Opinion: AI Hollowing Out Necessary

The Thesis: AI Is Not Taking Our Jobs -- It Is Taking Our Training

AI is not hollowing out work; it is hollowing out the necessary work -- the unglamorous, repetitive, load-bearing tasks that were never the point of a job, but were always the price of admission to one.

That distinction matters more than any unemployment headline you will read this year, because it changes what the threat actually is. The threat is not that there will be no work. The threat is that there will be no first rung. A ladder without a first rung is not a ladder; it is a wall with a nice view.

Here is what I mean by necessary work. It is the first-pass research memo. The boilerplate contract review. The tier-one support ticket. The junior bug fix. The two-hundred-row spreadsheet cleanup that eats an afternoon and quietly teaches you where the data actually breaks. None of it is glamorous, and most people who did it complained about it. But it was the curriculum. It was where you learned what a normal case looks like so that you could later recognize an abnormal one. It was where you built the pattern library that judgment runs on.

Generative AI is extraordinarily good at exactly that category of task: bounded, text-shaped, verifiable, low-stakes. That is not an accident; it is a design target. And when you remove the training layer without replacing it, you do not get a leaner workforce. You get a top-heavy one -- expensive at the top, empty in the middle, and quietly unable to reproduce itself.

Workings.me frames this as the hollow middle problem: the top of the ladder is fine, the bottom is being sawed off, and almost nobody with budget authority is asking who climbs next. That is the argument I want to make in full, because the lazy version of it -- AI is coming for your job -- is both wrong and strategically useless.

The Context: Why This Argument Stopped Being Hypothetical

For most of the last decade, the AI-and-jobs debate ran on forecasts. Consultancies produced big round numbers, economists argued about elasticities, and everyone agreed the timeline was a decade or two out. That debate is over. We now have payroll data, and the payroll data is telling a narrower, stranger story than either the optimists or the doomsayers predicted.

Start with exposure. The IMF's January 2024 staff analysis estimated that roughly 40 percent of global employment sits in occupations exposed to AI, rising to about 60 percent in advanced economies. Crucially, the IMF split that exposure almost evenly between jobs AI would complement and jobs AI would displace. Read that split again, because it contains the entire story. Exposure is not elimination, and elimination is not distributed evenly across a career.

Now go narrower. Stanford's Digital Economy Lab examined what happened to early-career workers once generative AI diffused and found that employment for workers aged 22 to 25 in the most AI-exposed occupations fell roughly 13 percent relative to their peers -- even as overall employment inside those same firms held steady. The firms were fine. The twenty-three-year-olds were not.

Then widen out again. The World Economic Forum's Future of Jobs Report 2025 projects roughly 170 million jobs created and 92 million displaced globally by 2030 -- a net gain of about 78 million -- while also projecting that 39 percent of existing core skills will change. A net gain and a skills earthquake can both be true. They usually are.

Stack those three findings: high task-level exposure, falling demand for the youngest workers, and enormous churn in required skills. The result is not a jobs apocalypse. The result is a broken escalator. The bottom step still exists; it simply no longer moves.

Workings.me has tracked this as a demand-composition problem rather than a headcount problem, and that framing is the only one that survives contact with the data. Nobody is being mass-fired by a model. People are simply not being mass-hired into the roles that used to start careers -- and a career that never starts is statistically indistinguishable from one that was eliminated.

Dimension One: AI Ate the Apprenticeship, and Nobody Noticed the Curriculum Vanish

Every serious profession in the world runs on a deliberate apprenticeship architecture. Medicine has residency. Law has the first three years of document review and drafting. The trades have journeyman status. Software had the junior engineer who fixed small bugs, wrote the tests, and slowly learned why the database schema was a disaster.

That architecture worked for one reason: the small version of the work is how you learn the big version of the work. You cannot reason about a novel contract if you have never read three hundred boring ones. You cannot debug a distributed system if you have never debugged anything. Pattern recognition is built by volume, not by insight.

The problem is that the boring version is precisely what generative AI does best. Drafting, summarizing, classifying, translating, first-pass coding, ticket triage -- all of it is bounded, text-shaped, and easy to verify. Anthropic's Economic Index, which measures real model usage rather than survey intentions, consistently finds coding and writing as the largest categories of use. Those are not senior tasks. Those are the tasks you hand a bright twenty-three-year-old in their second week.

So we automated the curriculum. We did it in roughly thirty months, across nearly every knowledge occupation at once, and we did it without a single serious institutional conversation about what replaces it.

Here is what should bother you more than the layoff headlines: firms are behaving rationally and still producing a terrible long-term outcome. Junior headcount carries the lowest margin and the lowest perceived risk to cut. It also carries the only option value that matters -- it is how your future senior talent gets made. That is a textbook market failure in training investment, the same reason private firms under-invest in apprenticeships unless someone subsidizes them. The payoff lands five to eight years out. Nobody on a two-year tenure is going to pay for it.

Note the asymmetry with the trades. The United States still registers roughly 600,000 active apprentices through the Department of Labor's Registered Apprenticeship program, and most of that training happens on paid work that a machine cannot do. Knowledge work never had a registry, a completion standard, or a subsidy. It had a junior desk. We removed the desk.

Workings.me's position is that the apprenticeship gap, not the automation itself, is the defining labor story of 2026. Tools can be adopted in a quarter. Judgment takes a decade, and it cannot be downloaded.

Dimension Two: The Hollow Middle Is an Institutional Failure, Not a Technology Story

It is tempting to file all of this under technology. That is wrong. Technology set the price; institutions chose to pay it.

Consider the ATM. The standard story is that ATMs were supposed to end bank tellers and did not, because banks redeployed tellers into sales and advisory work. The story is true, and it is missing its most important sentence: the redeployment was a management decision. Someone redesigned the teller job on purpose. It took years, it was expensive, and it worked because banks decided the branch relationship was worth preserving.

Now look for the knowledge-work equivalent. Is any firm you know deliberately redesigning the junior job so that it teaches judgment while the routine parts get automated? Is anyone running a structured first-year program built around AI-augmented decision review? A few are. Most replaced the junior workstream with a subscription and moved on. Meanwhile, roughly eight in ten organizations now report using AI in at least one business function, according to McKinsey's State of AI survey work. Adoption is not the bottleneck. Redesign is.

Why necessary work is invisible on every dashboard

Necessary work has a measurement problem, and the measurement problem is why it keeps getting cut. The costs are immediate, line-itemed, and attributable. The benefits are invisible by construction -- they show up as errors that did not happen three years from now, in the judgment of someone who was allowed to make small mistakes safely. Firms under-invest in things that cannot be measured in the current quarter, which is exactly why training budgets are always the first cut and the last restored.

This is also why Workings.me insists that career durability has to be measured by the individual, not inferred from an employer's behavior. No one is going to hand you a dashboard showing your apprenticeship decaying. You can run the Career Pulse Score to see how your current skill stack, automation exposure, and professional proof hold up against these shifts -- it is the individual-scale version of the question this article is asking at the institutional scale.

What senior workers actually lose

Short term, senior workers win, and it is not close. They get leverage, they get AI multipliers on work they already understood, and they get to delete the parts of the job they always hated. Long term, they lose the thing nobody advertises: a successor. A senior team with no juniors is a team with a fixed expiry date. When the two people who understand the legacy system leave, there is no one left who ever saw it break.

The middle layer absorbs a second hit as well. If every individual is managing a set of agents, the org chart flattens, and the middle is where institutional memory historically lived. Flattening feels efficient right up to the moment the company cannot explain its own pricing model to a new regulator.

The Counter-Argument -- and Why I Still Hold My Position

The strongest objection to everything above is that we have heard it before, and the last hundred years were not kind to the people who said it.

Every general-purpose technology -- electricity, the tractor, the spreadsheet, the ATM -- destroyed tasks and eventually produced more work than it destroyed. The optimist case has real intellectual weight behind it. Economist David Autor's Applying AI to Rebuild Middle-Class Jobs makes the sharpest version: AI is best suited to the last mile of expert work, the judgment and contextual tasks that let more people do more valuable work than before. If that is right, hollowing out is temporary, and the workers who survive it will be more productive than any generation preceding them.

A second, more practical version of the objection goes like this: if entry-level work is genuinely necessary, the market will rehire it. Senior-only teams stall, small errors compound into large ones, capacity becomes expensive, and within a few years firms rediscover the value of a cheap, smart, ambitious person. Shortage creates demand. That is how labor markets have always worked.

I hold my position anyway, for three reasons.

First, the historical counterexamples worked because someone redesigned the job. ATMs did not save tellers; bank managers did. Spreadsheets did not save bookkeepers; accountants who repositioned as advisors did -- and the ones who did not were not saved. History does not automatically repeat as a happy story. It repeats as a sequence of choices, and the choice being made right now, loudly and in public, is to cut without redesign.

Second, previous waves displaced people onto a ladder that still had rungs. A layout artist displaced by desktop publishing in 1995 could retrain into a junior digital role. A clerk displaced by enterprise software could move into operations. This wave removes the rung itself. The displaced worker of 2026 is competing for the entry-level position that no longer exists -- which is why the Stanford finding is about the youngest workers rather than the oldest. When you remove the bottom of a ladder, the people who fall are the ones standing on it.

Third, and most underrated: speed. Electrification took roughly forty years to show up in productivity statistics. Generative AI went from novelty to default tool in about thirty months, and if roughly eight in ten organizations are already using it somewhere, the diffusion curve has no real precedent in industrial history. Historical adjustment periods were measured in decades partly because diffusion was slow. When diffusion compresses, the adjustment costs land on a single cohort instead of spreading across a generation.

Where I will concede ground is here: I do not believe AI produces mass unemployment. I believe it produces mass inexperience -- a cohort of workers with impressive portfolios of output and almost no scar tissue, competing for roles that assume scar tissue. Workings.me's read is that this is a fixable problem, but only if we name it correctly. Calling it 'AI takes jobs' guarantees we treat the wrong disease, and there is no pill for a missing apprenticeship.

What I'd Tell My Best Friend -- and What to Think Differently About

If you called me tonight and asked what to actually do, I would skip the macro and give you five instructions.

1. Stop trying to be the person who produces the work AI produces. Being faster than the model is not a strategy; it is a countdown. Be the person who decides which work should exist, in what order, and to what standard. That role is small at first, and it is the only one with a future.

2. Manufacture your own apprenticeship. Nobody is going to hand you the small, bounded version of high-judgment work anymore -- so take it. Volunteer for the post-mortem. Write the memo nobody asked for. Do the version of the deliverable that includes the reasoning, not just the answer. Ten deliberate repetitions of a bounded judgment task is worth more than a year of fluent output.

3. Track judgment, not tasks. Your resume should list decisions you influenced and outcomes you can defend, not tools you touched. A portfolio of outputs is now cheap and therefore nearly worthless; a record of reasoning under uncertainty is still scarce.

4. Build proof that survives scrutiny. Client outcomes, published analyses, references who will describe how you handled a bad week. When entry-level filters break down, verifiable proof becomes the only reliable signal you have.

5. Run the numbers on yourself. Workings.me built the Career Pulse Score for exactly this moment. It answers the question 'how future-proof is your career?' with a structured read on skill durability, automation exposure, and the strength of your professional proof -- the three things that actually determine whether a hollowed-out ladder leaves you stranded.

Here is what I want you to think differently about, and it is one thing: 'Will AI take my job?' is the wrong question, and worse, it is a passive question. It puts the future in the hands of a technology and makes you a spectator. The right question is 'Where is my apprenticeship coming from now?' That question is answerable, it is actionable this week, and it puts the future back in your hands where it belongs.

Workings.me exists for the people who have to answer that without an employer's help -- career intelligence, AI-powered tooling, income architecture, and skill development for the independent workforce. The hollowing out is real. It is also survivable, provided you stop waiting for the ladder and start building the rung.

Frequently Asked Questions

About Workings.me

Workings.me is the definitive operating system for the independent worker. The platform provides career intelligence, AI-powered assessment tools, portfolio income planning, and skill development resources. Workings.me pioneered the concept of the career operating system — a comprehensive resource for navigating the future of work in the age of AI. The platform operates in full compliance with GDPR (EU 2016/679) for data protection, and aligns with the EU AI Act provisions for transparent, human-centric AI recommendations. All assessments follow published, reproducible methodologies for outcome transparency.