The Entry-Level Job May Survive. The Learning May Not.

Companies are still hiring junior workers. The harder question is whether AI is removing the work that once taught them how to become experts.

Image generated by the author

Ford thought it could formalize the problem further.

The company had design requirements, automated quality systems, and AI tools meant to catch quality issues earlier. Feed the right specifications into the machinery, the logic went, and better vehicles should come out the other end.

But engineering knowledge does not live only in requirements documents.

It lives in the memory of someone who has watched the same component fail across three product cycles. It lives in the engineer who notices a measurement that reads fine but feels wrong. It lives in the technician who already knows which small defect turns into a recall once heat, vibration, weather, and real customers get involved.

Ford eventually concluded that too much of that knowledge had left the building. The company hired, promoted, or brought back more than 350 experienced engineers to rebuild that layer of expertise. According to Ford executives, those engineers now help guide younger engineers, improve data collection, and strengthen the AI training behind Ford’s automated systems. Ford did not abandon AI. It put human experience back around it, as The Verge reported.

The machines needed training.

So did the next generation of humans.

That second problem is much easier to miss.

The public argument about AI and entry-level work usually starts with a number: how many junior jobs will disappear.

It is probably the wrong number.

The reassuring answer to the wrong question

A recent Ramp–Revelio Labs working paper offers a surprisingly upbeat picture. The researchers connected AI spending to workforce records across more than 21,000 U.S. firms. Among companies making the largest AI investments, total headcount grew about 10% over the two years following adoption. Entry-level headcount grew 12%.

That does not look like an entry-level extinction event.

It is not proof that AI created those jobs, either. Ramp’s own writeup says the gains were concentrated among high-intensity adopters, while low-intensity adopters saw no statistically significant change. It also notes that AI adopters were already larger, more engineering-intensive, more likely to be venture-backed, and faster-growing than non-adopters.

Still, the finding matters.

It complicates the cleanest version of the jobs-apocalypse story: companies can pour money into AI and keep expanding their junior workforce at the same time.

The entry-level job may survive.

A headcount just cannot tell you what happens inside it. It cannot tell you whether a new employee attempts the work or simply receives the result, whether anyone explains why an answer failed, or whether the beginner sees enough ordinary cases to recognize the strange one.

A company can keep the junior title on the org chart while quietly removing the path that used to lead from junior to senior.

The missing organizational object

Call that path the training ladder.

Nobody draws it on an org chart. It is not a formal curriculum. It is the sequence through which a beginner watches, attempts, fails, gets corrected, compares approaches, handles harder cases, and eventually acts without supervision — usually buried inside work that looks, on paper, like ordinary output.

A junior accountant learns what a clean reconciliation looks like by wading through hundreds of messy ones. A young lawyer discovers which clause actually matters because a senior partner hands back a draft bleeding red ink. An engineer learns to distrust a spec sheet after watching a technically compliant part fail under real-world conditions.

None of that makes every repetitive assignment sacred.

Plenty of junior work is tedious simply because junior labor is cheap. Some assignments teach almost nothing, and some professional traditions have long confused exhaustion with development. Automation should absolutely eliminate work that burns hours without building capability.

The danger shows up when an organization removes the assignment and assumes the learning attached to it will vanish just as harmlessly.

Production and learning used to happen in the same place. The employee finished the work, and the work changed the employee.

AI can preserve the first outcome while quietly erasing the second.

Sometimes the AI really does teach the beginner

That outcome is not inevitable — and one of the strongest workplace studies of generative AI points the other way entirely.

A peer-reviewed Quarterly Journal of Economics study examined the staggered deployment of a generative AI assistant across 5,172 customer-support agents. Access to the assistant increased productivity by 15% on average, and the gains accrued disproportionately to less-experienced and lower-skilled workers. The authors also found evidence that AI assistance facilitated worker learning and helped disseminate behaviors associated with more productive agents.

This is not a footnote to the apprenticeship argument.

It is the argument’s necessary counterweight.

AI can distribute expertise that used to be locked inside a handful of people’s heads. It can deliver guidance in the exact moment a worker needs it, drawing on patterns from thousands of prior conversations instead of whichever supervisor happens to be free.

In this case, the technology did not take the conversation away from the agent. The agent still faced the customer, still had to interpret the situation, and still had to choose a response. The AI supplied support inside the task instead of removing the worker from it.

That distinction is the whole essay in miniature.

The danger was never assistance.

It is assistance designed only to finish the job.

The polished answer creates a new kind of beginner

The problem gets sharper once the AI starts producing work that looks more competent than the person reviewing it.

A peer-reviewed Organization Science experiment involving 758 Boston Consulting Group consultants tested AI assistance across realistic knowledge-work tasks. The study found that AI improved performance for some tasks and worsened it for others, even inside the same knowledge workflow. Its authors call this the “jagged technology frontier.”

On 18 tasks inside the model’s capability frontier, consultants using AI completed 12.2% more tasks, worked 25.1% faster, and produced higher-quality results. But on a task selected to sit outside that frontier, consultants using AI were less likely to reach the correct conclusion than consultants working without it, according to the Harvard Business School summary.

The study tested only one outside-the-frontier task, which limits how far that specific failure travels. But the failure points at something bigger: AI capability is jagged. Two assignments that look equally difficult to a professional can sit on opposite sides of a model’s competence boundary, with nothing obvious on the surface to tell you which side you are on.

That creates a strange new burden for junior workers.

They are expected to review systems that are often more articulate, more widely read, and faster than they are — systems that can also fail in ways nearly impossible to predict from the surface of the task.

Reliable review takes more than reading the answer carefully. It takes knowing which assumptions deserve suspicion, recognizing the exception hiding inside an ordinary-looking request, and having enough accumulated exposure to mistakes to notice when a persuasive recommendation quietly violates a constraint the model never saw.

Those instincts are usually the product of an apprenticeship, not something a beginner arrives with.

Here is the trap: organizations remove the early work because AI can now perform it, then ask inexperienced employees to supervise the output using judgment that the early work used to build.

Review becomes the first assignment instead of the eventual result of one.

The work moves upward before the worker is ready

Accounting shows how this shift can happen without anyone deliberately dismantling a profession.

Recent accounting research found that generative AI adoption was associated with productivity gains and changes in how accounting work was allocated, with less emphasis on routine data-entry work and more emphasis on higher-value tasks such as communication and quality assurance. The same research also warns that human judgment remains necessary when evaluating AI output.

On its face, that is good news.

Less data entry can mean more time for client conversations, exception handling, and the kind of analysis that actually uses professional judgment. Nobody should defend hours of copying numbers just because a previous generation of accountants endured them.

But moving the worker upward is not the same as preparing them for what is up there.

Routine transaction work exposed people to the raw material of the profession: normal patterns, common errors, odd classifications, reconciliations that do not balance, the whole path from an individual record to a finished report.

Not every hour of that exposure becomes expertise. Removing the contact still changes what a worker has seen before they are handed quality-assurance responsibilities.

The question was never whether data entry deserves protection.

It is whether its developmental value — where that value actually existed — was identified before the task disappeared, and replaced with something else.

Skip that step, and an organization gains a faster reporting process while quietly producing fewer people who understand how the report got built.

Save the learning function, not the task

Learning science has a better vocabulary for this problem than “people need to struggle.”

Struggle by itself teaches nothing. Confusion can just stay confusion. Repetition can curdle into numbness. A disengaged supervisor can let the same mistake calcify into habit.

What matters is the design wrapped around the attempt.

Cognitive apprenticeship describes expertise forming through modeling, coaching, scaffolding, reflection, and progressively independent work. The expert’s reasoning has to become visible. The learner needs a real chance to act. Support should be available — but not permanent.

Research on productive failure backs this up from a different angle. A meta-analysis of 53 studies found an advantage for designs where learners attempted a suitable problem before receiving instruction, especially when that attempt was followed by careful consolidation and feedback. That evidence comes from classrooms, not workplaces, but it points to the mechanism: a genuine attempt can prepare the mind to actually absorb the explanation that follows.

Generative AI can break that sequence simply by arriving too early.

A 2026 preprint involving 1,222 participants found that AI assistance improved immediate performance across several tasks but was followed by weaker unaided performance and lower persistence once the assistance was removed. It is a preprint, and it does not study professional careers, so it cannot prove workplace deskilling — but it does show why completed output and acquired capability need to be measured as two separate things.

Peer-reviewed research in high-school mathematics adds a sharper warning. Students given an unrestricted GPT-style assistant performed better during practice and worse once the tool was removed. A more constrained tutor, built with pedagogical guardrails, largely reduced that damage, according to a PNAS study.

The point is not that employees should be treated like students.

It is that timing, constraint, and task design change what assistance actually does to the person receiving it.

The same model can act as a coach or as a replacement.

The difference is not only in the model.

It is in how the work around it is built.

AI as coach. AI as replacement.

AI acts like a coach when it keeps the learner inside the work: asking for an attempt before offering a solution, giving a hint instead of finishing the assignment, explaining why an approach failed, generating a slightly harder case, asking the worker to defend the final call — then stepping back.

AI acts like a replacement when it produces the deliverable immediately, hides the intermediate work, corrects errors before the beginner even encounters them, and stays present for every future attempt, indefinitely.

An employee can become extremely effective inside that second system.

Effectiveness with permanent assistance is not the same thing as independent competence, and the gap between them tends to stay invisible until the assistance disappears.

This is not an argument for tool abstinence. Accountants do not need to prove themselves by abandoning spreadsheets, and engineers do not need to discard simulation software to stay sharp.

The real test is whether someone understands the work deeply enough to catch an invalid input, a broken assumption, an exceptional case, or a wrong output — and whether they can keep going when the tool fails, and the problem needs to be escalated instead of trusted.

A productivity dashboard cannot see any of that.

Some professions are trying to rebuild the ladder

Law is one of the few industries visibly experimenting with replacements for the training experiences AI may be changing.

Working with Stanford Law School’s liftlab, the firm Vorys developed AI personas modeled on 19 of its partners. Associates can question them, pressure-test documents against them, and get editing feedback before a draft reaches a human partner’s desk. The same Stanford lab is also exploring simulations that let associates practice transactions or depositions before they face the real thing, as Reuters reported.

These are early pilots, not proof of long-term professional development. The personas can hallucinate. Synthetic feedback is not the same as a partner putting their name behind a junior lawyer’s growth — nobody at the firm is accountable for what an AI persona teaches badly in the way a human supervisor is accountable for a real trainee.

Ropes & Gray is experimenting from another angle: protected time. Reuters reported that first-year associates can devote nearly 400 hours of their annual billing requirements to AI training and simulations, up to 20% of their required billable hours. The firm described the program as a deliberate investment in early-career lawyers’ AI skills.

One experiment tries to scale expert feedback.

The other protects time to learn.

Neither proves the profession has solved the problem. Together, they show the shape of the response.

The strongest AI-era training system would not try to preserve every old assignment. It would preserve the sequence that makes responsibility possible: attempt the problem first, use AI for feedback or comparison, require the worker to explain the final choice, expose them to failure cases and not just standard ones, raise the difficulty gradually, and withdraw support on selected assignments to test whether the worker can diagnose an error without being told where it is.

Keep human experts involved at the points where judgment, accountability, and consequences turn real.

And measure progression over time — not just whether junior employees complete more work with AI, but whether they are getting harder to fool.

The organization must produce more than work

Ford’s experienced specialists do not prove that humans beat AI.

What they prove is less theatrical and more consequential.

A production system runs on knowledge that may never have been written down cleanly enough to automate — knowledge held by people who have watched failures repeat under different names. Unless that knowledge moves into younger workers, the company will eventually face the same shortage again, wearing a different label.

AI can help with that transfer.

It can make expert patterns available to more people. It can build simulations that would once have been too expensive to run. It can hand a beginner feedback at the exact moment a human supervisor is not in the room.

That transfer does not happen on its own.

A workflow optimized for today’s output can quietly consume the experiences required to build tomorrow’s judgment — and a company can watch it happen without ever seeing the invoice.

That is why entry-level hiring is not enough.

The junior worker can stay on the payroll. The productivity chart can keep climbing. The deliverables can keep arriving early.

And the training ladder can still be missing, one rung at a time, with nobody in the room noticing which rung is gone.

When an experienced engineer walks into a design review, the immediate goal is to catch the defect before it reaches production.

But something else is supposed to be happening in that same room.

Someone less experienced is supposed to be learning why the defect looked harmless, which pattern exposed it, and when to stop trusting the system that had already approved it.

The real test of an AI workflow was never whether it produces the right answer today.

It is whether anyone in that room is still learning how to recognize the wrong one tomorrow.