The career pyramid is turning into a diamond: how to grow new experts and leaders in the age of AI
Anyone rolling out AI today hears the same question: "Why do we need juniors if the model does their work faster and cheaper?" Looking at this year's budget, it makes sense. Five years out, it turns into a different question: where will your seniors come from if nobody grew the juniors?
I think cutting juniors today is one of the most expensive mistakes companies make with AI. Below I'll explain why, and what to do about it: how to train young people when the model does the simple work, and how to grow leaders out of them.
Work doesn't disappear, but people pay for the transition
Technology doesn't remove work altogether, it changes it. Six out of ten of today's jobs didn't exist in 1940: David Autor and his MIT colleagues counted this from US census data (MIT News). And some jobs that were in demand back then are gone.
But this kind of shift doesn't happen in a day or a month. It can take years or decades. During the British industrial revolution, output per worker grew by 46% over sixty years, while real wages grew by only 12% (Robert Allen, "Engels' pause"). It worked out in the end, but a whole generation spent their working lives inside that pause.
And the bill for such a shift is paid by people nobody helps. So whoever implements or builds AI is responsible not only for the effect, but also for the move to the new paradigm. Above all, for moving the people.
The pyramid becomes a diamond
A classic career is shaped like a pyramid: many juniors at the bottom, fewer mid-level people above them, a handful of seniors at the top. Juniors do the simple work, learn on it, and some of them move up.
Now AI takes exactly that simple work. And the career pyramid turns into a diamond. Companies have fewer junior positions, but without early experience nobody can become a mid-level or senior specialist.
Each company on its own is better off not hiring beginners: it's cheaper to hire a ready specialist and give them AI. Brynjolfsson calls this a prisoner's dilemma, and I agree. If everyone does it, in a few years nobody will have ready specialists, and companies will start poaching the few who managed to grow. You saved on juniors, and now you buy seniors from your competitors, only at a higher price.
The first signs are already here. According to fresh Stanford data, employment of 22- to 25-year-olds in the occupations most exposed to AI is 19% below what you'd expect, up from a 15% gap a year earlier (Stanford Digital Economy Lab, August 2026). It's too early to panic: the authors call these descriptive patterns, and the Yale Budget Lab sees no broad disruption of the labor market so far. What matters more to me in the Stanford data is this: the decline shows up where AI replaces people, and where it supports them, employment is flat or growing. So it's not the technology that decides, it's how you use it.
I already described this risk for engineers in the article "The Engineer in Five Years": we risk winning three years of productivity and losing a generation of expertise. Now it's clear this goes well beyond engineers.

Who gains most from AI
In my experience, AI helps most those who combine a flexible, open mind with enough expertise to set the task and check the result.
A beginner has flexibility, but not the expertise to check the model. So they easily fall into an error trap. The model is confidently wrong, and they believe it, because they don't know themselves how it should be.
AI can make an expert three, five, even ten times stronger: they know what to ask and see at once where the model is wrong. But experts are often stiff. They aren't ready to trust it, and many are snobbish about it: "I've been doing this for twenty years, what can a machine tell me?" Then AI gives them almost nothing.
Mid-level people are the most interesting layer in this picture. They already have enough expertise to check the model on a typical task, and they haven't lost their flexibility. In my experience, they are the fastest to get a return from AI. But they face two risks. First, their experience was often built on the very typical tasks AI is now taking over. If they don't move into complex work, they get stuck: less routine, and not yet the depth of a senior. Second, mid-level people used to grow by starting to lead juniors. With fewer juniors, that step is gone too. So I'd give a mid-level person two things: complex tasks where AI works as their hands, and the role of mentor for the juniors you do have.
The common view is that AI helps beginners most. On typical tasks that's true, AI gives a speed boost there. In a call center with an AI assistant, novices' productivity rose by 34%, while the most experienced barely changed (QJE, 2025). Among BCG consultants, on tasks AI handles well, the weaker ones gained 43%, the stronger ones 17%, and together they worked about a quarter faster (Dell'Acqua et al., 2023).
On complex tasks the picture is different, and AI sometimes simply gets in the way. The same BCG experiment had a task deliberately built so the model would get it wrong. The spreadsheet looked complete, but the key details were hidden in interview notes, and they pointed to a different conclusion. Without AI, about 85% of consultants got it right. With AI, 60 to 70% did.
Why? Not for lack of experience: these were strong BCG consultants. The trap beginners usually fall into caught experienced people too. The authors write that the ones who lost tended to blindly adopt the model's output and question it less. The most telling part: the group that got a short briefing on how to work with AI before the experiment made more mistakes than the rest, 24 percentage points against 13. Faith in the technology beat their own experience. So expert snobbery has a point: sometimes AI really does lead you astray.
But the conclusion isn't "AI is no good for complex tasks". On a complex task, AI becomes the expert's hands. And the boost should come not in the number of tasks closed, but in the depth of the work: more research, more options considered, more hypotheses tested in the same time. Here a lot depends on who works with AI, how, and how they check the result. The same authors describe two working styles and call them centaurs and cyborgs. Centaurs split the work: this I do, this goes to the model. Cyborgs weave AI into every step and keep checking it against themselves. And one more finding: ideas from people who used AI were noticeably similar to each other. If you just ask the model for "some options", you'll get the same options as everyone else. You have to demand depth from the model, which means knowing what to demand.
So I'd ask not "who does AI help more" but what holds each of them back. A beginner lacks the expertise to check, and the five moves below deal with that. An expert lacks flexibility. What helps there is a task where AI is clearly faster, a colleague nearby who already works this way, and a manager who asks for work done with AI. A mid-level person lacks complex tasks and juniors of their own to lead. And the biggest winner is the pair: a flexible junior and an expert who stopped resisting.

How to grow new experts: five moves
Juniors used to grow on grunt work: they did simple things, made mistakes, and a senior corrected them. Now the model does the simple things, and the path that used to produce experts has to be built again. Juniors learn not on routine, but on checking what AI produced and on exceptions. Here are five moves I suggest to managers.
1. The junior checks the AI. They don't do the work themselves. They accept or reject the model's output and explain why. A good drill: a draft document from the model with mistakes hidden in it in advance, which the junior goes through with an acceptance checklist. You can only check it if you understand how it should be. That's the learning.
2. Exceptions, not the flow. AI takes the typical cases. The junior gets the unusual ones, with a senior nearby. In a few months they see what used to come up once a year.
3. A training mode without AI. For a while, the junior does the core operations by hand, to understand what exactly the model is checking. In aviation this has long been the norm: pilots are encouraged to fly manually on a regular basis, because constant autopilot weakens the ability to take over quickly (FAA, 2013).
4. Pairing with a senior or mid-level colleague on a real task. Training without support is an expense, not an investment. A course gives knowledge, but skill only appears on your own task with feedback. The mentor explains out loud why they made a decision, and AI frees up their time for it. For a mid-level person, this is also their own step up: they learn to lead people.
5. Rotation across the process. Growth used to go deeper into one operation. Now it goes wider: the junior goes through the whole process and learns to see the connections. The model will give depth on any single topic on request. Seeing the whole picture and asking the right question is still a human job.
There's one metric for the manager: how many months until a junior starts catching the model's mistakes on their own. If that number isn't going down, you haven't rebuilt training. You've just removed the step your future seniors climb.

How to grow leaders
With leaders it's the same story, only more visible. The road into management used to go through a first team: you were given two or three people, and on them you learned to set tasks, check work, and own the outcome. Now the first "team" shows up earlier, and it's made of agents.
Setting a task for an agent and accepting its output is the same management work, just on a small scale. You need to figure out what's actually needed, write the task with boundaries and criteria, then honestly check what came back, and not click "accept" without looking. Execution got cheaper, framing the task got more expensive. That's exactly what future leaders need to learn. I covered the environment in which an agent works without going off the rails in "The Agent Harness".
So I'd give young specialists tasks to set for AI as early as possible, but not without oversight. Let them write the task, get the output, defend it in front of a senior, and work out where the model went wrong and why they didn't notice. In a year you'll have someone who can set tasks and own the outcome. That's the foundation of a manager.
Markets differ
Where entry-level work is a commodity, the diamond is already forming. In the UK, 36% of employers in a Work Foundation survey said they had cut entry-level jobs, partly because of AI and automation (Reuters via Allwork.Space, August 2026).
But the question "what will they learn on" is the same everywhere. You'll hire a young process engineer or shift supervisor, and the simple tasks they used to learn on are already handled by an agent. A veteran's know-how, once it's inside a system, won't pass itself on to a junior.
What a manager should do
If you're rolling out AI, add two lines to the plan next to the savings line.
The first is the move of people. How many employees whose tasks went to a model or an agent are working on new tasks inside the company six months to a year later. With an owner and a deadline. If your plan only shows payroll savings, you're optimizing the budget, not transforming the company.
The second is the entry into the profession. How many junior positions you keep for the next two or three years and how you train them. "Working with AI" courses for everyone don't count. What counts is what I described above: checking AI, exceptions, manual mode, pairing with a senior or mid-level colleague, rotation.
And one metric for all of it: how many months until your junior catches the model's mistake on their own. Start measuring it now. In five years, it will decide who your seniors are.
Further reading
Dzhimsher Chelidze's book "Artificial Intelligence. Freefall" is about how to implement AI in a company and which people it takes: those who know the domain and can set a task for AI and check its result. Download the free edition.
Training and advisory on AI implementation: on the home page.


