The Apprenticeship Paradox
Researched by an agentic pipeline · reviewed and gated by the author
The biggest measured gains from AI showed up in exactly the people companies are most tempted to stop hiring.
The paradox
One of the most important findings in the economics of AI has received surprisingly little attention in boardrooms. In a study of more than 5,000 customer service workers, researchers at Stanford and MIT measured what happened when the workers were given an AI assistant. On average, productivity rose about 14%. But the average hides the real story. The newest, least experienced workers improved by around 34%. The most experienced workers barely improved at all.
Read that not as a technology statistic, but as a hiring statistic. The biggest measured gains from AI showed up in exactly the people companies are most tempted to stop hiring.
Because that is what many companies are now doing. The reasoning sounds sensible: AI can do a large share of junior work, so we need fewer junior people. The first half of that sentence is often true. The second half does not follow from it. In fact, it may have the economics exactly backwards.
What junior employees are actually for
Junior employees have always been a trade-off. They are cheap in salary but expensive in attention. They need supervision, they make mistakes, they take longer, and their work is inconsistent. Companies have always accepted that cost for one reason: junior people are not just labour. They are the “raw material” from which experienced people are made.
Every organisation runs an apprenticeship, whether it calls it one or not. A junior analyst becomes a senior analyst. A young engineer becomes an architect. A graduate salesperson becomes an account director. There was never any other way to produce experienced people, so the inefficiency was simply the price of having a future.
Then AI arrived, and something changed. The cost of inexperience started to fall. If a new employee with AI support can perform much closer to an experienced one, then AI hasn’t just made work cheaper. It has made producing experienced people cheaper. That is a completely different opportunity — and most companies are about to walk past it.
A junior job is not a pile of junior tasks
Here is where I think companies are making a category error.
Look at any entry-level role and you will find plenty of tasks AI can now do: research, first drafts, basic analysis, formatting, routine responses, document review. If those tasks are what the junior employee is for, the conclusion is obvious — automate them and hire fewer juniors.
But that confuses the tasks a junior person does with the purpose the junior role serves. Much of that routine work was never really about the output. A young lawyer reviewing hundreds of documents wasn’t just producing reviews — she was becoming a lawyer. A new salesperson losing deals and watching senior colleagues rescue them wasn’t just generating activity — he was acquiring judgment. The task produced an output. It also produced a future expert.
The right question is therefore not “can AI do this task?” It is: was this task only producing output, or was it also producing expertise? Those two answers lead to very different decisions.
The savings arrive now. The bill arrives later.
There is a second study worth knowing. The same Stanford researchers examined payroll data across the US economy and found that employment for workers aged 22 to 25 has fallen noticeably in the jobs most exposed to AI — but only where AI was used to replace their work. Where AI was used to help people do their work, young employment did not fall.
That distinction matters, because a hiring freeze on juniors has a very convenient shape for executives. The saving lands in this year’s budget. The damage arrives years later — the senior manager who doesn’t exist in four years, the technical leader who was never developed, the succession plan with no names on it. By then, the executives who made the decision may have moved on. The spreadsheet captures the saving perfectly and the loss not at all.
IBM saw this clearly. In 2026 it announced a major increase in entry-level hiring — including in technical roles everyone assumes AI will eliminate. Its reasoning was not sentimental. It was simple arithmetic: if you stop developing junior people now, who is your experienced workforce in five years? Companies cannot all buy senior talent from each other forever. Someone has to make it.
The honest objection
The strongest argument against all this deserves to be stated properly.
A sceptic could say: if AI lifts a novice 34% of the way toward expert performance, maybe AI itself is the new apprenticeship. Maybe people will learn faster with an AI coach than they ever did doing grunt work, and companies genuinely will need a smaller pipeline.
Part of that is probably right — and it actually strengthens the argument rather than defeating it. If AI makes people learn faster, then junior employees become a better investment than before, not a worse one: you get an experienced professional in less time, for less money. What the objection does correctly narrow is the shape of the pipeline. The answer is not to preserve today’s junior jobs. Some tasks should disappear. Some roles should shrink. The point is not nostalgia. The point is to redesign what “junior” means, rather than simply deleting it.
What redesigning it looks like
I would split early-career work into three buckets.
- Work AI should eliminate. Moving data between systems, formatting, routine summaries, administrative preparation. There is no virtue in making a young person spend two years doing what a machine now does better. Automate it. The goal was never to preserve tasks — it was to preserve learning.
- Work AI should accelerate. Give junior people the tools to perform closer to senior colleagues: analyse more cases, test more ideas, get instant feedback, draft something, critique it, and understand why the better version is better. The metric that matters stops being “how many analysts do we need?” and becomes “how fast can we turn a new analyst into someone with independent judgment?” AI can dramatically improve that number — if you point it there.
- Work humans must deliberately practise. This is the bucket most companies will forget. When AI removes the routine work, it also removes the accidental learning that came with it. So companies will have to create those experiences on purpose: real customer contact, decisions with incomplete information, defending a recommendation, handling the case the model gets wrong, taking responsibility for an outcome. These aren’t leftover tasks AI happens to be bad at. They are how senior judgment gets built.
Today, most companies are asking AI to remove friction from junior jobs. The better question is: which friction was actually the learning?
The choice
Every major technology removes work. AI will too, and it should. But there is a difference between removing unnecessary work and removing the machinery that produces expertise — and on a spreadsheet, the two look identical. A junior doing a task slowly looks inefficient. An AI doing it instantly looks superior. At the level of the task, that’s true. At the level of the company, it may be dangerously incomplete.
The companies that get this right will not protect outdated junior jobs. They will build new ones — with less drudgery, more customer contact, more real decisions, faster feedback, and a deliberately designed path from novice to judgment. The companies that get it wrong will book an immediate payroll saving and discover, a few years later, that something far more expensive has quietly disappeared: their pipeline.
The first generation of AI workforce strategy asked one question: which jobs can this technology do? The next generation needs to ask a harder one: what kind of people can this technology help us create?
Because AI may not make junior talent obsolete. It may make developing great people cheaper and faster than it has ever been. And if that is true, cutting entry-level hiring at the very moment AI arrived may turn out to be one of the most expensive savings a company ever booked.
Sources
- Brynjolfsson, Li & Raymond, Generative AI at Work, NBER Working Paper w31161 https://www.nber.org/papers/w31161 — Study of 5,179 customer support agents. Productivity rose 14% on average and roughly 34% for novice and low-skilled workers, with minimal gains for the most experienced.
- Brynjolfsson, Chandar & Chen, Canaries in the Coal Mine, Stanford Digital Economy Lab https://digitaleconomy.stanford.edu/publication/canaries-in-the-coal-mine-six-facts-about-the-recent-employment-effects-of-artificial-intelligence/ — Payroll data showing an employment decline of roughly 13% relative to trend for workers aged 22–25 in the most AI-exposed occupations, concentrated where AI automates rather than augments work.
- IBM to triple US entry-level hiring, February 2026 https://www.ibm.com/think/news/entry-level-roles-get-reset-ai — IBM’s expansion of entry-level hiring and redesign of early-career roles toward judgment work, framed explicitly as a pipeline decision.