For most modern professions, the path to competence has long involved doing a considerable amount of work that did not look particularly sophisticated from the outside.
Junior consultants collected data, checked numbers, prepared presentations and produced first analyses that would later be reviewed by someone more experienced. Young lawyers read contracts and assembled case material. Analysts built models, reconciled spreadsheets and produced early drafts that would pass through a manager's hands.
In many professions, the route towards good judgment began, paradoxically, with tasks that required relatively little judgment.
The system was not especially efficient. But it was educational.
The value of these activities was never limited to their output. They placed inexperienced people close to real problems and, perhaps more importantly, close to people who already knew how to solve them.
A junior consultant did not simply learn to make a slide. They learned why a partner rejected one argument and kept another, which information mattered to a client, when a number looked suspicious, how an experienced professional talked about uncertainty in front of a CEO, and how a technically correct answer could still be the wrong recommendation.
Much professional expertise has always been transmitted this way: indirectly, unevenly and often without anyone calling the process learning.
Artificial intelligence is beginning to intervene in exactly this mechanism.
Not because it can replace senior professionals. The immediate problem may be almost the opposite.
AI is increasingly capable of doing precisely the work organizations traditionally gave people before they became senior.
The hidden curriculum of work
A recent Financial Times article on the UK consulting industry captures the tension well. As AI takes over more research, analysis and other technical work, firms such as EY, KPMG and PwC are placing greater emphasis on human capabilities such as professional judgment, empathy, storytelling and leadership.
At the same time, leaders interviewed by the FT are again emphasizing the need for early-career professionals to work in contexts where they can observe and learn directly from more experienced colleagues. Sayeh Ghanbari, managing partner for consulting at EY UK & Ireland, explicitly connects the issue to the weakening of consulting's traditional apprenticeship model during remote work.
The discussion has predictably drifted into another round of the return-to-office debate. But the more interesting issue is not where people work. It is how people learn to work.
Professional apprenticeship is not simply about proximity. It is about participation.
Research on workplace learning has made this point for decades. Alison Fuller and Lorna Unwin showed that professional development depends heavily on the social and pedagogical relationships between learners and more experienced colleagues.
Their distinction between expansive and restrictive learning environments remains useful: organizations differ not only in how much formal training they provide, but in the access people have to experienced colleagues, varied activities, knowledge and progressively more demanding forms of participation.
Professional learning does not happen only when someone attends a course. It happens while they work.
Junior work produces two things
We tend to treat a professional task as if it produced one result: an analysis produces an analysis, a financial model produces a model, software produces software and a client presentation produces slides.
But early-career work actually produces two things. The first is obvious: the output. The second is harder to see: the person who gradually becomes capable of doing more difficult work.
Organizations have strong economic reasons to automate work at the bottom of professional hierarchies. Routine research, first drafts, document review, information synthesis and initial analysis consume enormous amounts of junior time.
If AI can do a meaningful part of that work in minutes, asking a graduate to spend six hours doing it manually can look absurd. Sometimes it is.
There is no good reason to preserve inefficient work merely because previous generations had to endure it. But removing a task and replacing the learning embedded in that task are two different decisions. The first is relatively easy. The second is much harder.
Productivity is not the same as learning
Imagine a junior analyst spending several hours building a first market analysis. They search for information, encounter contradictory sources, make bad assumptions, notice an odd number, ask a more experienced colleague why it matters, rebuild the analysis and defend the conclusion when a manager challenges it.
After dozens of experiences like this, something difficult to measure begins to happen. The analyst no longer merely knows how to produce the analysis. They begin to know what to look for.
Now suppose AI produces the first version of the same analysis in three minutes. The organization has saved several hours. But what has the analyst learned?
The answer may be a great deal if they verify the output, challenge assumptions, compare alternatives and discuss the reasoning with someone more experienced. It may be almost nothing if they accept the analysis, polish the wording and pass it on. The difference is not primarily technological. It is pedagogical.
A well-known field study involving more than 5,000 customer-support agents found that generative AI increased productivity by about 14% on average, with substantially larger benefits for less experienced workers. That is excellent news for productivity. But performing more like an expert and becoming an expert are not necessarily the same thing.
More recent experimental work is beginning to capture this distinction. When AI takes over too much of the cognitive effort, people can achieve better immediate results without developing conceptual understanding and independent problem-solving to the same degree.
Organizations therefore face an uncomfortable question: are we optimizing today's performance or tomorrow's capability? Ideally both. AI does not guarantee that automatically.
Efficiency and development operate on different timescales
The efficiency benefits of AI are immediate and easy to measure. We can estimate hours saved, count outputs, compare costs and measure task completion time.
The loss of developmental experiences is much harder to see. Quarterly dashboards do not contain a metric called professional judgment that never developed.
The consequences appear years later, when organizations discover that technically capable employees have had too few opportunities to build contextual understanding, work with ambiguity, develop client judgment or acquire the judgment expected from senior professionals.
Professional hierarchies have always assumed that people become capable of difficult work by accumulating experience through progressively more complex work. If we remove enough of the lower rungs, the ladder begins to look rather strange.
The relevant question is therefore not only whether AI will eliminate junior jobs. Even if the jobs remain, what developmental experiences remain inside them?
We do not need to preserve pointless work
The easy conclusion would be that organizations should protect junior tasks from automation. That would be a mistake.
We do not need to preserve hours of pointless research, mechanical checking or PowerPoint production so that new generations can experience the character-building benefits of aligning text boxes at 11:47 p.m.
The challenge is to separate work required for production from experience required for learning. Historically, the two came bundled together. AI is beginning to separate them.
If AI can produce the first draft, the developmental task can become evaluating that draft. If AI can conduct initial research, the junior can compare sources, identify missing evidence and defend why one interpretation is stronger than another.
If AI can build the model, learning can move towards examining assumptions, testing scenarios and identifying what the model cannot know. If routine execution becomes cheaper, junior professionals may need earlier exposure to ambiguity, clients and real decisions.
In theory, this could produce a better development model than the old one. But it will not happen by itself. It has to be designed.
Managers will have to become teachers again
There is a less comfortable implication for managers. The traditional model allowed a great deal of learning to happen accidentally, as a side effect of production. The junior did something, the manager corrected it, and the junior tried again.
If AI removes part of that work, managers will have to become much more intentional about developing people.
They will need to explain reasoning that used to remain implicit, review AI-generated work with juniors rather than simply celebrate a faster deliverable, make their own thinking visible, create access to difficult conversations and ask people why they agree or disagree with an AI recommendation.
They will also need to give people responsibility before they feel completely ready, while providing enough support for the experience to remain developmental rather than reckless.
In other words, some of the time saved through AI will need to be reinvested in development.
The financially tempting version of AI adoption is obvious: automate 30% of a junior employee's work and immediately fill the released capacity with 30% more output. The developmental version is less impressive in a spreadsheet: automate part of the work and use some of the recovered time to deliberately build capabilities that the old work used to develop accidentally.
The first approach optimizes this year's utilization. The second protects the expertise the organization will need five years from now.
From accidental learning to designed learning
For decades, organizations outsourced a surprising amount of professional development to the structure of work itself.
People learned because they had to do things they were not yet very good at, close to people who were. Responsibility increased gradually. Mistakes generated feedback. Repetition created pattern recognition. Observation transmitted things no competency framework could fully describe.
We did not have to design this system particularly carefully because the work produced it naturally. AI changes that.
As technology becomes capable of doing more of the work traditionally given to beginners, organizations will have to identify which experiences people still need in order to become experts and then create those experiences deliberately.
The result could even be better than the old model. We can eliminate genuinely pointless work and give young professionals earlier access to analysis, decisions, real problems and client interactions.
But that outcome does not arrive with the AI licence. Technology can eliminate a task in seconds. It cannot decide what the person doing that task was supposed to learn from it.
That becomes a management problem, a leadership problem and, perhaps above all, a learning problem.
Organizations that understand the difference will not ask only how much junior work they can automate. They will have to answer a harder question: if AI does the work through which people used to learn to become experts, how will they become experts?
Leadership in the AI Era
Explore Leadership in the AI EraSources and context
The article starts from recent observations in the consulting industry and places them in the context of research on workplace learning, apprenticeship and the effects of AI on productivity and learning.