Part 5 of five, the finale. "AI, leadership and the human structures of work" is a series on what AI actually changes about leading people, and why those changes are choices, not inevitabilities.
Walk into a team that has quietly stopped hiring juniors and it looks fine. Better than fine: leaner, faster, all senior, nobody to train. It looks like that for years. It keeps looking like that right up until the seniors start to retire and there is no one behind them, and by then the problem took a decade to make and cannot be fixed in a quarter. This is the choice with the longest shadow in the whole series, and the one the industry is sleepwalking into.
We have argued that AI's costs, to wellbeing, to expertise, to how we manage, and to how we organise, are all leadership choices dressed up as inevitabilities. Here is the biggest one.
Across knowledge work, the entry-level door is closing. This is not a vibe; it is now measured. A Stanford study tracking payroll records for millions of workers found that those aged 22 to 25 in the most AI-exposed jobs have seen real declines: employment for young software developers is down roughly 20% since late 2022, with customer service among the other hardest-hit occupations, while employment for experienced workers in the same fields held up. The juniors are the canaries in the coal mine.
A separate line of evidence points the same way from entirely different data. PwC's 2026 AI Jobs Barometer analysed more than a billion job postings rather than payroll records, and found the entry-level rung has not so much vanished as risen out of reach: entry-level roles in highly AI-exposed occupations are now seven times more likely to demand skills that historically appeared later in a career, and 52% of newly appearing entry-level skills were ones previously associated with experienced workers, against just 7% in the least-exposed fields. Traditional entry-level positions shrank 10% since 2019. Two different methods, one direction, and the second tells you how the door is closing: not by removing the job title, but by loading it with prerequisites no beginner can have.
But the same study contains the fact that turns this from a lament into an argument. The decline showed up in jobs where AI automates the work. In jobs where AI augments the worker, entry-level employment did not fall. Same technology. Two deployment choices. Two completely different outcomes for a generation of workers. That is the whole thesis of this series, made visible in the labour data: automate or augment is a choice, and it is landing on the youngest first.
First, be honest about the cause
It is tempting, and self-serving for anyone senior, to pin the collapse of junior hiring entirely on AI. So let me steelman the sceptic: 2022 to 2025 also brought higher interest rates, a correction after pandemic over-hiring, and tighter budgets. Plenty of the entry-level squeeze is the ordinary business cycle, not the technology.
The Stanford authors ran exactly that challenge, which is why their finding is worth citing rather than the usual hand-waving. They checked interest rates and found that AI-exposed jobs are actually less sensitive to rates than average, so rate rises do not explain why those specific jobs shed their juniors. The AI-exposed decline holds up after controlling for the obvious economic factors. AI is not the only thing happening in the labour market. But on the best evidence we have, it is doing real, specific work here, and "it's just the economy" no longer covers it.
It is also genuinely contested, and worth saying so. The Financial Times recently argued that AI is not destroying entry-level work so much as changing it, pointing to US employers who expect to hire around 5.6% more new graduates this year. Both things can be true at once. Aggregate graduate hiring can rise while the specific, AI-exposed junior roles fall, which is exactly what the Stanford data shows: entry-level employment is climbing in the least AI-exposed jobs and dropping in the most exposed. The average hides the sorting.
The counter-argument, taken seriously
The sharper objection is the optimistic one: even if juniors are being automated, maybe that is fine. Maybe we simply need fewer people, and AI will conjure new entry-level roles we cannot yet see, as every technology wave eventually has.
Two problems. First, "new roles will appear" is a hope, not a plan, and it asks a specific cohort of real people to absorb the gap while we wait to be proven right. Second, and more concrete, the best evidence says automating juniors is not even the efficient move. In the Harvard and BCG study I cited earlier in this series, the people who improved most with AI were the least experienced; the floor rose faster than the ceiling. So the economically rational play is to pair a junior with AI and get near-senior output at junior cost, exactly the "augment" path the Stanford data shows protects entry-level jobs. Cutting the junior instead is not the efficient choice. It is the cheap-this-quarter one. Which makes it a choice, and a short-sighted one.
The pincer
Now put this next to something I planted earlier in the series and deliberately left unresolved: intuition rust. In a year-long study of cancer specialists using AI in clinical work, the researchers found early productivity gains quietly masking a dulling of the clinicians' own judgment. The expertise eroded without symptoms, until it was gone. That is a narrow population and I would not stretch it further than it goes, but the mechanism is not specialty-specific: skill atrophies when you stop exercising it, and AI is very good at letting you stop.
Hold both facts at once. You are cutting off the supply of new experts at the bottom, because AI can do junior work. And you are quietly de-skilling your existing experts at the top, because AI is doing their work too. No one coming in. The ones you have, rusting. That is a pincer on expertise, and it closes slowly enough that no single quarter's numbers ever show it. There are now names for both jaws of it: "intuition rust" for the experts quietly losing their edge, and "never skilling" for the juniors who never build one, because the work that would have built it was automated before they arrived.
There is a farming phrase for exactly this: eating your seed corn. When times are hard you can eat the grain you saved to plant next year. It feeds you now. It guarantees there is nothing to harvest later. Automating the juniors while offloading the seniors' judgment to the machine is eating the seed corn of your own profession, and calling the fuller belly a productivity gain.
Here is the question the pincer forces, and almost nobody in the "automate the juniors" conversation is asking it. Where do the seniors of 2035 come from? The senior you are hiring today became senior by doing, badly and then well, the junior work you are now automating away. Remove the rung and you do not just lose this year's juniors. You lose the mechanism that produces every future senior. And the people you are counting on to supervise the AI, whose judgment is the last line against a confidently wrong machine, are the exact people quietly losing that judgment.
What this asks of a leader
Treat your junior pipeline as capability infrastructure, not a cost line. Pair juniors with AI rather than replacing them, the augment path, not because it is kind but because the data says it is the smarter economics and the only way to keep making experts. Protect the deliberate, inefficient practice, in juniors and seniors both, that builds the judgment no model has. Some slowness is not waste. It is how expertise is made, and it is the first thing an efficiency drive deletes.
That is the whole series in one idea. AI hands you options, not outcomes. It can lift your weakest people or hollow out your strongest. It can make work feel owned or make it feel like babysitting a machine. It can build an enabling organisation or a queue. It can grow the next generation of experts or eat the seed that would have become them. Every one of those is decided by a person, not the technology. The costs are real, the evidence is clear, and the choice, every time, is yours.
The series: AI, leadership and the human structures of work
- The psychological cost of AI is a leadership choice, not a technology outcome
- Who's the authority now? Leading in the age of the jagged generalist
- Managing a team of agents: leadership when roles become software
- Org design for AI: why your Centre of Excellence becomes a bottleneck
- Cutting juniors is a choice, not an AI inevitability
You are reading part 5, the finale. Links added as each publishes.
Written by Richard Atkins.
Sources: Brynjolfsson, Chandar & Chen, "Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of AI", Stanford Digital Economy Lab (ADP payroll microdata; the overall relative decline for ages 22–25 in the most AI-exposed occupations was revised from 13% to 16% in the February 2026 update) — https://digitaleconomy.stanford.edu/publication/canaries-in-the-coal-mine-six-facts-about-the-recent-employment-effects-of-artificial-intelligence/ ; the authors' interest-rates follow-up — https://digitaleconomy.stanford.edu/news/canaries-interest-rates-and-timinga-more-on-recent-drivers-of-employment-changes-for-young-workers/ · PwC, 2026 Global AI Jobs Barometer (analysis of 1bn+ job postings) — https://www.pwc.com/gx/en/news-room/press-releases/2026/pwc-2026-ai-jobs-barometer.html · Dell'Acqua et al. (2023), Harvard/BCG jagged-frontier study — https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4573321 · Ehsan et al. (2026), "From Future of Work to Future of Workers", the "intuition rust" study (year-long study of cancer specialists; CHI 2026) — https://arxiv.org/abs/2601.21920 · "never-skilling": Ke et al. (2026), "AI-induced never-skilling in medical education", Nature Medicine 32(6) — https://www.nature.com/articles/s41591-026-04438-y · Class of 2026 graduate hiring projection (+5.6%): NACE Job Outlook Spring Update, April 2026 — https://www.naceweb.org/job-market/trends-and-predictions/ ; "AI isn't destroying entry-level jobs. It's changing them", Financial Times (2026).
Top comments (0)