In the last fourteen days, seven CEOs wrote essentially the same memo. Different companies. Different industries. Different continents. The message landed in shareholder letters, earnings calls, podcasts, and securities filings, and when you lay them side by side, the synchronicity is unnerving.
Snap cut a thousand jobs on April 15, framed by its CEO as a shift to "AI-powered squads." Block had already laid off four thousand people in February, roughly forty percent of its workforce, and is now running a public experiment in what its CEO calls a "mini-AGI" company. Amazon has cut thirty thousand corporate roles across two rounds since last October. Oracle let thirty thousand go on March 31. Atlassian announced sixteen hundred layoffs in a March securities filing that explicitly named AI as the driver. JPMorgan's boss told shareholders the future belongs to "small, Navy SEAL" teams. Meta's AI team is now, by some accounts, running at a roughly fifty-to-one employee-to-manager ratio. OpenAI's founder predicted in 2024 that the first ten-person, billion-dollar company was coming. That prediction is no longer a prediction.
None of it was coordinated. That is what makes it historic. When enough independent executives arrive at the same playbook in the same quarter, you are not watching a trend. You are watching an operating system rewrite.
This is the story of the Great Unstaffing. Who is being deleted. Who is being rewarded. What the companies using AI to shrink are getting wrong. And the one question every knowledge worker should be asking themselves starting Monday.
Part One: Why this is happening right now, not five years ago
For two decades, corporate software and corporate hiring moved in lockstep. More product meant more engineers. More engineers meant more managers. More managers meant more reporting layers. More layers meant more meetings, more status updates, more information routing. The middle of the org chart swelled because it had a real job to do: translate messy human work into legible executive reports, and translate executive decisions back into coordinated team action. For most of the twenty-first century, that translation layer was the only way a company of any size could function.
AI did not just automate a set of tasks. AI automated the translation layer itself.
A single modern agentic system can now route information across teams, summarize the state of a project, flag anomalies, write the status update, schedule the meeting, draft the memo, tag the relevant stakeholders, and produce the executive-facing brief. None of that requires judgment. All of that used to require a director, a program manager, and a small army of senior individual contributors writing slide decks for an audience of one. Take out the translation layer and the company still runs, but suddenly it runs at a very different shape.
That is the shape that seven CEOs are all converging on, independently, in April 2026. It looks like this: a small handful of humans with real judgment at the top. A large number of humans at the edge, closer to customers and to the physical world. And a thin, but dense, layer of AI in between, doing what the middle used to do, faster and without coffee breaks.
The reason it is happening this quarter and not two years ago is simple. The tools finally work. Agentic coding systems, coding copilots, browser-operating agents, voice agents, meeting recorders that actually summarize correctly. A CEO who flattened her org in 2023 would have been running blind. A CEO who flattens in 2026 has infrastructure.
Part Two: Who is being deleted
Be honest about what is happening. The "Great Unstaffing" is not uniform. Specific roles are being removed and the pattern is clear.
The middle manager of the middle manager. The VP who ran the team that ran the teams. The director whose whole job was to produce the dashboard the C-suite skims every Monday. Block's CEO has stated his goal publicly, no euphemism: a flat company where all six thousand employees report to him, with at most two to three layers between any worker and the chief executive. He says it sounds ridiculous only because old hierarchies made it seem ridiculous. When the intelligence layer is handling the routing, he argues, six thousand direct reports stop being theoretical.
The coordinator. The role whose whole function was to make sure the left hand and the right hand were not contradicting each other. Job postings for middle management fell double digits year over year in 2025. That is not noise. That is a category being zoned out of the employment market.
Junior talent. This is the wound nobody is talking about, and it is deeper than the manager cut. When you automate the work that junior employees used to learn on, you break the on-ramp to senior work. A Harvard Business School researcher put it plainly in one of the reports that surfaced this month: "You don't have as many people to promote to seniors. Junior talent is not going to be junior forever." Companies that cheerfully announce a sixty-six times productivity multiplier in one quarter are also, in the same move, quietly removing the pipeline that would have produced their next generation of staff engineers, staff designers, senior account leads, and VPs of product. The math works great for this fiscal year. It is catastrophic on a ten-year horizon.
The producer of artifacts no one reads. The PowerPoint artisan. The status-report compiler. The meeting-minutes rewriter. If your job output could be described in one sentence as "producing a document that someone more senior will glance at," your job was already under threat. AI just accelerated the clock.
Notice who is not on this list. Account executives who understand a customer's politics. Nurses. Construction foremen. Designers who sense when a layout is one pixel wrong. Comedy writers. Negotiators. Trial lawyers. Ops leaders who can look at a room of exhausted people and know when to push and when to let up. The people AI cannot replicate, or replicate badly enough that it is obvious.
Which raises a question the optimistic press release never answers: what happens to a company when it strips itself down to mostly the executive tier plus edge operators, with no training ground in the middle?
Part Three: The counter-evidence nobody wants to publicize
Here is a number that most tiny-teams evangelists will not put in the tweet: the European Central Bank studied more than five thousand euro-area firms this year and found that companies investing in AI for research and development are four percent more likely to hire than firms that are not. They are nearly two percent more likely to grow total headcount. In Europe, only fifteen percent of companies say cost reduction is their reason for adopting AI. The rest are trying to build something new, and they are doing that with more people, not fewer.
The ECB data reveals a fork in the road that American headlines are missing. There are two kinds of AI adopters.
Type one uses AI to do what they already do with fewer people. The goal is margin. The signal is in the layoff notice. The visible output is a share-price bump and a headline.
Type two uses AI to do things they could not previously do at all. The goal is new product, new market, new capability. The signal is a hiring plan. The visible output is revenue growth. And according to five thousand European firms, the type-two companies are winning on employment and, by implication, on the only metric that matters over five years: are you building something worth building?
It is worth reading the same sentence twice. The companies using AI to shrink are shrinking. The companies using AI to build are hiring. Same tool. Opposite outcomes. The distinction is not the technology. The distinction is the intent of the person deploying it.
That means the Great Unstaffing narrative, as currently framed in the financial press, is telling you only half the story. The half that gets clicks.
Part Four: The historical mirror
This is not the first time a workforce has been restructured around a new kind of machine.
In the 1910s, Henry Ford's moving assembly line took what had been a skilled, generalist workforce of automotive craftsmen and split it into two tiers: a small engineering elite that designed the process, and a large army of narrow-task operators who performed one motion, over and over, eight hours a day. Overnight, a generation of skilled mechanics had no place in the new org chart. The public response was a fury that reads familiar a century later.
What actually happened next is the part we forget. Ford paid his narrow-task operators five dollars a day in 1914, more than double the going rate, and in doing so invented the American middle class. The assembly line did not just restructure manufacturing. It restructured the social contract. The operators who had been displaced eventually retrained into new roles: line foremen, quality inspectors, industrial engineers, union leaders. The new roles could not have existed without the assembly line. They also could not have existed without the displacement.
In the 1980s, the same movie ran again. Desktop computers rolled into corporate offices and erased a generation of secretarial roles, typing-pool staff, carbon-copy clerks, and mimeograph operators. Companies underwent what was then called, without apology, the Great Flattening. Ranks of middle managers were culled. The same arguments were made then that are being made now. "There's no need for all these layers." "One person with a PC can do the work of a department." The layoffs were brutal. The predictions of a permanent underclass were loud.
What actually happened next, again, is the part we forget. The PC created the knowledge-worker economy. Web designer, software engineer, data analyst, digital marketer, UX researcher. Entire professions that could not have existed without the technology that had just gutted the prior generation of office roles.
History does not tell us AI will create as many jobs as it destroys. History tells us we have to actively build the bridge, because nobody else will. The workers displaced by the assembly line did not fall into the middle class by accident. They fought for it, organized for it, and in Ford's case, were paid into it because Ford needed them to eventually buy his cars. The clerical workers displaced by the PC did not become software engineers on their own. A massive retraining apparatus, public and private, was built around that transition. It was painful, it was incomplete, and it worked anyway.
Someone has to build that bridge again in 2026 and 2027. Companies that delete their junior pipeline in April are, by the end of the decade, going to be begging for experienced talent in a market that no longer produces any. The rational move, for a company that plans to still exist in ten years, is to protect the bottom of the ladder even as you thin the middle. The companies doing the opposite are making an explicit bet that in ten years they will not need experienced humans at all. That is either the boldest prophecy of our lifetime or one of the great strategic errors of the century. Possibly both.
Part Five: What the data actually says about the productivity multiplier
The number that is driving every boardroom conversation right now came from one paragraph in one shareholder letter. Amazon's CEO wrote that when his company needed to rebuild the inference engine underneath its Bedrock AI platform, the traditional approach would have required a team of forty engineers working for a year. Instead, six engineers using an agentic coding service delivered the new engine, which they named Mantle, in seventy-six days.
Do the math. Forty engineer-years versus roughly one and a quarter engineer-years. Sixty-six to one, on a single project, at a specific moment in time. That is the number that launched a thousand layoff memos.
Before you treat that as the law of nature, read the footnotes that did not make the press release.
First, the six engineers who were assigned to Mantle were, in the CEO's own words, "very skilled." They were not a random sample of the Amazon engineering workforce. They were the top fraction of a heavily curated talent pool that Amazon spent two decades building. The productivity multiplier is not coming from the agentic coding tool alone. It is coming from the tool in the hands of engineers who would have been ten-times developers even without it.
Second, the forty-engineer baseline comes from Amazon's own historical experience building the same kind of thing in a pre-AI era. It is not a rigorous controlled experiment. It is an after-the-fact comparison by an executive with an obvious incentive to make the comparison flattering.
Third, and most quietly, the rest of the Bedrock platform still exists. The Mantle team did not replace the Bedrock org. It delivered one component that used to require a larger build. The total headcount implications for the broader platform are not in the shareholder letter.
None of this is a case against Mantle. The story is real. The productivity gain is real. But sixty-six times is not the number you should use to plan the next decade of your career. The more honest framing is this: in the hands of top-tier operators, modern AI tools multiply output by at least one order of magnitude on the kinds of tasks where the model can see a correct answer. On tasks where there is no training example, where the correct answer is a judgment call, where the stakes depend on reading a specific human being, the multiplier drops to zero. Or goes negative, when the model confidently hallucinates.
The question every individual contributor should be asking is not "can AI do my job." The question is "what fraction of my job is a task the model has been trained on ten thousand times, versus a task that has never been done before."
The first fraction is disappearing. The second fraction is where your compensation lives for the next decade.
Part Six: What makes a worker undeletable
Read through the counter-voices buried in the tiny-teams reporting and a pattern emerges. The experts who are skeptical of the flat-company revolution are not skeptical of AI. They are skeptical of any organization that assumes the remaining humans are interchangeable.
A Stanford economist who has spent his career on labor and technology put it the cleanest: the winners of this era will not be the leanest organizations. They will be the ones that redesign work so that humans and AI complement each other. Leanness is easy. Redesign is hard.
A Harvard Business School professor, writing about the same trend, warned that cross-functional teams, even small ones, need what she calls "bridgers." People who can translate across disciplines, who can host a useful disagreement between two specialists, who can sense when a room is heading in the wrong direction. Those people are rarely the loudest in the meeting. They are almost never the ones producing the most artifacts. They are expensive to identify and nearly impossible to replace. They also do not show up in the kind of data a cost-cutting CFO looks at.
A former senior leader at Netflix, who helped write what became one of the most influential HR documents of the century, made a point this month that is worth writing on a Post-it and sticking to your monitor. AI, she said, is not going to take humans out of the equation. It is going to take some tasks out, just like factories did. The equation still has a human variable. The question is what that variable represents.
Another voice, a researcher at an HR technology company, gave us the phrase of the quarter: "silicon sycophant." Modern AI is tuned to keep you engaged. It is therefore biased toward telling you what you already want to hear. Remove the human friction from your decision-making and you will, eventually, remove the one thing that used to push back when you were wrong. The flatter an organization gets, the fewer people there are to say the emperor has no clothes, and the more the remaining voice in the room is the AI that has been trained to agree with the person in charge.
A political psychologist, researching employee engagement at a workforce-recognition company, reminds us of something softer but just as real. Engagement is driven by interaction with colleagues and with leaders. Strip that interaction out too aggressively and "inspiration can decay." Productivity tools are not the same as belonging. A company of individuals who each have a perfect agent on their laptop can still feel, to each of those individuals, like an empty office with a humming server.
Put these voices together and the picture is not cynical. It is coherent. The undeletable human is not competing with the model on the model's turf. The undeletable human is doing the thing the model cannot do: exercising judgment in a new situation, building trust across people who have reason not to trust each other, sensing cultural context, noticing what the data is not saying, making the hard call that cannot be justified with a spreadsheet.
If you can describe your work in one sentence, and that sentence ends with "and then I send the result to someone else," you are in the wrong half of the fork. If the sentence is "and then I make a judgment call that requires context nobody else in the room has," you are in the half that is getting more valuable, not less.
Part Seven: What to do on Monday
Forget the macro story for a minute. Zoom into your own week.
Open your calendar. Look at the last five working days. Every meeting. Every email thread. Every deliverable. Every deck you touched. Now put each one into one of two buckets.
Bucket A: tasks whose job was to route information from one human to another. Weekly sync. Status update. Cross-functional alignment email. "Here are the takeaways from our customer call." "Here is where we landed on the roadmap." Anything that an agent with good tools and a complete information graph could produce without you.
Bucket B: tasks whose job was to make a judgment call, hold a difficult conversation, build trust across a friction boundary, or read the room in a way the data alone could not predict. The ten-minute phone call where you talked a frustrated customer off a ledge. The one-on-one where you told a direct report the hard truth. The design review where you said "something is off here, I can't articulate it, we need another week."
Look at the ratio. If Bucket A is more than sixty percent of your week, your current role is, quietly, in the red zone. The work you are doing will shrink over the next twelve months. This is not a prediction. It is already happening at the companies named at the top of this essay.
If Bucket B is where most of your energy lives, you are in the right lane for the decade.
Either way, Monday's homework is the same. Start moving time from Bucket A to Bucket B. Delegate the routing to an agent, even a dumb one, even a flaky one. Reclaim the hours. Put those hours into the Bucket B work that was getting crowded out by meetings. Have the harder conversation. Take the customer call that does not scale. Write the memo that says what the data is not saying. Make the judgment call you have been avoiding because there was no bulletproof spreadsheet for it.
And above all, protect the people on your team whose Bucket B work is not currently getting measured. The junior on your staff who is learning the hard skill of reading the room. The middle manager whose real job has always been acting as a bridger and who is about to be written out of the organizational chart because the title on the badge does not match the work being done. Those people are the pipeline to the next decade of your company. The ones losing them do not realize what they are losing until it is too late to replace.
Closing
A lot of people will write this decade the way they wrote the last one. "AI took my job." There will be true stories of displacement. There will be cruel layoffs. There will be waste. There will be a policy debate that arrives, as usual, a year too late to help anyone.
But there is a parallel story that is already being written, in Europe, in the companies that are using AI to build new things instead of to trim the old ones. In the worker who realizes she is twenty percent coordinator and eighty percent bridger, and decides to lean into the eighty. In the junior engineer who skips the meetings the agent can attend for him and uses the reclaimed hour to ship something a human has never shipped before. In the CEO who, six months after flattening her company, realizes she needs her mid-layer back, but in a new shape, with a new name, and a new job description that centers on judgment instead of on routing.
The Great Unstaffing is not a one-time event. It is a sorting function. And it is sorting right now.
Go be one of the humans they cannot delete.