piece you’ve read this year.
Most companies are doing one thing: either using AI to automate work, or laying people off for other reasons. The two overlap more often than the press releases acknowledge, but they’re usually presented as separate.
What Meta did in April 2026 is both, at the same time, with a documented mechanism connecting them.
The MCI surveillance is the extraction mechanism. It takes tacit knowledge — the how of human work, the things that experienced employees do without thinking — and encodes it into training data. The agents being built in Meta Superintelligence Labs then train on that data. That’s not coincidence. That’s a pipeline.
Meanwhile, 8,000 of those employees are gone by May 20. The data collection continues. The model training continues. The capital investment continues.
The employees who remain know they’re being monitored. They know their workflows are being logged. They know they’re graded on AI adoption. A number on the wall tells them what percentage of their engineering output needs to be “agent-assisted.”
You are simultaneously being surveilled to build the agents, graded on your adoption of AI tools, and watching 10% of your colleagues be removed. That’s a specific kind of workplace pressure that Taylor would have recognized instantly.
This Is Not Just a Meta Story
Q1 2026 tech layoffs: 78,557 workers across the industry. AI cited in nearly half the cases.
The Microsoft parallel deserves more than a bullet point.
Over 2025 and into 2026, Microsoft cut approximately 15,000 employees — engineers, product managers, customer support staff across multiple divisions. The press releases cited efficiency. Restructuring. Operational streamlining. AI was not prominently named.
What was named, in financial filings and investor calls: Microsoft’s deepening commitment to OpenAI. Their total investment has exceeded $13 billion — the largest bet on external AI capability in the industry. The two facts — 15,000 people out, $13 billion into OpenAI — are the same sentence written in different languages.
Here’s what makes the Microsoft case its own story: they chose a different mechanism than Meta. Meta is building its AI capability in-house. Capture how employees work via MCI, feed it to the models, train the agents internally — vertical integration, from human workflow to machine replacement. Microsoft is buying its AI capability from OpenAI: license the models, embed Copilot across every product surface, let OpenAI’s training infrastructure do what MCI is doing for Meta.
Buy vs. build. Different execution, identical logic: the future runs on machines, and the humans currently doing the work are the budget item.
Amazon’s ~30,000 cuts followed the same pattern. AWS is the dominant cloud provider for the AI infrastructure boom that is displacing the workers — they’re selling the picks and shovels while laying off the miners.
The pattern is consistent: profitable companies cutting headcount by choice. Not because the business is failing. Because they believe AI infrastructure generates better long-term returns than the people on payroll.
Business Insider called it “the classic layoff switcheroo.” You frame it as efficiency. You don’t mention the machines. You move the capital and the journalists write about the numbers.
What makes Meta’s situation unusual is the documentation. The internal memos. The Reuters exclusive. The Bloomberg Opinion headline. The CNBC confirmed employee list of tracked platforms — which included, originally, ChatGPT and Claude, before internal blowback got them removed. Most companies doing this do it quietly. Meta got caught with their clipboard out.
The Cumulative Picture
It helps to zoom out slightly and see the full shape of what’s happened at Meta since 2022.
Approximately 25,000 employees have been cut since that year. The 8,000 in May 2026 are not an isolated event — they’re the latest in a sustained, multi-year reduction. The 2022 and 2023 cuts happened during a period of genuine business pressure. The 2026 cuts are happening at peak profitability.
What changed between 2022 and 2026 isn’t the business outlook. It’s the capability horizon. AI agents are now close enough to doing real knowledge work that the calculus has shifted. In 2022, you cut because you had to. In 2026, you cut because you’ve decided machines are the better long-term bet. That’s a different motivation, and it deserves to be named as such.
The $115–135 billion capex figure isn’t going to Meta’s shareholders. It’s going into infrastructure — data centers, compute, energy. New division: Meta Compute, explicitly targeting “tens of gigawatts” of AI capacity this decade. Alexandr Wang, 29 years old, brought in from Scale AI to run Superintelligence Labs, with a memo that closed: “Superintelligence is coming.”
Meta is not hedging. They’re going all in. The employees are the cost being cut to fund the bet.
The Honest Counterargument
There’s a real version of the pushback here, and it deserves acknowledgment.
The $135 billion bet could fail. Superintelligence timelines could slip. AI capability improvements could plateau. If the agents don’t deliver on their promise, Meta will be sitting on $135 billion in infrastructure and 8,000 fewer experienced employees — and the institutional knowledge those employees carried walked out the door in May 2026.
The Guardian put it plainly in April: “Tech companies are cutting jobs and betting on AI. The payoff is far from guaranteed.”
That’s the real risk. Not to workers — they’re already bearing the cost — but to the companies. An overfitted bet on superintelligence timelines, against a workforce that knows how to do things the agents haven’t learned yet. MCI is filling that gap. Whether it fills it fast enough, or well enough, is genuinely uncertain.
The Question That Actually Matters
This is the part that won’t be in the quarterly earnings call.
At what point does a company have an obligation to tell its employees what their work is being used for?
Meta employees did not know, before Reuters reported it, that their daily workflows were being captured for AI training. They found out from a news article. They cannot opt out. And the announcement of 8,000 layoffs came 48 hours after the surveillance story broke.
If you worked at Meta in April 2026, here’s what you experienced: you learned that your computer was recording everything you did, and two days later you learned that 10% of your colleagues were being let go. The connection between those two facts was not explained to you by the company. You had to read Bloomberg Opinion to get the headline.
That’s the story that matters beyond Meta. Not whether the $135 billion pays off. Not whether the agents end up being capable. Whether companies, in the middle of one of the most significant workforce transitions in economic history, have any obligation to be honest with the people they employ about what’s happening to the value of their labor.
Meta chose not to be. They let Reuters do it for them.
The pattern being set in April 2026 is going to matter for every employer and every employee for the next decade. The technology will spread. The MCI approach — capture how experts work, encode it into training data, build the agents, cut the headcount — is not a Meta-specific experiment. It’s a template.
The question isn’t whether this happens more broadly. It’s already happening. The question is whether the people it happens to will know what’s being done with their work while it’s happening.
So far, the answer at one of the most profitable companies in the world is: no. They won’t know. The workers at Bethlehem Steel didn’t get a memo either.
Sources: Reuters (April 17 and 21, 2026), Bloomberg, Bloomberg Opinion, CNBC, The Verge, BBC, Gizmodo, WSJ, Guardian, TechCrunch, Data Center Dynamics. All claims drawn from verified tier-1 reporting.