Issue 14 · Nov 2–8, 2026Agent Age News magazine · Other editions · Search · Method · HU / ENG

Work changes twice in the age of AI: first what people do, then what they are paid for.

Essay

Who teaches the profession when AI already does the first tasks?

If AI performs the preparatory tasks, early-career staff may lose precisely the path along which professional judgment is built. Replenishing expertise is therefore an organizational task that must be designed separately, and its preconditions are access to real work, visible reasoning and protected mentoring time.

Varga Zoltán · 2026-11-21 · 9 min read

I see two separate products in the early tasks of entry-level staff, and only one of them appears in the organisation's accounts. In their first months, a junior analyst cleans spreadsheets, searches for sources and prepares comparisons, and then an experienced colleague corrects their work. These tasks have a business result, and meanwhile another process is under way: the new colleague gradually learns which data is suspect, which questions are worth asking, and when an apparently regular answer is insufficient.

In work transformed by AI, the first product is produced faster, but the second no longer arises as a by-product. I see this learning process as something that therefore has to be designed separately, because along with today's efficiency, the organisation is also reshaping the channel through which the next generation of experts grows up.

The expert also sees the significance of the deviation

Imagine a proposal in which every calculation is correct, yet the company takes on too much risk with it. The junior employee checked the figures and the template. The experienced colleague, however, notices that the deadline the client requested had regularly led to disputes in an earlier project. The difference shows up in how attention is directed and how the situation is interpreted, since even in a properly prepared document one can recognise a badly managed business relationship.

Michael Polanyi's 1966 book, The Tacit Dimension, offers a conceptual foothold with its theory of tacit knowledge. People use certain parts of their knowledge in action without being able to turn them fully into statements. Relying on the parts, they recognize a whole, while their attention is mostly on the whole. In professional judgment, therefore, practiced perception and personal experience may play a significant role alongside the stated rule 1.

In his 2002 book Complications, the surgeon Atul Gawande gives a consistent account of his own learning. During practice, for days he saw only fragments of what had to be done, then one day the whole thing came together, and conscious learning turned into knowledge whose origin he himself could not pinpoint 2. The account shows that tacit knowledge builds up through repetition, mistakes and feedback, so producing it has its own workflow.

It is easy to make a romantic story out of this about inexpressible expert intuition. The usable conclusion is stricter. Even a good professional should try to show which cues they attended to, which alternative they rejected, and what experience their decision rested on. That part of knowledge is hard to put into words implies a need for more joint work; it hardly implies that professional decisions cannot be examined.

Learning requires access to real work

Jean Lave and Etienne Wenger's 1991 book Situated Learning: Legitimate Peripheral Participation describes learning as the transformation of participation in social practice. In legitimate peripheral participation, the newcomer gets access to the shared activity, from which fuller participation can gradually develop. In this sense, the periphery is a learning position. The point is that the beginner should come into contact with the meaning, the actors and the consequences of the work 3.

Imagine a consulting firm where the junior prepares every analysis with AI and then hands over a finished document. If the junior is left out of the client conversation and the final decision, their learning opportunity remains narrow despite the fast output. Correcting the document's errors shows what was inaccurate, and the client's reaction can also show why the question mattered.

When designing a learning path, it is therefore worth tracing which phases of work a beginner sees: whether they can be there at the first formulation of the problem, whether they hear the professional debate, and whether the result comes back to them. Besides access to information, the opportunity to speak also matters. A junior's own interpretation develops when it becomes clear how it relates to the others' deliberation.

A warning from an operating room

Matt Beane's ethnographic research on robotic surgery training, Shadow Learning, published in 2019, showed how trainees' hands-on participation can narrow in a technically more advanced workflow. The experienced surgeon could work more independently with the help of the robot, while the resident was given fewer substantive tasks. Some of the trainees who developed successfully sought practice opportunities alongside the established training routes, in ways that pushed against norms. Beane called this shadow learning 4.

The study analysed a specific professional setting, so its result cannot be carried over to every form of knowledge work without separate evidence. The organisational problem it raises can, however, be stated precisely: a technology may make today's expert's work more efficient while transforming the path along which the next expert can grow up. Ethan Mollick's 2024 book Co-Intelligence describes the same entry path in more general terms. Expertise has traditionally been built from the bottom up: the carpenter's apprentice, the newsroom intern and the resident learned alongside more experienced professionals, by trying and making mistakes, and according to Mollick this order is expected to change quickly under the influence of AI 5.

A company should aim to prevent progress from depending solely on the personal workarounds of especially persistent employees. An accessible learning task, a mentor's time and growing responsibility are resources that can be planned. Voluntary evening practice can be a personal choice, but an organization's capability building is better placed on more predictable foundations. Moreover, how the opportunity to learn is distributed also affects how later advancement is distributed.

Part of thinking can be made visible

The 1989 work by Allan Collins, John Seely Brown and Susan Newman, Cognitive Apprenticeship, uses the concept of cognitive apprenticeship to describe an educational approach in which expert problem-solving becomes more observable. Modeling, supported practice, gradual fading of help, articulating one's own procedure and comparative reflection together may help learners become independent. The authors develop the approach in the areas of reading, writing and mathematics 6.

In a corporate application, a mentor can talk through why they stop at a particular piece of data, what check they choose and why they leave a question open. The newcomer then gets a similar case, first with more support and later with less. In the discussion, alongside the final answer, it is worth attending to the points of choice, because two correct results can come from reasoning of differing reliability.

In this process, AI can produce examples, alternative solutions and deliberately flawed materials. The mentor must decide which error teaches something about the profession and which is merely artificial difficulty. A missing source, a wrong basic assumption and a mistyped number test different skills. The value of the task comes from the professional decision it targets.

AI can take on several learning roles

In a well-designed exercise, the system can first serve as a comparison partner. The junior produces their own draft, then compares it with another version. They mark what they would adopt, what they would reject, and what evidence they base the decision on. Keeping their own first attempt preserves the possibility of seeing, from later differences, how their thinking changed. The order matters. Chaeyeon Lim's 2025 preprint, DeBiasMe, is a position paper that proposes metacognitive support for university students' use of AI, meaning support built on observing one's own thinking, including deliberate friction against anchoring and confirmation bias 7. I regard putting one's own draft first as this kind of friction: whoever sees the machine version first can easily have their judgment anchored to it.

In another situation, the system can play the role of the client in a pre-designed practice conversation. The aim of the task may then be to clarify an incomplete requirement or to formulate a difficult follow-up question. The simulation makes sense if the mentor ties it to dilemmas that occur in real work and afterwards shows where the exercise oversimplified the situation.

The third role can be a check on one's own work. The system suggests a different interpretation, and the learner examines whether this changes their decision. The shared value of these three uses is that they prompt choice and justification. The outcome of all three suggestions can be judged by later independent task solving.

Enjoyable practice on its own tells us only about the participant experience, and the literature currently knows little more than that. In their 2024 study, Navigating STEM careers with AI mentors, Chi-Ning Chang and colleagues tested an AI mentor for career planning with eighteen postgraduate students in science and engineering fields. Participants reported a favourable experience, highlighting the instant responses and up-to-date information, and also raised concerns about incorrect information, bias and data protection. From this, the authors outlined two hybrid models combining human and AI mentors 8. The sample is small and what was measured is the students' own account, so this study says nothing about later independent performance. The 2026 systematic review by Fatima Fazal Abbasi and colleagues, Exploring the role of generative AI in skill development, based on 79 peer-reviewed articles, names precisely this gap: there is little longitudinal evidence, that is, evidence that follows the same people over time, on whether skills acquired with generative AI persist 9. To me this means that an organisation has to rely on its own measurement, and that a learning plan's success is best measured by what a colleague can solve without help months later.

The mentor's time should be treated as a business resource

In the short term, it is often simpler for the expert to do the task themselves, especially with a fast tool. Teaching, however, means waiting for the other person's attempt, uncovering the misunderstanding and returning to the decision. An organization that rewards only daily output therefore comes into tension with its own training goal, and this tension grows with AI, because the expert's independent work has become even cheaper relative to teaching.

My conclusion is that management needs to handle the situation with concrete agreements. In a project, it can be specified which parts prioritize fast execution and which get protected space for learning. In evaluating a mentor's work, it can be shown what their colleague can later do independently. Knowledge transfer then has a visible result, and the time spent on teaching becomes explainable.

A 2025 study by Bettina Falckenthal and colleagues, Intergenerational Tacit Knowledge Transfer: Leveraging AI, drew on 36 in-depth interviews to identify the success factors of junior–senior pairs: clarity of roles, complementary skills, compatible personalities and trust. The authors suggest that AI could support these factors as a supplementary service 10. Each of the four factors concerns the relationship between two people, so I see AI's role here as easing the pair's work; creating and sustaining the pair remains a management decision.

Expertise distributed among several people can also be valuable to a customer. The organization depends less on a single key person and finds it easier to continue work if someone else is assigned a task. This is a possible business consequence that is worth measuring in one's own operations. The training investment should therefore be assessed together with the operation of the whole expert system.

The entry path shapes the profession's future

Reworking junior tasks is an opportunity to build professional learning more deliberately. Some repetitive sub-tasks can be sped up, leaving more attention for decision points and feedback. This requires preserving the gradual taking over of real responsibility. A learner becomes a professional when, in more and more situations, they can recognize on their own what needs to be examined.

Alongside AI, the good training question is therefore about the next independent decision: what situation will the colleague be able to handle in a few months, and what experiences prepare them for it. The answer includes tasks, relationships, attempts and consequences. These build the expertise that the organisation can rely on at the next change of tools. I therefore regard the replenishment of expertise as an organisational task just like producing today's performance, and it likewise needs an owner, time and measurement.

Recommended reading

Sources

  1. [1] Polanyi, Michael (1966): The Tacit Dimension. Doubleday. Reissue, University of Chicago Press, 2009. book
  2. [2] Atul Gawande (2002): Complications: A Surgeon's Notes on an Imperfect Science. book
  3. [3] Lave, Jean; Wenger, Etienne (1991): Situated Learning: Legitimate Peripheral Participation. Cambridge University Press. book
  4. [4] Beane, Matthew (2019): Shadow Learning: Building Robotic Surgical Skill When Approved Means Fail. Administrative Science Quarterly, 64(1), 87–123.
  5. [5] Ethan Mollick (2024): Co-Intelligence: Living and Working with AI. book
  6. [6] Collins, Allan; Brown, John Seely; Newman, Susan E (1989): Cognitive Apprenticeship: Teaching the Crafts of Reading, Writing, and Mathematics. In Resnick, Lauren B. szerk., *Knowing, Learning, and Instruction*, 453–494. Lawrence Erlbaum.
  7. [7] Chaeyeon Lim (2025): DeBiasMe: De-biasing Human-AI Interactions with Metacognitive AIED (AI in Education) Interventions. arXiv. preprint, not yet peer-reviewed
  8. [8] Chi‐Ning Chang, John Hui, Cammie Justus-Smith, Tzu‐Wei Wang (2024): Navigating STEM careers with AI mentors: a new IDP journey. Frontiers in Artificial Intelligence.
  9. [9] Fatima Fazal Abbasi, Muhammad Usman, Bilal Latif (2026): Exploring the role of generative AI in skill development: a systematic literature review and bibliometric analysis. European journal of training and development.
  10. [10] Bettina Falckenthal, Manuel Au‐Yong‐Oliveira, Cláudia Figueiredo (2025): Intergenerational Tacit Knowledge Transfer: Leveraging AI. Societies.

Essays and reviews →

Weekly email

One email a week with the main stories from the Agent Age editions. You confirm by email and can unsubscribe in one click.

Almost done: we sent you an email. Open it and press the confirm link. If it does not arrive within a few minutes, check your spam folder.

↑