If juniors stop getting hired, where do seniors come from?

Sources checked on 8 October 2026.

A hiring plan can look sensible one vacancy at a time. The team has an AI assistant. The assistant handles the first draft, the routine fix and the initial research. The remaining work needs someone experienced. So the next hire should be senior.

Repeat that decision across several years and a harder question appears: where will the next experienced person get their experience? The work being removed may have been unglamorous, but some of it was how a beginner learned what a competent result looked like.

Read the 19 per cent carefully

Stanford's August update puts employment for highly AI-exposed US workers aged 22 to 25 at 19% below the path implied by less-exposed peers, against a 15% gap in July 2025. Hiring, rather than separations, appears central. These ADP patterns are descriptive: education, prior trends and sample sensitivity complicate a causal reading. [1]

Saying employment fell behind a comparison group's path is different from saying one in five young workers lost a job to AI. It is also different from measuring the Saudi labour market. Neither of those claims follows from the chart.

Underneath the gap, the two groups moved in opposite directions. Between November 2022 and June 2026, employment of 22- to 25-year-olds in the two most exposed quintiles fell about 11%, while in the three least-exposed quintiles it grew about 10%. [1]

For a team deciding how to hire, the useful question does not require certainty about the entire economy. It requires looking at its own work: which tasks are disappearing, which people used to learn from them, and what will replace that learning?

The first draft was doing two jobs

Consider a junior engineer asked to investigate an intermittent bug. The visible output is a fix. Along the way, they have to reproduce the problem, decide which evidence to trust, discover a misleading assumption and explain why a change addresses the cause. A senior's review can then connect the proposed fix to consequences the junior has not seen yet.

Now imagine a tool produces a plausible patch immediately. That can be valuable. But if the junior simply submits it and the senior repairs the mistakes, the junior has skipped much of the work that would have made the next investigation easier. The patch has moved forward. The person's judgement may not have moved with it.

That risk comes from how the work is organised; it is no argument for preserving every slow task. Copying boilerplate for the fiftieth time is a poor training programme. The useful part is the part where the learner makes a decision, sees its consequences and receives feedback specific enough to change the next attempt.

Keep the learning inside the redesigned task

These are proposed assignments; no training trial has tested them. Their purpose is to keep a piece of consequential thinking with the learner while allowing the tool to remove work that does not add much learning. The right assignment depends on what the person already knows. A novice may need to construct an explanation that an experienced colleague can safely delegate.

The review has to change as well. If a senior only approves the final output, the learner can hide behind a good-looking result. Ask what they expected to happen, what evidence changed their mind and which alternative they rejected. Those questions make the reasoning inspectable without pretending every task needs an oral exam.

There is evidence for helping beginners too

The published Generative AI at Work study followed 5,172 customer-support agents. AI assistance raised issues resolved per hour by 15% on average; less experienced and lower-skilled workers improved both speed and quality. [2][3] That is a useful counterweight to the assumption that beginners necessarily have less value once a tool arrives.

It does not tell us that every coding assistant trains every junior engineer. The setting and the work matter. It does give a team a different hypothesis to test: perhaps the tool lets a beginner handle useful work sooner, provided the work includes feedback and responsibility they can understand.

A company can use the same capability to reduce entry-level hiring or to expand what an entry-level employee can do. Those choices may produce similar output this month and very different teams later. Calling both choices AI adoption hides the management decision.

Give apprenticeship a place in the plan

Start with one recurring task and a bounded area of ownership. For a new engineer, that might be a small service with a reviewer and a clear way to reverse a change. For a support employee, it might be a class of questions with explicit escalation rules. Give the person enough context to understand the outcome, not just a fragment that disappears into somebody else's queue.

Before using the assistant, ask for a short prediction or a proposed approach. Then let the person use the tool, check its work and discuss a specific decision with the reviewer. On a later task, change an important condition. If the person can adapt, there is more evidence of learning than another polished response would provide.

Make review time part of the staffing calculation. A junior who requires guidance has not failed to be senior; they were hired into a role that includes learning. If every experienced employee is booked at full delivery capacity, mentorship becomes work that exists only in the job advert.

Track things that expose progress: which categories of work the person can now own, whether the same misunderstanding keeps returning, and whether they recognise when to ask for help. Output volume still matters, but a count of completed tickets cannot tell you who is ready to take responsibility for the next difficult one.

There is no requirement to keep an old junior role unchanged. Some tasks should disappear. But if the hiring plan assumes a continuing supply of people with judgement, the organisation needs a credible account of how someone acquires it. Before removing the next entry-level position, write down where that practice will happen instead. An empty answer is a staffing problem worth seeing early.

Sources

• [1] Stanford Digital Economy Lab: Canaries, August 2026 update. 12 August 2026. ADP data through June 2026; the reported comparison is descriptive, not causal.

• [2] Brynjolfsson, Li and Raymond: Generative AI at Work. The Quarterly Journal of Economics, 2025, 140(2), 889 to 942. Published version: 5,172 support agents and a 15% average productivity increase.

• [3] Stanford: Generative AI at Work, publication summary. The authors' affiliated summary of the published study; distinguishes effects by worker experience.