An analyst sends her manager a market sizing model on day three of a project. It is clean, well structured, and roughly right. Two years ago that would have taken her a fortnight and three rounds of corrections. Her manager is pleased. Six weeks later a client asks why she chose one assumption over another, and she cannot answer, because she did not choose it.
This is happening quietly across professional services, engineering, marketing and finance. The work that used to train people is the work AI now does fastest. Nobody sat down and decided to dismantle the apprenticeship model. It is dissolving by convenience, one delegated task at a time.
Where judgement actually came from
Junior work was never valuable because the output was valuable. A first year lawyer reviewing documents, a graduate engineer checking calculations, a junior planner pulling competitor research: most of it was slow, and a good deal of it was redone by someone more senior. Organisations tolerated the inefficiency because it produced something other than the deliverable. People learned what good looked like by producing bad versions of it and having them corrected.
That correction loop is where judgement came from. You made an assumption, someone senior asked why, you could not answer, and you learned to ask yourself the question before anyone else did. The output was the by-product. The reasoning was the point.
When AI produces a competent first draft, the loop breaks in one specific place. The junior still receives feedback, but the feedback lands on something they did not build. They can accept a correction without ever having held the reasoning that produced the error, which means they cannot carry it to the next problem. They have learned an answer rather than a way of thinking.
Why the bill arrives sooner than people expect
The common response is that this will sort itself out, because juniors will learn to work alongside AI and that will become its own skill. There is something in that. But there is a timing problem underneath it.
Organisations are cutting junior headcount now, on the basis of this year’s productivity figures, while the cost of the missing training appears in five to eight years, when there is nobody ready to step into the senior roles. The people making the headcount decision will mostly not be the people who inherit the gap. That is exactly the kind of cost that gets deferred until it becomes structural.
The nearer cost is simpler. Someone has to be able to tell when the AI output is wrong. That judgement is precisely what the old apprenticeship built. If you remove the training ground and keep the review responsibility, you end up with people who are accountable for checking work they were never taught how to check. They will approve it, because it looks right, and looking right is the thing these tools are best at.
What the organisations handling this well do differently
The firms getting this right have not banned the tools or slowed anyone down. They have changed what they ask for and what they inspect.
The shift is from reviewing deliverables to reviewing reasoning. A senior person who only ever sees the finished model has no way of knowing whether the junior understood it. A senior person who asks what the analyst expected to see before she ran the numbers finds out immediately.
They have also become deliberate about which work stays slow. Not all of it, and not for its own sake. But a small, protected amount of work where the point is the learning rather than the efficiency, and where everyone knows that is the arrangement.
Five practical things you can change this quarter
If you manage people early in their careers, these are worth trying.
Ask for the prediction before the output. Before anyone generates the analysis, ask what they expect it to show and why. It takes two minutes. It forces a hypothesis into the open, and it gives them something to be surprised by, which is when learning actually happens.
Review the reasoning, not only the artefact. Change the review question from “is this right?” to “walk me through why this assumption and not the obvious alternative.” You will find out quickly whether the person understands what they have handed you.
Hand juniors problems where the tool fails. Ambiguous briefs, thin data, contested definitions, situations where the answer depends on context the model does not have. These are the problems that build judgement, and they are the ones most often reserved for senior people.
Make the AI a first draft, never a final one, and say so explicitly. Teams drift into treating output as finished because nobody said otherwise. Naming the expectation out loud changes behaviour more than a policy document does.
Give feedback on what was rejected. Ask what the tool suggested that the person chose not to use, and why. The decision to discard is where the judgement lives, and it is invisible in the deliverable.
What this looks like in a real organisation
An engineering consultancy noticed that its graduate intake was producing more work than any cohort before it and asking noticeably fewer questions. The technical directors could not point to anything wrong in the outputs. What they could point to was a change in the conversations: graduates were defending outputs rather than exploring problems.
They made two changes. Graduates were asked to write a short note before each piece of analysis, setting out what they expected to find and which assumption they were least confident about. And review meetings started with that note rather than with the finished work.
The first month was uncomfortable. Several graduates found they could not articulate an expectation at all, which was itself the finding. Within a quarter, the directors reported that the questions had come back, and that the quality of challenge in project meetings had improved across the whole team, not only among the juniors. The senior engineers had started writing the notes too.
Common traps
Restricting the tools instead of changing the teaching. Bans push usage underground and remove your ability to see how people are working. The problem is not access. It is that nobody has been shown what good use looks like.
Assuming juniors will pick it up informally. Apprenticeship worked because it was structured, repeated and supervised, even when nobody called it training. Remove the structure and the informal learning does not survive on its own.
Measuring only throughput. If your dashboards show output per person and nothing about capability, you will keep making the trade that looks good this year and costs you later. Track something about development, even if it is imperfect.
Treating this as a junior problem. Senior people lose calibration too. If a director has not built a model in three years and now reviews AI-generated ones, their judgement is also decaying, more slowly and less visibly.
Reflection questions
Worth sitting with, on your own or with your leadership team.
- If your most junior people had to defend their work without the tool in the room, how many of them could?
- What did you learn in your first three years that your current juniors will never have the chance to learn?
- Who in your organisation is accountable for checking AI-assisted work, and where did they build the judgement to do it?
- Which of your development happens by design, and which of it was happening by accident inside work you have now automated?
- If you cut a junior role this year, who is doing that person’s job in 2033?
The productivity gain is real and it is not going away. The question is whether you are spending it or investing it. Right now most organisations are spending it, because the saving is measurable this quarter and the shortfall is not.
