You Only Want Seniors. Who Should Train Them?

No company should keep a pointless task alive just because, twenty years ago, someone learned the trade by doing it. If an assistant can write a reliable first draft in a few minutes, making a new hire spend the afternoon on the same draft doesn’t become a good idea because we call it training. It may just be how we justify our habits.

I’ll say this up front, because this is not a defence of paying your dues. Nobody entering a profession today owes us a repeat of our inefficiencies. A programmer doesn’t need to know every function of a library by heart, and an analyst’s worth isn’t measured in hours spent tidying spreadsheet cells. Some of what we called dues was dull work handed to whoever was least able to say no, and calling it school never made it one. If AI takes it, I won’t miss it.

The problem starts afterwards. Once we’ve decided what no longer needs doing, we tend to treat the question of what still needs learning as closed too.

Take a company that hands a growing share of its work to AI and concentrates oversight in a few experienced people. It may work very well, and there’s no need to invent a disaster to question it. You only have to ask what the person joining that company today will do to become, one day, capable of exercising that oversight. “They’ll use AI” is not an answer. AI is the tool they’ll have. It isn’t the path by which they’ll learn when to trust it.

I’m talking about a specific kind of organisation, not all of them: the one that cuts the ways in, and the chances to learn, while still running on the judgement of experienced professionals. In its plans, experience shows up as something available from outside. Today we need people who are already capable, tomorrow we’ll need others, and somebody will have trained them.

Even the Engine Maker Makes People Practise

On October 2, Anthropic announced Claude Frontier Academy: $100 million to train 10,000 engineers by the end of 2027. It isn’t a programme for beginners. Companies nominate their best engineers, who spend a few intensive days with Anthropic’s engineers, take a practical test on a scenario they haven’t seen, and then spend twelve weeks leading a real project inside their own organisation. At the end they are assessed again.

I don’t read it as a confession, along the lines of “AI doesn’t work without humans”, and I don’t read it as philanthropy either. It’s a commercial investment: whoever sells these systems needs people who can adopt them well, and is building them. What interests me is the method. Even with professionals who already know their trade, the company that makes the model isn’t content to hand out access to the tool. It offers practice on real cases and a check on what people can actually do.

If using these systems well takes a path even for people with years of experience, the question for everyone else is simple. Where is that path in your plan?

The Work Delivered and the Person Who Did It

A task given to someone who is still learning produces two things. One is the delivered work, which has a client and a deadline. The other is what that person becomes able to do, and that has neither client nor deadline: it exists only if somebody decides to invest in it.

AI can improve both. It can also improve only one of them, and looking at the delivery won’t tell you which.

A study by Judy Hanwen Shen and Alex Tamkin, released in January and presented by Anthropic itself, helps make the difference visible. Fifty-two developers, mostly early in their careers, had to work with Trio, a Python library none of them had used before. Some could get help from an assistant, the others couldn’t. On the comprehension quiz that followed, the assisted group averaged 50% and the group without AI 67%, with the widest gap on debugging questions. The time saved, on the other hand, was not statistically significant. It should be read for what it is: seventeen points less understanding, measured right after that task. Not 17% of someone’s skills gone. The sample is small, the task narrow, the measurement immediate, and it says nothing about years of work or about skills people already have. The authors also note that participants who used the assistant to get things explained, rather than to get them done, did better. But nobody was randomly assigned to a way of using it, so there’s no recipe here of the “just ask for explanations and the problem goes away” kind.

Another piece of work points the other way, and deserves the same space. Erik Brynjolfsson, Danielle Li and Lindsey Raymond studied the arrival of an assistant in a customer support operation with 5,172 agents, in a paper published in 2025 in the Quarterly Journal of Economics. Issues resolved per hour rose by about 15% on average, and by about 30% among the least experienced, while the most experienced barely moved. Looking at periods when the system was down, the authors find signs consistent with learning that persists without the assistant. They warn that those outages are rare and not necessarily random.

Put side by side, the two studies don’t say that AI makes people worse at their jobs, or that it trains them. They say it can do either, depending on the work and on how it’s used. So the question for anyone running a team isn’t whether AI makes people better. It’s this: when someone gets more work done thanks to AI, what is your belief that they’ve also become more independent actually based on?

An Export That Takes a Few Minutes

Here is an invented example, but an ordinary one. A company needs to add a document export feature to its business management software. The task goes to someone who joined recently and does it with an assistant. The AI produces a solution, the planned tests pass, a senior colleague reviews it and approves it. The task has been completed correctly.

Now let me change the question. A user starts an export that takes a few minutes. Meanwhile, their access to those documents is revoked. Should the file still be generated and delivered?

The AI doesn’t need to have got it wrong. It may have handled the case perfectly. What I want to know is about the person. Had they noticed the problem? Could they say which requirement needed clarifying, and with whom? Could they write a test that tells one behaviour apart from the other?

In the first organisation, the senior colleague corrects or approves, and everyone moves on to the next task. In the second, before looking at the solution, the person writes down what they expect to happen in that case. Then they compare their prediction with the code, and the colleague talks them through the point where their reasoning fell short.

The two organisations can ship the same product on the same day. Only the second contains at least one explicit chance to learn something, and even that doesn’t prove the chance was taken. To know, you’ll need to see the same person facing a different case, a few weeks later.

There’s no incident in this story, no hopeless beginner, no stupid machine. The output is good, and the question about growth is still open. In your team, who notices the difference between the two?

Who Pays for the Apprenticeship

I’ve already written about the apprenticeship paradox: automation risks eating the process through which professional judgement is passed on, and that process is better made deliberate. In Measure What Stays Scarce I argued that value is shifting towards judgement and verification. Recognising that doesn’t explain how the capability gets built. Here I want to go one level down, from professional culture to a company’s accounts.

Think of a task that is economically useful, contained enough to give to someone inexperienced and interesting enough to teach them something. The client pays for the result. The company uses part of that work to grow someone, and absorbs the cost of supervising them. It doesn’t always pay off, but at least production and learning sit on the same invoice.

If AI does the hands-on part much better, automating it may be reasonable, and the manager who decides to do so may be making an excellent production decision. What’s left to decide is how to fund the experience that used to be built inside that production. In cases like this, training doesn’t disappear because someone thinks it’s useless. It disappears because nobody has an immediate reason to pay for it in its old form any more.

“People need to keep their skills up to date” doesn’t fill the gap. Updating your skills is an individual matter. Being able to take on gradually larger responsibilities inside an organisation is something else, and it depends on how the organisation hands out work.

The same goes for a line that turns up in a lot of adoption plans: juniors will review the AI’s output. That describes a job. It doesn’t explain how anyone becomes capable of doing it. Which error should that person spot? What evidence can they rely on? What happens when the solution is formally correct but the problem was badly defined? I’m not claiming that to assess a solution you always need to be able to write it from scratch; that would be too much. I’m claiming you need to know which skills that review requires, and then check that everyday work actually builds them. Nobody learns to supervise by being appointed supervisor.

Then there’s the pressure to deliver. A company can encourage people to use AI to learn while judging them only on closed tickets. In that setting, stopping to ask for an explanation, or digging into an anomaly, becomes a choice you pay for yourself. Whoever delegates straight away finishes first; whoever tries to understand has to justify the extra time. There’s no need to call the first lazy or the second virtuous. The organisation is making one behaviour pay and recommending another, and until it puts the cost of learning into how work is assigned, “use AI mindfully” stays a piece of advice without the conditions to follow it.

The most serious answer to all of this is that seniors can be hired from the market. For a single company, that’s legitimate. Not every firm has to grow every skill in-house, and a ten-person company couldn’t even try. But it should be called what it is: a strategy that depends on the future availability of people trained by someone else. If everyone did it, buying experience wouldn’t be enough to produce it. You don’t have to predict a collapse in the labour market to see the dependency. You just have to write it into the plan, next to the others.

Then there’s the oldest objection: if we train them, they’ll leave. True, the return is uncertain, and employees have every right to change jobs. But the comparison has to be done properly, between the cost of growing someone in-house and the cost of depending on a few people who are already experienced, who can leave too and cost more. In between there are other routes, such as sharing part of the investment with schools, universities or other companies.

And what if verification itself can be automated? Where reliable criteria exist, doing so is a good idea, and I’m not defending human reading of every line. Someone still has to decide whether those criteria fit the case, and someone has to notice the requirement nobody wrote down. If that too gets automated well one day, the skills worth training will change again. I won’t declare them irreplaceable on principle.

Eight Weeks to Find Out Whether Anyone Is Learning

We don’t have to preserve the path we took. We do have to be able to explain why whatever replaces it should work. I’m not asking anyone to prove a technology harmless before using it. I’m asking whoever designs the work to put learning among the expected results, and to check it.

The strongest objection at this point is that a junior with AI can grow faster. Maybe. By removing part of the hands-on work, we might bring newcomers into problems that matter sooner. An assistant can explain things when no colleague is free, and make available the kind of discussion that used to depend on the luck of finding a good mentor. That is exactly one of the possibilities worth testing. What I dispute is the unproven equivalence between having access to the tool and growing in the job.

Research in education suggests that how the help is built matters a great deal. In an experiment published in PNAS in 2025, Hamsa Bastani and colleagues gave nearly a thousand maths students at a Turkish high school two tools based on GPT-4: a generic assistant and a tutor designed not to hand over the answer. Both improved results during practice. Once the help was taken away, students who had used the generic assistant did worse than those who’d never had any, while those who had used the tutor ended up roughly level with them. A high school isn’t a software company, and I won’t transfer the numbers. I take something smaller from it: the way we organise assistance changes what people learn, and the differing results in the literature come partly from there. Learning can’t be inferred from tool adoption. It has to be looked at in your own setting.

The concrete proposal is an eight-week organisational experiment. The length is a design choice, not a figure taken from the studies. The point isn’t to certify new seniors. It’s to find out whether a particular way of working is building a capability we’ve defined in advance.

You pick one narrow responsibility, for example judging whether an access control check is correct. You pick a person who needs to learn to exercise it and a colleague able to examine their reasoning. The colleague’s time goes into the assignment, instead of being asked for as a favour between two emergencies. The person learning isn’t necessarily young. It could be a developer with fifteen years in the trade who has never worked on permissions and roles. Junior, here, describes where someone stands in relation to a problem, not their age.

At the start you watch how the person handles a representative case. AI stays available, because it’s part of the real work. Before seeing the solution, though, the person writes down a prediction: what should happen, and what evidence would show they’re wrong. Over the following weeks ordinary cases alternate with variations, and the colleague doesn’t just approve the result. They look for the point worth discussing. The question isn’t whether the explanation sounds convincing, but whether it leads to a relevant test.

Towards the end you bring in a new case, not a copy of an earlier exercise and not a memory quiz. You watch whether the person recognises the problem without anyone pointing at it, and how much help they need to reach a well-founded decision. A short test without the assistant can help separate what has been learned from what is still delegated, as long as it doesn’t turn into a general ban. Real work is done with tools, and it includes knowing how to use them well.

The experiment has to answer only a few questions. Does the person find the missing information on their own more often than before? Can they build a test that genuinely puts the solution under pressure? Do they recognise the moment to stop and bring someone in? The ability to withhold approval counts too. Making someone more confident without making them better at seeing their own limits is the worst possible outcome.

All of this has a cost, and it has to be measured alongside the effect on delivery. If after eight weeks there’s no visible improvement, you change the method. This isn’t a practice to defend because it’s “educational” by definition. Nor do you need to turn into a school: one responsibility learned properly is worth more than a generic AI training programme.

Where You Put the Cost

Hiring only people who are already experienced can be a reasonable choice. It means buying, along with their work, the result of an investment someone else made before you. There’s nothing scandalous about that. It is, however, different from believing that experience produces itself.

Nor is this a problem to offload onto employees. It would be incoherent to ask people to invest in their own learning while the organisation treats that time as a detour from delivery.

If AI makes the work we used to learn on unnecessary, we have a chance to build a better path, and maybe to reach more useful skills sooner. To know, we have to design that path and look at what it leaves people with, beyond the work delivered. We can let the old dues-paying go without regret. What replaces it can’t be called training until we’ve checked that somebody is learning.

Your AI adoption plan says who will oversee the machines. Where does it say how they’ll become capable of doing it?

The author

Andrea MargiovanniAI

I follow the relationship between AI and European regulation as a political fact, not a technical spectacle. I work with teams that have to make AI compliant with AI Act, CRA, NIS2 without reducing compliance to a checklist.

See the guide →