What happens to the junior when AI does the tasks that used to train seniors?

AI can absorb part of the research, first drafts, simple tickets, standard analysis or lines of code that historically made up entry-level work. The biggest risk, though, isn't just these tasks disappearing. It's the disappearance of the learning path they used to make possible.

For HR leaders and managers, the question becomes less "how many juniors will we need tomorrow?" than: how do you build experienced professionals when experience no longer gets built the same way?

The answer isn't necessarily to artificially preserve the old tasks. Some data even suggests AI can speed up learning for less experienced people. But that means explicitly rebuilding what used to be picked up almost without thinking: learning to produce, watching mistakes happen, getting feedback, handling exceptions, understanding the client, and gradually taking on decisions.

In brief

• •A worrying signal exists on youth employment in roles heavily exposed to AI, but it would be an overreach to conclude junior roles are disappearing across the board. In the US, Stanford observes in June 2026 an employment gap of around 19% for 22-to-25-year-olds in highly exposed occupations, relative to the trajectory of same-age workers in less exposed ones — a descriptive finding, not a causal estimate, as the researchers themselves make clear.
• •The problem is especially interesting because AI performs well on codified knowledge — rules, procedures, documented know-how — while much of seniority rests on tacit knowledge built through experience, incidents, context and interactions.
• •But AI can also become a learning accelerator: in a study of more than 5,000 support agents, the gains were markedly larger for novice and lower-performing employees, and the authors find evidence the tool passes along some of the best agents' practices to them.
• •AI doesn't necessarily remove the need for juniors. It weakens the implicit curriculum that used to turn juniors into seniors. The challenge is replacing that curriculum with a far more intentional learning architecture.

The real subject isn't the junior task — it's what that task taught

A simple task carries two very different kinds of value. The first is its productive value: analysing twenty files, writing a first draft, handling a standard ticket, running a benchmark, preparing a SQL query, documenting a component. The second is its formative value.

By doing a fairly well-defined piece of work a hundred times, a junior gradually learns what's normal and what isn't. They run into variants. They make small mistakes. They see a senior's corrections. They discover what information is always missing from a client request. They understand why a solution that looks elegant on paper can be a bad one in the real context.

Part of that repetition was tedious. Another part produced experience. And AI doesn't naturally tell the two apart.

Several organisations observed document a shift of work from processing toward supervision, review and exception-handling: engineers becoming more like agent orchestrators, support staff becoming supervisors, more porous role boundaries. But this body of evidence remains a panel of thirty companies selected for their organisational signals, not representative of the market — an essential point when reasoning about juniors.

If a task gets automated because it's economically automatable, you can, at the same time, remove a step through which an employee used to learn their job. The gain shows up immediately in productivity metrics. The skill loss, on the other hand, may not show up for three more years.

There's now a signal on youth employment, but the quick conclusion should be resisted

The revision of Stanford's Canaries in the Coal Mine study, published on 12 August 2026, offers one of the most interesting data points available to date.

Using ADP payroll data in the United States, the researchers find that employment among 22-to-25-year-olds in occupations heavily exposed to AI is now around 19% lower than it would have been had it evolved like that of same-age workers in less exposed occupations. The gap stood at 15% with the data available in July 2025. The mechanism observed seems to come more from reduced hiring than from higher lay-offs.

The result is strong enough to be taken seriously. But not strong enough to write: "AI cut junior employment by 19%".

The authors more or less rule that out themselves. They describe their results as descriptive, state the study can't establish how much of the gap was caused by AI, and note some estimates get more sensitive when the specifications change. Nor do they find a massive employment shift across the economy as a whole.

It's also worth avoiding importing this result directly into a French scale-up: it aggregates very different American occupations and companies, and isolates neither startups nor any specific organisational model.

Other work also calls for caution against quick generalisations. A study conducted in Denmark observed rapid chatbot adoption and changes in tasks, without finding, two years after ChatGPT, any detectable effect on pay or hours worked at the scale studied.

What the data lets us say, then, is narrower: some young workers seem to feel the adjustments in heavily exposed roles earlier than their older peers. It doesn't yet let us declare the end of junior work.

The most useful distinction may be between codified knowledge and tacit knowledge

The new version of the Stanford study raises a particularly fertile idea for organisations. Young workers' difficulties show up more in roles heavily dependent on codified knowledge: what can be formalised, documented, taught, and looked up in digital systems. Conversely, experience seems to protect workers better in activities where tacit knowledge matters more — knowledge built through practice and repeated exposure to real situations.

This distinction sheds light on several transformations observed elsewhere. An agent can look up a procedure. It can produce a first support reply. It can scan a codebase, prepare an analysis, or propose a fix. These are precisely some of the tasks that let a new employee start contributing.

But what generally sets a senior apart isn't just their ability to do the harder version of the same task. It's knowing a client doesn't say exactly what they want. Recognising an incident that looks like three others but where one detail changes the diagnosis. Knowing when a rule can be bent and when it absolutely can't. Choosing between two imperfect solutions. Spotting an answer that's technically correct but organisationally unworkable.

That kind of knowledge doesn't write into a prompt easily. And, above all, it's hard to acquire without exposure to real situations.

The same worry exists in consulting: AI now quickly produces first-draft deliverables that experienced professionals know how to critique, while younger ones haven't necessarily yet built the reference points needed to spot the errors.

This is exactly where the skills-debt risk lies.

Skills debt: saving effort today, running short on expertise tomorrow

This is an organisational risk identified by inference, not an already-established statistical indicator.

An organisation accumulates this debt when it immediately captures AI's productivity gain, but stops producing the experiences through which its people develop the judgment future responsibilities will need.

The mechanism can be insidious. A team notices an agent does the first drafts correctly. So it asks less of its juniors. Then the agent produces a first review: the junior reads the result but works from scratch less often. Managers, now able to get more output from fewer people, spend less time on feedback.

The junior still looks very strong: their output is good and fast. Three years later, the organisation asks them to make the call on an exception the agent can't resolve. That's when everyone may discover assisted performance and acquired skill aren't the same thing.

An experiment run with 758 BCG consultants indirectly illustrates this distinction. On tasks sitting inside the tested model's capability "frontier", using GPT-4 sharply raised both speed and quality. But the study also shows this frontier is uneven: the tool can be very strong on one task and misleading on another.

Knowing how to produce with AI is one skill, then. Knowing when to stop trusting it is another.

But the reverse story also exists: AI can speed up a junior's learning

This is the counter-example that stops us concluding automation necessarily produces skills debt.

In Brynjolfsson, Li and Raymond's study of more than 5,000 support agents, the generative assistant raised productivity by around 15% on average in the study's published version, with especially large gains for the least experienced and lowest-performing workers. The authors also find evidence the tool spreads behaviours from the most effective agents and eases learning.

In the detailed working paper version, agents with two months' tenure using the assistant reached performance comparable to agents with no assistant who had more than six months' experience.

The result is striking. It's also very narrow: one support centre, one company, a tool built from existing conversations, repeatable tasks, and one specific productivity measure. It doesn't prove a software engineer, a recruiter or a Product Manager will learn four times faster thanks to an LLM.

But it usefully flips the question. AI can remove opportunities to learn. It can also make patterns previously locked inside top performers' experience accessible to juniors.

The outcome, then, depends heavily on how the work is designed. Used as a replacement for the junior, it potentially shrinks their learning ground. Used as a simulator, an explainer, a sparring partner, and a gateway to expert practices, it can widen it.

A handful of companies already show both sides of this transformation

Alan: enabling production, but organising the review

In its Everyone Can Build programme, PMs and designers can contribute directly to code. Alan has, alongside this, created the Engineering buddy role: an engineer supports the non-engineer on their contributions, and the final merge decision stays on the engineering side.

The case isn't specifically about juniors. But the mechanism is instructive: widening autonomy is paired with a human review infrastructure. That's exactly what's missing when people assume simply making an agent available is enough to train someone.

Finary: AI can expand the scope entrusted very early

Finary created a Design System Engineer role initially given to an apprentice, able to cover a very wide scope thanks to agents, with human validation before anything gets merged.

This case obviously doesn't prove every apprentice can replace a team. It suggests a different possibility: AI can hand a junior profile end-to-end responsibility earlier, provided the review protects the critical decisions.

Joko: ring-fencing learning time

Joko says it sets aside half a day per week per person for building AI skills. The data comes notably from the company's own communications and job postings: it documents an announced practice, then, not its economic return.

The organisational signal remains interesting: when tools change fast, learning can no longer just be what you do "when you have time".

bsport: automating doesn't force you to stop hiring juniors

According to an account gathered by TPC, bsport chooses to keep hiring junior profiles while heavily transforming the work of developers and tech leads with AI. Here again, this is a reported company choice, not proof of economic superiority.

These cases tell us less about "the future of the junior" than about a more interesting rule: there's no automatic consequence of AI on early careers. Organisations partly manufacture that consequence through their own staffing, review and training choices.

The new junior path: learning less through volume, more through feedback loops

For a long time, part of professional development followed a fairly simple logic: do a lot → make mistakes → get corrected → recognise the patterns → gain more autonomy.

If AI now handles part of the "do a lot", the rest of that loop needs to be preserved.

For a startup or a scale-up, it's useful to shift learning's centre of gravity toward four grounds: review, exceptions and incidents, client context, and gradually exercising decision-making.

The goal isn't for the junior to "master AI" after a year. It's for them to be able to produce with it, detect when it's wrong, explain a decision, and own a small real scope of responsibility.

A 12-month path to rebuild learning

This path doesn't mean artificially making juniors work "the old way". It means replacing mechanical repetition with a deliberate sequence of high-formative-value experiences.

The manager's role changes too: they no longer just hand out tasks, they hand out experiences

An organisation can sharply raise its production capacity with AI while shrinking its training capacity. That's the trap.

When the agent produces the first version, the senior can be tempted to simply approve the final result. But an approval click transmits almost nothing. The review needs to become an act of management.

Why this solution rather than another? What clue should have raised doubt? Which implicit assumption was wrong? Which decision could be delegated? Which one needed knowledge of the context? This takes time.

This is exactly the problem: AI lowers the cost of production, but it doesn't necessarily lower the cost of passing on expertise.

A useful distinction emerges from this: a manager's span of production — the volume of work they can oversee with agents — can grow far faster than their span of learning, meaning the number of people they can genuinely help grow.

An organisation that conflates the two can post remarkable productivity while eroding its pipeline of future seniors.

HR leaders should track something other than the number of AI training sessions

Counting licences, logins, prompts, or employees who completed training measures exposure to the tool. Not career-building.

To manage the skills-debt risk, you need to watch far more organisational variables: do juniors still see enough decisions? Do they have access to clients and users? Are they actually reviewing, or just consuming outputs? Do they take part in post-mortems? Can they explain why a generated answer is wrong? Do they have room to make reversible decisions? Is the time seniors spend on feedback shrinking as productivity rises?

These questions also have a consequence for hiring. AI fluency is probably becoming a baseline skill in many roles; Stripe, for instance, describes it as a criterion that's become important in its hiring.

But hiring solely the candidate who produces the best result with AI can be misleading. The most interesting evaluation looks at how they work with an imperfect system: do they know how to question the answer, identify what they don't know, ask for missing context, check an assumption, and justify what they decide to keep?

We'll probably need to test less for the ability to get an answer, then, and more for the ability not to get fooled by an answer that only looks good.

What the data does not yet let us claim

We don't know whether the signal observed among young American workers will grow, stabilise, or reverse. The Stanford study's authors say so explicitly.

Nor do we know whether companies hiring fewer juniors today are doing so mainly because of AI, the economic climate, profitability requirements, a different funding market, or a mix of these factors.

The Alan, Finary, Joko and bsport cases don't prove any specific model is superior. They only show that several organisational answers are possible.

Finally, no available study lets us properly quantify the amount of skills debt a company would accumulate by cutting its learning opportunities too far. This is an organisational-foresight hypothesis.

It still deserves to be taken seriously, because its asymmetry is strong: short-term savings are visible immediately; the shortage of judgment will show up late.

Self-diagnostic: are you automating away your pipeline of future seniors?

• •For every automated junior task, do we know which skill it used to help develop?
• •Do our juniors sometimes work out their own reasoning before consulting AI, to keep a personal baseline?
• •Are they genuinely reviewing, with a senior explaining the corrections, or just approving outputs?
• •Are they exposed earlier to client context, users and business constraints that never make it into the documentation?
• •Do they take part in the incidents, post-mortems and exception cases that build judgment?
• •Have we defined the reversible decisions a junior can gradually make on their own?
• •Is managers' coaching time protected and measured as their production capacity rises?
• •Do we evaluate AI fluency by the quality of judgment and verification, rather than by speed or tool mastery?
• •Can we show how a junior moves, over twelve months, from "producing with AI" to "controlling and owning a small scope with AI"?
• •If we sharply cut junior hiring today, do we know precisely where our seniors and managers will come from in three to five years?

The conviction

The junior doesn't necessarily disappear along with the junior tasks. But they won't become senior the same way anymore.

The old model let much of the learning emerge from sheer volume of work: lots of research, tickets, drafts, code, corrections and back-and-forth. That mechanism becomes less reliable once AI absorbs the volume.

The answer isn't to protect every repetitive task for pedagogical reasons. That would turn the organisation into a school for motions that have become pointless.

It's to deliberately design learning around what AI makes more important: verifying, understanding context, handling the exception, learning from the incident, talking with the client, and gradually deciding.

The organisation that pulls off this transition could even train some juniors faster than before. The one that doesn't see it coming can become highly productive for a few years, while quietly no longer producing its future seniors.