What we lose when AI takes the work that used to make people
There is a particular type of news story that has been circulating for a while now, and it always lands strangely.
A government somewhere releases a labor projection. The numbers say that, by some future year, office workers will be in surplus and AI-related specialists will be in shortage. Commentators line up. Some warn about an "AI job apocalypse." Others reassure us that retraining will solve it. A few argue the projections are wrong. The conversation cycles, and then it moves on.
The first time you read one of these stories, it makes sense. The second time, less so. By the fifth or sixth, you start noticing that something about the framing is off, even if you cannot immediately name what.
Recently, Japan's Ministry of Economy, Trade and Industry released one of these. The projection said that by 2040, the country would see roughly 4.4 million surplus clerical workers, against a shortage of about 3.4 million AI- and robotics-capable specialists, plus another 2.6 million missing in frontline trades. The headlines wrote themselves. Clerical workers in surplus. AI specialists in shortage. Retrain or perish.
I want to argue, gently, that this framing is misleading. Not because the numbers are wrong, but because the numbers describe a symptom and the conversation keeps mistaking it for the disease.
What is actually happening, I think, is something quieter and considerably harder to see. The work that AI is absorbing is not just work. In many organizations, it was also the substrate through which inexperienced people gradually became experienced ones. When that substrate disappears, you do not simply lose some jobs. You lose the conditions under which the next generation of judgment-capable humans was supposed to form.
This is the part that does not show up in any projection.
The shallow story
The shallow story goes like this. AI is automating tasks. The tasks happen to cluster in certain kinds of office work. Therefore certain office workers will be displaced. Retraining will move them into roles AI cannot yet do. The economy adjusts. Some friction, but ultimately fine.
Each link in that chain contains a small lie of compression.
Start with the assumption that "AI is automating tasks." It is more accurate to say that AI is automating a particular layer of work — the layer where inputs are structured, outputs are predictable, and the space for judgment is narrow. This layer cuts across job titles. It exists inside lawyering and engineering and design and consulting and journalism. It exists inside roles labeled "creative" and roles labeled "analytical." When we talk about clerical work being automated, we are using a 1950s job category to describe a 2026 phenomenon. The mismatch matters.
Then there is the assumption that workers can simply be moved. The implicit model is that a human in role A can be retrained into role B, the way you might recompile software for a new platform. This works for narrow skills. It does not work — and cannot work, for reasons I want to come back to — for the kind of capability that takes a decade of being slowly wrong about things before you become reliably right.
But the deepest lie of compression is in the word adjusts. The economy adjusts. As if the only thing being rebalanced is the supply and demand of labor units, rather than the underlying machinery by which a society produces capable adults.
The deeper problem may be that we are watching that machinery come apart, one piece at a time, and the labor-market language we have inherited cannot quite see it.
What entry-level work was actually for
If you have ever managed a junior employee, you know that their work, in the first year or two, is rarely for the work itself.
A junior analyst takes the meeting notes. The notes are useful, but the real point is that taking the notes forces them to follow the meeting closely enough to understand who in the room actually decides things, who is performing certainty, who is hedging, and what the unspoken disagreements are about. The notes are the tuition; the meeting is the school.
A junior associate puts together the first draft of a client deck. The deck is rough. The senior will rewrite half of it. But the act of producing it — choosing what to lead with, where to put the numbers, how to phrase the recommendation — is the first time the junior has had to make those choices at all. The senior's rewrite is the lesson. Without the first draft, there is nothing to be rewritten.
A junior engineer fixes the tiny bug nobody wants to look at. In the process, they have to read code that other people wrote, in styles they did not choose, solving problems they did not pose. The bug fix is the deliverable; the orientation through someone else's reasoning is the actual asset.
I could go on. Every white-collar profession I know of has some version of this. The visible output is one thing. The hidden curriculum is something else entirely. The two are entangled, and for most of the modern era, they were entangled by accident — the work happened to require junior labor, and the junior labor happened to produce learning as a byproduct.
The entanglement was never planned, which is one reason it is now disappearing without anyone deciding to remove it. AI does the meeting notes. AI does the first draft of the deck. AI fixes the tiny bug. The output gets produced faster and often better. The byproduct — the slow, granular, accidental education of a human being — simply stops happening.
Nobody is opposed to this. There is no faction defending the dignity of the unedited deck. The transition is happening because every individual decision to use AI is locally sensible. It is the aggregate that should worry us.
Skills and judgment are not the same thing
Most of the conversation about AI and the future of work is conducted in the vocabulary of skills. People need new skills. Retraining is about skills. The shortage is a skills gap. The mismatch is between the skills supplied by the workforce and the skills demanded by the economy.
This vocabulary works for some of what is happening. It is also doing a lot of damage by quietly conflating two very different kinds of human capability.
Skills are things you can be taught. Programming languages, statistical methods, financial modeling, project management frameworks, written persuasion in a particular genre. Skills travel. You can read a book about them, take a course, watch a tutorial, ask an AI tutor. Skills can be acquired in months, sometimes weeks. They can be lost too, but they can be reacquired.
Judgment is different.
Judgment is what you have after you have made several thousand small decisions under conditions of incomplete information, watched some of them go badly, sat with the consequences, adjusted, and tried again. It is not portable in the way a skill is portable. It cannot be transferred from one person's head to another's. It cannot be reliably manufactured by a curriculum, because the essential ingredient is the experience of being responsible for an outcome you only partially controlled.
Judgment compounds slowly. It is the kind of capability that, in the right environment, accumulates almost invisibly over a decade — and then, one day, manifests as a person who can walk into a meeting, scan the room, ask three questions, and identify the actual problem. The walking-in part takes thirty seconds. The accumulating part took ten years of being given just enough rope to occasionally make small mistakes.
There is no shortcut. There has never been a shortcut. The reason senior people are valuable is not that they possess some collection of facts a junior could be tested on. It is that they have run thousands of small cycles of decide-observe-revise, and the residue of those cycles is what we call experience.
When we talk about retraining people for the AI economy, we are mostly talking about skills. This is the easier conversation. It is also the conversation that lets us avoid the harder one: what happens to a society when the conditions under which judgment used to form, quietly and almost by accident, are being optimized out of existence.
Experience as a kind of capital
It helps, I think, to treat experience as a form of capital — not metaphorically but operationally, the way economists treat human capital or social capital.
Experience capital has some specific properties.
It accrues only through time spent doing the thing. There is no synthetic substitute. You cannot front-load it. You cannot inject it. You cannot purchase someone else's stock of it and have it transfer into your own balance sheet.
It depreciates. People retire. People die. People burn out. The total stock in any organization or society is constantly leaking, and it has to be replenished by people coming up through the layers and accumulating their own.
It compounds within an individual, but does not compound across individuals automatically. A senior partner's thirty years of judgment do not seep into the junior associate by proximity. The junior has to do their own decisions, make their own errors, and metabolize their own consequences.
Most importantly: it requires a specific kind of environment to form at all. You need real decisions, not simulated ones. You need real stakes, where being wrong has visible costs. You need someone who has been through it themselves, watching closely enough to tell you what you missed without rescuing you from having seen it.
This last condition is what entry-level cognitive work, for a long time, accidentally provided. The work was real. The stakes were real, even if modest. There was usually someone more senior whose job partially included noticing what the junior was getting wrong. The environment was not designed to produce judgment. It just did, as a byproduct of how organizations had to be staffed.
That environment is now actively being dismantled, in the name of efficiency, and the dismantling is going faster than anyone seems to have noticed.
What AI actually does to experienced people
I want to be careful here, because the picture is genuinely two-sided.
For people who have already accumulated significant experience capital, AI is something close to a miracle. A senior consultant who has spent twenty years thinking about how organizations make decisions can now use AI to produce in a week what used to take a team a month. A senior engineer can ship features that used to require coordinating with a frontend specialist and a designer. A senior researcher can read across literatures they never would have had time to read.
AI is, for these people, a kind of cognitive prosthesis. It absorbs the rote layer of their work and amplifies the judgment layer, which they happen to have in abundance.
This is the source of a lot of the genuine optimism in current AI discourse, and it is not wrong. If you watch what people who are good at their work are doing with these tools, it is impressive. The one-person company that does what used to require ten people is not hype — it is happening, in narrow domains, and it will keep happening.
The discourse about "solopreneurs" running fully leveraged operations with AI agents handling coding, sales, support, and operations is real. Sam Altman has talked publicly about the eventual emergence of the first one-person billion-dollar company. Whether or not that specific milestone arrives on schedule, the underlying trend — a single experienced operator running what used to be a small division — is already visible across several industries.
But here is what gets missed in that story.
The people running these leveraged operations are, almost without exception, people who put in their time before the tools existed. They became senior in an environment that still had the apprenticeship layer. They have the judgment because they once had the boring work. They acquired the pattern recognition because they once spent years in rooms where patterns were forming, taking notes nobody read.
AI amplifies them because they were already amplifiable. AI does not, in any current sense, manufacture the underlying judgment. It augments what is already there. It does not generate it from scratch.
You can watch this asymmetry in any field that has both senior practitioners and a thinning junior layer. Talk to a senior software architect who is shipping more code than ever and ask them where they would be without the twelve years of pattern recognition that lets them know, almost preconsciously, when an AI-generated solution is structurally wrong. They will usually pause, then say something like: I am not sure a younger me, using these same tools, would have caught what I just caught. Talk to a senior investigative editor reviewing AI-drafted copy and ask them what they are doing when they read it. They will tell you they are checking it against a kind of internal model of how a story should land, built from decades of seeing stories land in ways they did not intend. The model is what is doing the work. The model took thirty years to build. The AI did not contribute to building it.
This is the crucial asymmetry. The same technology that turns an experienced operator into a one-person company is simultaneously eliminating the entry-level work that would have produced the next generation of experienced operators. The amplification is real. So is the erosion. They are happening at the same time, and they are being driven by exactly the same tools.
The market for finished people
Watch any hiring conversation for long enough and you start hearing a strange pattern.
Companies say they cannot find good people. Specifically, they cannot find people who are already capable. The job postings ask for five years of experience, the ability to work independently, demonstrated track record, prior ownership of similar problems. The implicit message is: we want someone who is already finished. We do not want to do the finishing.
Job seekers, for their part, want to work at companies that will invest in them. They want training, mentorship, growth, a clear path. The implicit message there is: we want someone who will do the finishing for us.
Both sides are asking for the same thing from a different direction. Both sides assume someone else is the one doing the developmental work. Neither side is volunteering.
For a long time, it was companies that did this work, somewhat unwillingly, because the structure of employment required it. You hired juniors because you needed labor at that price point. The juniors got worse work and lower pay; you absorbed their slow ramp; eventually they became useful. The arrangement was not generous, but it was functional. Apprenticeship was a side effect of how organizations had to be staffed in order to operate at all.
That necessity has eroded from two directions at once.
From one direction, the rise of flexible labor — contractors, agencies, fractional executives, project-based hires — meant that companies could increasingly procure capability without ever having to grow it themselves. Why train a junior for five years when you can rent a senior for six months? The math, on any individual hire, almost always favors renting.
From the other direction, AI is now performing the bottom layer of the work that juniors used to do. Even if a company wanted to hire and develop a junior, there is increasingly less work for that junior to do during the period when they would have been getting worse than the senior at it.
The result is a market that wants finished people, and a development pipeline that is producing fewer of them. The mismatch is not immediately visible because the existing stock of finished people is still circulating. Senior practitioners are still around, still hiring out their judgment, still solving problems. The system runs on inventory.
What happens when the inventory runs out is the question nobody seems to want to ask out loud.
The illusion that the system is fine
For now, things look manageable.
The senior cohort is large, productive, and increasingly AI-amplified. Companies that need work done can find someone to do it. AI handles the routine layer. Output is, by many measures, going up. There is no obvious crisis. The labor projections suggest some friction in transitions, but nothing catastrophic.
This is, I think, a temporary equilibrium that disguises a structural problem. The senior cohort will retire. The current pipeline is producing fewer fully-developed replacements per year than it used to. The work that would have developed those replacements is increasingly being done by tools that do not transfer their capability back into any human.
If you imagine a tank with water flowing in at the top and water draining out at the bottom, the level looks steady as long as the rates match. We are watching the inflow narrow, slowly, while the outflow continues at its usual rate. The level looks fine. It will keep looking fine for some years. Then, fairly suddenly, it will not.
The years before the level visibly drops are exactly the years in which it would be possible to do something about it. They are also the years in which it is hardest to make the case that something is wrong, because nothing visible is wrong.
This is the part that I think is being missed in the broader conversation. The crisis, if it comes, will not arrive as a sudden shortage. It will arrive as a slow, generational thinning of the layer of people capable of doing the work that AI cannot do — the work of figuring out what is actually going on, what to do about it, and how to do it under conditions the tools were not designed for.
By the time that thinning is visible in the aggregate numbers, it will have been compounding for fifteen or twenty years.
I think this is why the conversation feels so off. The optimists are looking at the current senior cohort, watching them do astonishing things with AI, and concluding that the future of work is more or less fine, perhaps even unprecedentedly good. They are not wrong about what they are seeing. They are looking at the right thing. They are just looking at a snapshot of a stock, and missing the change in the flow.
The pessimists, for their part, are looking at the job losses at the bottom and warning about mass unemployment. They are also not wrong about what they are seeing. But they are framing it as a redistribution problem — too few jobs, too many people — when the underlying issue is something else: too few jobs of a particular kind, the kind that used to convert people into the kind of people who could do the other jobs.
Neither framing quite captures what is happening, because what is happening is structural rather than cyclical, and our vocabulary for talking about the labor market is overwhelmingly cyclical. We think in terms of booms and busts, hot and cold markets, employer's and employee's years. We do not have a comfortable way of talking about the slow disassembly of a developmental machine that took half a century to build.
What is being optimized away
It might be worth naming, plainly, the categories of organizational behavior that are being optimized out of existence.
The first is the act of giving a junior person a piece of real work and letting them do it badly enough to learn from doing it badly. This was always inefficient. It was inefficient in 1980, and it was inefficient in 2010, and it is inefficient now. The difference is that in 1980, the inefficiency was unavoidable — there was no faster way to get the work done. Now there is, and the unavoidable inefficiency has become a chosen one. Most organizations are not choosing it.
The second is the long, unhurried conversation between an experienced person and a less experienced person, in which the less experienced person learns not just what the senior thinks but how the senior thinks. These conversations were never billed. They happened in offices, in hallways, on the phone, in margins of meetings. They are also among the first things to disappear when work shifts to asynchronous, remote, AI-mediated workflows where every interaction is logged and optimized for throughput.
The third is the willingness of organizations to carry people through periods when they are not yet productive enough to justify their cost. This was historically a feature of corporate life, especially in professions with steep learning curves. It is now widely regarded as poor capital allocation. The framing has changed; the underlying need has not.
None of these are dramatic losses. Each one looks, in isolation, like reasonable cost discipline. Each one is being decided by people who are operating rationally within their incentives. The aggregate effect is what should worry us, and the aggregate effect is, almost by construction, invisible to any individual decision-maker.
This is the shape of the problem. It is not a villain story. It is not even really a technology story, except in the sense that technology is accelerating a tendency that was already there.
It is a story about what gets quietly stripped out of a system when every part of it is asked to justify itself by short-term productivity, and there is no longer anyone whose job it is to defend the parts that only pay off across decades.
Open questions
I do not want to pretend I know what to do about this.
A few directions seem worth thinking about, though I hold them tentatively.
One is that organizations may need to start treating apprenticeship not as a byproduct of operations but as an explicit function — staffed, budgeted, and protected from the pressure to optimize. This is hard to do, and harder to sustain across leadership transitions. It is also probably necessary in any organization that intends to exist in twenty years with capable senior people in it.
Another is that the developmental layer may need to be socialized, in some form, given that no individual employer has a strong incentive to provide it. This is essentially the public-goods argument for education extended into early career formation. It runs into all the usual difficulties of public-goods provision, plus the additional difficulty that the beneficiaries are diffuse and the costs are concentrated.
A third is that we may need to think much harder about what kinds of work AI should not be allowed to absorb, not because AI cannot do them but because the human doing of them is itself the developmental mechanism. This cuts against the dominant logic of efficiency, and it is uncomfortable to articulate without sounding like a Luddite. But the underlying point is serious: not all efficiency is gain, if what is being made efficient was secretly doing two jobs.
None of these are answers. They are at most the beginnings of conversations that, as far as I can tell, are not being had at the level they need to be had.
What does not show up in the projections
The labor projection that started this essay shows surplus clerical workers and shortage AI specialists. That is what it can measure. It cannot measure the absence of the next generation of judgment-capable humans, because that absence will not be visible for another decade or two, and by then the projections will be measuring something else entirely.
The deeper observation, the one I keep coming back to, is that we have built a story about AI and work that is almost entirely about substitution. AI replaces human at task X. Therefore human moves to task Y. The model is mechanical and tidy.
The reality, as best I can see it, is messier. AI is not just substituting for tasks. It is dissolving the developmental scaffolding that, almost invisibly, used to convert inexperienced humans into experienced ones. The scaffolding was made of small inefficient tasks, slow conversations, tolerable failures, and the patient attention of people one or two career stages ahead. None of that scaffolding shows up in any productivity metric. All of it is being removed.
What we are losing, I think, is not work. We have plenty of work, and we will continue to have plenty of work. What we are losing is the conditions under which a person could enter that work as an apprentice and exit it, fifteen years later, as someone whose judgment was worth deferring to.
If that loss is real, and if it is as broad as I suspect, then the question is not how to retrain people for the AI economy. The question is what kind of humans an AI-saturated economy can still grow.
I do not think we know the answer to that yet. I am not sure we have noticed we should be asking.
Ryoji Morii is the founder of Insynergy Inc., a Tokyo-based advisory firm focused on AI governance and organizational decision architecture. He writes about how institutions are quietly being reshaped by AI, and what gets lost when the changes happen too fast to see clearly.