We are automating the apprenticeship

A Harvard study came out recently that I think deserves a lot more attention than it got. In firms that started using AI, the hiring of junior people fell by 7.7% over six quarters, while senior hiring kept rising. Nobody was let go. We just stopped bringing younger people in. The study followed 62 million workers. It worries me, because it affects something that matters: how the next generation learns to do our work.

A few of the things I read last week:

• Saudi Arabia's SAMAI programme passed one million people trained, and the Cabinet named 2026 the Year of AI. (SPA)
• Eurostat counted one in five EU firms using AI in 2025, up from 13.5% a year earlier, and 55% among large firms. (Eurostat)
• Junior hiring at AI-adopting firms fell 7.7% over six quarters, against firms that did not adopt — a study of 62 million workers across 285,000 firms. (Lichtinger & Hosseini Massoum, Harvard)
• Early-career job postings are down about 35% since 2023, and 37% of organisations say they plan to replace early-career roles with AI. (Korn Ferry)
• Anthropic published a policy framework that plans, out loud, for AI pushing unemployment to 10% and beyond. (Anthropic)

Two of those are about how many people we are pushing in at the top. One is about the bottom slowly emptying out. It is the second one that worries me.

 The week in numbers.

The rung that makes our managers

We have become very good at counting the wrong things. Saudi Arabia has trained a million people. Europe can tell you, almost to the decimal, what share of its firms use AI. These are real achievements and I do not want to wave them away. But notice what we are not counting. In the firms that have actually brought AI in, the entry-level roles are quietly going unfilled. No round of layoffs, no announcement. The bottom of the ladder simply stops being built, one graduate we never hired at a time.

An entry-level job was never only about the work it produced. It was where judgement was formed. You do something by hand, someone more experienced pulls it apart, you do it again, and slowly you learn to tell when the work is good and when the machine is wrong in a confident voice. If we hand all of that to AI, we keep the output, but we quietly close the place where our future colleagues were meant to grow up. The cost does not show up this quarter. It shows up in a few years, when we look around for someone ready to take the next step and find we never gave anyone the chance to become ready.

I saw this happen with a team last year. They gave nearly all of the first-year work to AI, and for a while everyone was pleased with how much faster things moved. Then, maybe six months in, they noticed that nobody below the senior people was ready to take anything serious on. They had gained a lot of speed and, without quite meaning to, traded away their own succession. I do not think they were careless. It is just very easy not to see.

IBM changed its mind, and the timing is hard

It does not have to go this way, and IBM is the example I find encouraging. After a couple of years of automating hard, the company turned around in February and said it was tripling its entry-level hiring, including the software roles we are forever told AI can do. Its head of HR was very direct about it, calling them "all these jobs we're being told AI can do." To me that reads as a company that went a long way down one road, looked clearly at what it had given up, and chose to rebuild the bottom of the ladder on purpose. The point is not that AI cannot do junior work. Of course it can. The point is that an organisation is more than one quarter's output, and someone there remembered that in time.

There is a fair argument on the other side, and I want to give it its due. Junior work was often slow and expensive, so if AI can do it faster, trimming it looks like plain good sense. And for the task in front of you, it is. But we are not only buying efficiency. We are, a little at a time, disinvesting in the very people we were counting on to run things later.

The Digital Education Council and Google.org recently mapped this, and it is a useful way to see it. As AI rises, deep expertise becomes both more valuable and harder to build, because the entry-level work that used to build it is the first thing AI takes over. They call the space this opens an expertise gap. In the same study they notice something else: the work of judging AI's output, of deciding when to trust it, is sliding down from senior people towards those at the very start of their careers.

 The expertise gap, after the Digital Education Council × Google.org AI Skills Opportunity Map (2026). And the technology is not standing still. Ethan Mollick wrote this month that the friendly assistant, the human kept in the loop, was always a temporary stage, and that what is arriving now are agents that run hours of work on their own. Anthropic, which builds some of those agents, plans openly for unemployment reaching ten per cent and beyond. An agent that acts needs someone to decide where it is allowed to act, someone to check it, and someone to answer for it when it gets things wrong. That is judgement again. And we are quietly defunding the apprenticeship that grows judgement, at the very moment we are going to need much more of it.

If you work in L&D, HR, or transformation

I think this is squarely ours to pick up, and I am not sure enough of us see it that way yet. Put the junior pipeline up beside your adoption and licence numbers, and treat it as a measure of capability, because that is what it is. How many people did we bring in at the start this year? And are we giving them real work and real feedback, or are they sitting beside us watching AI do the very things they were meant to learn on?

If AI has taken the everyday work juniors used to grow through, then we have to put that learning back on purpose. Give them the genuinely hard problems. Ask someone experienced to mark their work properly, rather than just signing it off. Build the rotations and the moments of stretch that grow judgement instead of merely clearing the queue. None of this happens by itself, and no prompting course will do it for us. It is the slow, unshowy part of building capability, and it is the part that quietly compounds.

The provocation

So, the question I would leave you with is this. If we freeze junior hiring and let AI take the entry-level work, who is learning, right now, to make the difficult calls five years from now? And who will be ready to govern the agents we are about to switch on? If we cannot name those people, we are not really saving money. We are borrowing from our own future to make this quarter look a little better. Maybe the most useful thing any of us could do this week is to choose one early-career role and decide, on purpose, to protect it and help it grow.