“Experience is the name everyone gives to their mistakes.”
Oscar Wilde
There was a time when the first years of a career came with a slightly humiliating bargain. You did the research nobody else wanted to do. You cleaned the spreadsheet. You made version 17 of the deck, then watched a senior colleague remove half of it in ten minutes.
At the time, it could feel like a ritual designed by someone who hated graduates. Looking back at my years in agencies, I see something else. The work was repetitive, sometimes infuriating, often badly paid. But it was also rehearsal. You learned what a weak argument sounded like before you had to defend one in a client room. You learned that a good strategy is not a slide with more arrows. You learned how much context disappears when a neat sentence leaves out the people who have to live with it.
Now AI can produce the first draft, clean the data, summarize the meeting, generate the research plan and turn rough notes into a competent-looking presentation before the coffee has cooled. I am not nostalgic for junior people spending three nights aligning logos. But I am worried about the apprenticeship system we are removing without designing a replacement.
The question is not whether AI will remove entry-level tasks. It already is. The question is who learns judgment when those tasks no longer create the repetitions from which judgment grows.
The work nobody romanticised was still a workshop
In most professional services businesses, the pyramid had a practical function. Senior people framed the problem, juniors did the first pass, the work came back with comments, and the next pass was less naive. The model was not noble by default. It also created overwork, hierarchy and more unnecessary PowerPoint than any society should tolerate.
Still, the mechanism mattered. A junior analyst who has compared twenty competitors begins to notice that the supposedly obvious category story rarely survives contact with the market. A young creative who has written thirty headlines begins to recognize when an idea is merely loud. A product manager who has sat through usability sessions learns that users are remarkably inventive at misunderstanding the feature everyone in the room considered self-explanatory.
Those repetitions create pattern recognition. More importantly, they create the emotional muscle of being wrong in public, revising a point of view and carrying the consequences into the next decision.
The old model trained people through production. AI is shrinking the production layer much faster than companies are rebuilding the training layer.
That is the practical side of the shift I explored in The skills that will define the AI-augmented workforce. We are moving from execution toward orchestration. The catch is that an orchestrator still needs enough experience to know when the orchestra is playing the wrong piece.
The disappearing rehearsal creates a judgment debt
The temptation is to say that the answer is simple: juniors can use AI and become more productive. Of course they can. They should. Refusing useful tools because previous generations suffered through slower workflows is not a career philosophy, it is hazing with a nostalgic soundtrack.
But productivity and learning are not the same thing.
If someone asks a model for an analysis, receives a polished answer and edits a few sentences, they may have completed the task without acquiring the underlying mental map. They have not necessarily seen the false starts, assessed conflicting evidence or learned why one data point mattered and another did not. The work looks finished. The learning may not have started.
This is the judgment debt. A company saves time now by automating the junior layer, then discovers later that it has too few people who can challenge the model, coach a client through ambiguity or decide when the elegant output is simply wrong for the situation.
AI can compress production. It cannot compress the experience of noticing what does not fit.
That distinction becomes visible when the system makes a plausible error. The person who has only seen good-looking outputs may not have enough scars to spot it. The person who has done the messy work, argued with the data and been corrected by someone more experienced often does.
The evidence is more complicated than the slogans
The International Labour Organization offers a useful antidote to the familiar theatre of certainty. Its 2025 global index finds that one in four workers is in an occupation with some exposure to generative AI. Clerical work remains most exposed, and some professional and technical roles have become more exposed as the technology handles increasingly specialized tasks. But the ILO’s central conclusion is not mass replacement. Because most occupations contain tasks that still require human input, job transformation is the most likely outcome.[1]
That is reassuring only if we pay attention to the word “transformation.” A role can survive on paper while losing the tasks that made it a viable entry point.
IBM’s recent account of its Forward Deployed Units makes the shift tangible. It describes small teams of three or four specialists working with 10 to 20 AI agents. The agents may generate requirements, code, test cases or defect analysis. The humans define the problem, add business context and decide whether the work meets the client’s needs.[2]
That model contains both promise and warning. It can reduce pointless labor and bring expertise closer to the client. Yet it also asks less experienced people to develop judgment faster, in a setting where the visible evidence of how work gets made may be thinner.
BCG’s 2026 Applied AI Index reaches a similar tension from the boardroom. Respondents expect AI to reduce workforces by roughly 10% to 15%, while 89% expect it to generate new work. The most mature companies are far more likely to practice deliberate workforce planning, meaning they decide which roles should exist in an agentic organization rather than hoping the org chart will sort itself out.[3]
The data do not prove that every junior job will disappear, or that every company will need fewer people. They do show that task exposure is not a neutral event. When a task disappears, its hidden learning function can disappear with it.
This is a quieter version of the concern behind The white-collar apocalypse nobody is talking about (yet): the immediate loss is not always a role. It can be the route by which someone became capable of doing that role well.
Stop treating learning as a by-product of busywork
The answer is not to preserve inefficient work for sentimental reasons. It is to make learning an explicit product of the operating model.
I would start with four practical choices.
| Old default | Better AI-era design |
|---|---|
| Juniors learn by producing first drafts alone | Juniors observe how seniors frame a decision before the agent starts |
| Quality review happens at the end | Review happens around assumptions, sources and trade-offs |
| AI output is treated as a shortcut | AI output becomes material for critique, comparison and revision |
| Career progression follows time served | Career progression follows demonstrated judgment in real decisions |
Give juniors ownership of the brief, not only the output. Before asking the agent to work, ask the junior person to explain the decision at stake, the missing information and the people affected by a wrong answer.
Make critique visible. A senior person should not merely replace a paragraph generated by AI. They should explain why it fails the client context, the evidence standard or the commercial reality. That commentary is now part of the apprenticeship.
Create decision reviews, not just delivery reviews. The useful question is not “was the deck good?” It is “what did we believe, what did the agent assume, and what would have changed our mind?”
Let people meet the consequences. A junior who presents an AI-assisted recommendation to a client, then listens to the objections, learns more than someone who watches a perfect answer arrive in a shared folder.
This is not a request for more meetings disguised as humanism. It is a request to move training closer to the places where judgment is actually formed.
The real management question
Companies like to describe AI as a copilot. Fine. But copilots do not make pilots unnecessary. They raise the standard for the person in the seat.
The danger is that organizations will buy the productivity, remove the routine, congratulate themselves on efficiency and only later notice that nobody knows how to diagnose an unfamiliar problem without asking the machine first.
In systems terms, the junior layer was not merely a cost center. It was a capability supply chain. Remove it and the organization may still ship work for a while. What it cannot replenish so easily is the stock of people able to judge the next difficult case.
Cui bono when we treat entry-level work as a cost to eliminate rather than a place where professional judgment is made?
Because if nobody is allowed to do the basic work any more, who will be ready to judge the agent when its answer looks convincing, arrives on time and is quietly wrong?
References
[1] International Labour Organization, Generative AI and Jobs: A Refined Global Index of Occupational Exposure, Working Paper 140, 2025.
[2] IBM Think, AI is tightening consulting teams and raising the bar for who’s on them, 21 September 2026.
[3] Boston Consulting Group, The Formula for Agentic AI Value: The Applied AI Index 2026, 30 September 2026.