
Who will learn to make decisions when we automate entry-level work
Automating entry-level tasks also changes how expert judgment develops. Renewing it requires learning pathways that are deliberately designed and tested.
The promise of freeing people from routine work so they can focus on higher-value decisions rests on an assumption worth making explicit: there will be people capable of making those decisions. In a team of experienced professionals, that condition may seem assured. They know the context, recognize unusual results, and understand when a technically correct answer is insufficient. If some of that judgment developed through work the organization is about to remove, we also need to consider those who will join after automation and never follow that path.
An experienced professional and a beginner can use the same tool, receive the same recommendation, and achieve similar results. That does not establish that they have developed an equivalent ability to judge the situation. The first may recognize an omission because they have seen its consequences; the second may accept the answer because they do not yet know what is missing. The tool might also help the beginner understand something that would have taken much longer to discover alone. The difference depends on what happens during the work: which problems they encounter, what they must resolve themselves, what they can observe, and how they find out whether their interpretation was sound.
Calling certain tasks “entry-level work” therefore tells us less than it might seem. It describes their place in a career but does not explain the role they play in it. Sorting requests, preparing an initial analysis, or investigating an incident can produce a useful result while also allowing someone to learn to distinguish situations, formulate a hypothesis, and receive corrections (opens in a new tab) before the stakes become higher. Or it may simply involve repetition with little meaningful learning. The fact that a task has historically been assigned to beginners does not make it a good learning experience, nor does it oblige us to retain it when there is a better way to do it.
The available evidence does not justify equating automation with a loss of capability either. In the study by Erik Brynjolfsson, Danielle Li, and Lindsey Raymond (opens in a new tab), published in The Quarterly Journal of Economics, the introduction of an AI assistant in customer support produced particularly large productivity gains among less experienced workers and those with lower prior performance. The authors also found signs of learning: during assistant outages, workers who had already used it continued to handle cases more quickly than before adopting it. This is limited evidence of learning: outages were infrequent, the estimates were imprecise, and the measure was handling time. How professional judgment develops over the long term remains an open question. Taken together, the findings suggest that technology can give beginners access to practices previously concentrated among more experienced people.
How the experience of learning changes
The immediate improvement leaves another question open: what learning does the new way of working enable, and what learning no longer takes place? Correct answers can teach; they can also conceal the decisions that made them possible. For someone who is learning, receiving a finished solution, examining why it works, and developing an interpretation that is then tested are different experiences. If we look only at the quality of the output (opens in a new tab), those differences may remain invisible precisely when results are improving.
In his research on learning in robotic surgery (opens in a new tab), Matthew Beane described how a technological reorganization could reduce residents’ participation in the practice on which their training depended. The work was still getting done, but the relationship between those who knew how to do it and those who needed to learn was changing. Although this involved a different technology and professional practice, it helps us recognize a relevant mechanism: access to formative experiences can diminish without that change registering as an operational failure.
This does not mean every simple query must be handed back to people. It does require us to ask how someone can learn to recognize an exception with little access to the situations that make it exceptional. They could, for example, examine resolved cases with the help of AI and test their own diagnoses against the evidence in each case. What matters is that there be a clear progression from joining the team to taking on the responsibility expected of them; immediately assigning “higher-value work” does not, by itself, create the capability that work requires.

What we learn as performance improves
An assistant that allows people to explore alternatives, consult earlier cases, and test a hypothesis can enrich the available experience if it keeps the beginner engaged with the problem and expands what they can understand. It might even make situations accessible that previously depended on encountering the right expert at the right moment. But that possibility requires attention to the task the person actually performs. An interface that offers explanations does not, simply by offering them, demonstrate that anyone examines them, understands them, or learns to challenge them.
The distinction between assisted performance and learning emerges clearly in an experiment by Hamsa Bastani and her coauthors (opens in a new tab) involving high school students practicing mathematics. Access to a GPT-based assistant improved their results while they had the help, but one version led to worse performance than the control group in a subsequent assessment without it. A version designed to guide learning mitigated that harm. In this short-term school setting, the way help was provided changed the relationship between solving and learning. For work design, the finding offers a useful caution: examine separately what a person achieves with a tool and what they are learning to do through it.
Nor would it make sense to define expert judgment as the ability to reproduce everything the system does by hand. Work changes, and the capabilities needed to take responsibility for it can change too. Someone overseeing an automated solution will need to understand aspects that may not have been central to the previous process: where the information comes from, the conditions under which a recommendation applies, or the difference between a plausible result and one that deserves trust. Some skills will become less important. Others will become indispensable. Designing a learning pathway requires deciding what those capabilities are and which experiences can develop them in the new practice.
Renewing the judgment that sustains the operation
The time lag makes this kind of gap difficult to notice. An implementation can rely for years on the judgment of people who learned their profession before the work changed. They interpret exceptions, detect what the system does not show, and decide when to intervene. As long as they do this well, operational indicators may confirm that automation is working. Yet that performance does not answer the question of whether someone joining under the new conditions is learning to take on that responsibility. An organization can retain a capability it has yet to demonstrate it can renew.
This is also the limitation of solving the problem by hiring experienced people from outside. For an individual company, doing so may be a reasonable decision. But it shifts the development of that expertise elsewhere and preserves the dependence on someone continuing to produce it. It does not establish that the new work design develops experts. If this option is chosen, it should be recognized as a deliberate dependency and its viability examined, rather than treating the current availability of professionals as a permanent feature of the market.
Responsibility for this continuity extends to those who decide what to automate. Eliminating tasks, redistributing cases, and changing access to context also change the conditions in which people can learn. Asking a training department to close the gap afterward, without revisiting those decisions, may leave it trying to compensate for missing experience with instructional content. The leaders who approve the redesign need to be able to explain the path they will offer new team members and how they will know that it is developing the required capability.
To begin testing this, it helps to distinguish the results of those who learned before automation from the development of those who join afterward. An indicator that combines both groups may improve thanks to accumulated experience while revealing very little about what the new pathway is producing. Tracking that development requires observing which situations new team members are beginning to understand and how much support they need to interpret them, with attention to the cases they actually encounter. The quality of their work still matters, but sustaining capability requires another kind of evidence: that people are learning to take responsibility for decisions that currently rest with others.
Automating entry-level work can open up a better way to learn and practice a profession. It can remove an unnecessary burden, accelerate access to knowledge, and allow someone to participate in meaningful problems sooner. For that promise to be defensible, the organization must take responsibility for the pathway it is creating. The decision remains incomplete if it determines how the work will be done but leaves to chance how the people who must answer for it will learn.





