There is a version of career development that we have always taken for granted. You start somewhere junior. You do work that is real but relatively low-stakes. You make mistakes that sting but do not sink the company. You learn, over several years, how to exercise judgement about people, about decisions, about what matters and what does not. And then, gradually, more is trusted to you.

That pipeline is now under serious pressure. Not because companies are failing at talent development. Because AI is doing an increasing share of what juniors used to do: the research, the first drafts, the data analysis, the synthesis, the initial thinking on a problem. The output improves. The efficiency gains are real. And the formative experience that used to come with that work quietly disappears.

This is a problem we are not taking seriously enough, and the consequences will not be obvious for another decade, which is exactly when they will become very expensive.

I have been thinking about this through the lens of my own two children, and the contrast between them makes the problem concrete in a way that abstract arguments do not.

My son is nineteen and in his second year studying biomedical science. For him, AI looks like an advantage. It accelerates research, surfaces patterns in data that would take years to find manually, and augments the kind of deep analytical work his field demands. The expertise he builds will sit on top of AI capability, not in competition with it. He will still need clinical experience, research judgement, the ability to ask the right question before reaching for the tool. But the tool is genuinely on his side.

My daughter is seventeen, finishing school, and wants to study law. That is a different conversation. Legal research, contract review, first-draft document work: the things junior lawyers have always spent their early years doing. These are exactly what AI now handles with increasing competence. The junior roles that used to exist as the training ground for the profession are shrinking. Which raises a question I do not have a comfortable answer to: how does she build the judgement and experience that law requires, if the work that used to develop it is no longer there for her to do?

That question is not unique to law. It is the question facing any profession where AI is absorbing the entry-level work that was never really about output. It was about formation.

What entry-level work was actually for

If you look back at a graduate role, a junior analyst position, an associate-level anything, the job was never really about the output. The output was fine. Often it needed significant rework before it was useful. The real value was in the formation happening underneath.

Doing a first draft and having it pulled apart teaches you something that no feedback session can replicate. Sitting in a client meeting where things go sideways and watching a senior colleague navigate it in real time teaches you something a case study cannot. Getting a decision wrong, owning it, and figuring out how to recover from it teaches you something that no amount of observation will.

Wisdom is built through consequence. You have to care about the outcome. You have to feel the weight of it. You have to be the one who has to live with what happens next. That is what makes experience formative rather than just informative.

Graduate roles put people in that position early. Low enough stakes that mistakes were recoverable. High enough stakes that they were real. That combination was never incidental. It was the design.

What gets lost when AI absorbs that work

When AI does the first draft, the junior reviews it. When AI does the analysis, the junior checks it. When AI maps the options, the junior presents them upwards. That looks like development. It is not. Reviewing someone else’s work is not the same as producing it. The struggle is where the learning lives.

There is also something more subtle happening. The judgement you develop as a leader is not abstract. It is built on a foundation of very concrete, very specific experiences of being in situations and having to navigate them. The more of that experience you have, the better your pattern recognition, your instincts, your read of people and situations.

If the next generation spends their early career supervising AI outputs rather than doing the work, they will arrive at mid-level roles with strong tool skills and thin foundations. They will be excellent at prompting, at reviewing, at orchestrating. They will be less practiced at the harder things: sitting with a genuinely ambiguous problem, making a call without enough information, holding a difficult conversation that has no good outcome.

Those are the things leadership actually requires. And they are not learned by proxy.

The code review problem

Software development is probably the most immediate and visible version of this. AI writes code now. It writes a lot of it, quickly, and much of it looks correct. The question nobody is asking loudly enough is: who reviews it, and do they actually know what they are looking for?

Reviewing code well is not a beginner skill. It is something you develop by writing code, writing bad code, shipping something that breaks, debugging things that should not have broken, and slowly building up the pattern recognition to know what trouble looks like before it becomes trouble. Security vulnerabilities, performance problems, architectural decisions that work fine at low volume and collapse under load: these do not announce themselves. They hide in code that passes a cursory read. Finding them requires the kind of experience that only comes from years of having made those mistakes yourself.

The junior developer who has spent their first two years prompting AI and reviewing its output has not built that foundation. They may not know what they are missing, and that is the dangerous part. The result is AI-generated code reviewed by people without the experience to catch what is wrong with it, approved, shipped, and quietly accumulated across the systems that underpin everything. The version of “AI slop” that applies to software has sharper consequences than bad writing: systems that are brittle, insecure, or inefficient in ways that are very hard to unpick once they are embedded.

The developers who will be genuinely valuable in ten years are the ones who understand what is happening underneath the code they are working with. Not just what it does, but why it is structured the way it is, where it is likely to fail, and what a subtly wrong implementation looks like compared to a correct one. That understanding does not come from reviewing AI output. It comes from building things from scratch, getting them wrong, and understanding why.

The incentive problem

Here is what makes this genuinely difficult. Companies are not failing to develop talent out of negligence. They are optimising rationally for the wrong time horizon.

An organisation that uses AI to handle graduate-level work gets a real and immediate benefit: more output, lower cost, faster delivery. The cost of that choice, a leadership pipeline that is thinner in ten years, is invisible right now and will land on a different leadership team at a different moment. The incentive to invest in formative development at the expense of short-term efficiency is very weak.

This is the same structural problem that shows up in most underinvestment in people. The returns are long, the costs are near, and the connection between the investment and the outcome is hard to draw a straight line through.

The organisations that get this right will be the ones where senior leaders feel personally accountable for who comes after them, not just for what gets delivered this quarter.

What deliberate development looks like now

The answer is not to stop using AI. That would be like telling a previous generation to stop using spreadsheets so the juniors would have to do more arithmetic. The tool is here. The question is what you build around it.

The organisations that will develop strong leaders in this environment are the ones that treat human development as a design constraint, not an afterthought. That means making deliberate choices about what work stays with people, even when it would be faster or cheaper to hand it to a system.

It means putting junior people in rooms where things are genuinely uncertain and letting them contribute, not just observe. It means creating accountability at a level that is real enough to matter, which means sometimes letting people fail at things that have actual consequences. Supervised failure with debrief is one of the most efficient development tools that exists. It is also deeply unfashionable because it looks like inefficiency from the outside.

It means mentorship that is active rather than nominal. Not “I am available if you need me” but “I am watching how you handle this and we are going to talk about it afterwards.” That is time-intensive. It does not scale. It is also how people have always learned to lead.

And it means being honest with early-career people about what they are building. Not just skills. Judgement. The capacity to navigate situations that are hard in ways that cannot be fully anticipated. That capacity takes years to develop and there are no shortcuts, even sophisticated ones.

The question for the next decade

We are at an early enough stage in this shift that the problem is still preventable, or at least manageable. The generation now entering the workforce will be the test case. In ten years we will be able to see clearly whether they were developed or just deployed.

None of this is an argument against AI. The tools are extraordinary. The people who learn to use them well, and who also have strong foundations underneath, will be more capable than any previous generation of practitioners and leaders. That combination is genuinely exciting. The concern is not the technology. It is the assumption that the human development piece will take care of itself while organisations chase the efficiency gains.

It will not take care of itself. Wisdom cannot be downloaded. It has never been able to be. The question is whether we are building the conditions for the next generation to acquire it alongside the tools, or whether we are quietly assuming that one replaces the other.

It does not. And the sooner we treat that as an operational problem rather than a philosophical one, the better.