What Happens When AI Removes the Work That Used to Develop Professionals?

Imagine a junior project professional preparing a risk register. Before generative AI, the work required reading the project plan, comparing dependencies, speaking with workstream leaders and translating uncertain concerns into explicit risks. The first register was rarely excellent. A senior colleague would question generic entries, expose assumptions and ask why a particular response was feasible.

Now AI can produce a credible register in seconds. It can organise risks into categories, assign probability and impact scores, and recommend responses in the language of professional practice. The junior employee edits the document, the manager approves it and the team saves several hours.

The immediate output may be better. But what happened to the learning that once occurred while producing it?

This question extends beyond project management. Junior lawyers review documents, researchers code sources, consultants structure analyses, educators prepare lessons, engineers inspect specifications and clinicians encounter routine cases. Some of this work is repetitive, inefficient or poorly designed. Yet it has also provided repeated exposure to the language, patterns, exceptions and consequences of a profession.

When AI removes the task, it may also remove the developmental pathway hidden inside it.

The governing distinction is between the operational value of work and its formative value.

Operational value concerns what the task produces now. Formative value concerns what the person becomes capable of doing later because they performed it.

These values can diverge. A task may contribute little to the final output while providing essential practice to a novice. Another task may consume hours without developing any capability worth preserving. Responsible automation requires organisations to distinguish the two before redesigning the work.

Work has always contained a hidden curriculum

Professional development is often discussed as though it occurs mainly through courses, qualifications and formal training. These activities matter, but much professional capability develops inside the work itself.

People learn by observing how experienced colleagues interpret an ambiguous situation, attempting a task before they fully master it, receiving correction and encountering cases that do not fit the standard procedure. They learn which details deserve attention, which mistakes have consequences and when a formally correct answer is contextually wrong.

Eraut's research on informal workplace learning distinguishes a range of processes through which people learn from experience, including participation in group processes, working alongside others, tackling challenging tasks, solving problems, receiving feedback and consulting knowledgeable colleagues.¹ Much of this learning is not labelled as education. It occurs because the structure of work exposes the learner to progressively more demanding forms of participation.

This is the hidden curriculum of professional work. It is carried through assignments, supervision, informal explanation, observed decisions, mistakes and increasing responsibility. A novice does not merely complete tasks. The novice gradually learns what the profession notices, values, questions and refuses to accept.

Automation decisions rarely include this curriculum in their productivity calculations. They measure the time removed from the task but not the exposure, feedback or calibration removed from the learner.

Junior work is not only cheap labour

The defence of junior work can easily become nostalgic. Some entry-level tasks are monotonous because organisations have historically transferred low-status burdens to people with the least power. Manual reformatting, duplicated data entry and avoidable administrative effort should not be protected under the language of development.

The relevant question is not whether a task is difficult or traditional. It is what cognitive and professional activity the task requires.

Reviewing contracts may develop sensitivity to wording, exceptions and downstream obligations. Coding qualitative data may force a researcher to confront ambiguity before themes become stable. Preparing a lesson may require an educator to anticipate misconceptions and sequence explanations. Building a project schedule may reveal how apparently independent activities constrain one another.

The visible artefact is only one outcome. The learner may also be developing:

• a domain vocabulary and conceptual map; pattern recognition across repeated cases; sensitivity to errors and anomalies; understanding of how upstream choices affect downstream work; calibration about what they know and when to seek help; a professional identity shaped by standards and consequences.

Not every instance produces these benefits. Repetition without feedback can consolidate poor habits. Observation without participation can remain passive. Excessive workload can crowd out reflection. The formative value of junior work depends on challenge, feedback, variation and increasing responsibility.

This is consistent with research on expertise. Ericsson, Krampe and Tesch-Römer argued that expert performance develops through extended, structured practice rather than experience alone.² Their formulation of deliberate practice should not be applied mechanically to every profession, but the broader principle is useful: capability develops when activity is designed to improve performance, includes feedback and stretches what the learner can currently do.

Routine work becomes developmental when the learner must notice, decide, receive correction and adapt.

AI can separate performance from formation

AI Can Give You the Answer Without Building the Capability distinguished assisted performance from capability. The professional-formation argument applies that distinction to professional development.

AI can allow a novice to produce work that resembles expert output before the novice has acquired the mental models that make expert judgment possible. The report is coherent, the categories are correct and the recommendation sounds credible. Supervisors see improved output and may reasonably assume that the employee is progressing.

Yet the polished result can conceal what the employee did not do. They may not have interpreted the raw information, compared competing possibilities, struggled with an exception or constructed the reasoning that connects evidence to conclusion. If the supervisor edits only the final wording, the missing capability may remain invisible.

This creates a formation gap: the distance between the standard of work a person can produce with assistance and the standard of judgment they can exercise when the situation changes.

The gap may not become visible during routine work because AI continues to provide adequate support. It appears when the system is unavailable, the case falls outside familiar patterns or the professional must decide whether the generated answer is wrong. By then, the person may hold a role whose apparent experience exceeds their independent calibration.

AI can accelerate the appearance of professional competence before the underlying judgment has had time to form.

The apprenticeship problem

Apprenticeship is not simply an expert telling a novice what to do. It combines observation, modelling, participation, coaching, correction and progressively more consequential responsibility. The learner sees not only the finished answer but how the expert frames the problem, responds to uncertainty and recovers from error.

Technology can disrupt this pathway even when it improves the primary work. Beane's research on robotic surgery provides a revealing example. In traditional surgery, trainees could participate progressively in the operation. Robotic systems concentrated control at the console and reduced opportunities for meaningful trainee participation. Approved learning practices became less effective, and only a minority of trainees developed competence through what Beane called “shadow learning,” which included opportunistic and sometimes undersupervised practices outside the formal pathway.³

Robotic surgery and generative AI are different technologies operating in different professional contexts. The transferable lesson is narrower: when technology changes who participates in the work and what they can observe, the old apprenticeship system may stop functioning even though the organisation continues to call it training.

Generative AI can place a polished output between novice and supervisor. Instead of observing the learner's reasoning, the supervisor sees an edited result. Instead of wrestling with the material and then receiving targeted feedback, the novice may compare two finished answers. The interaction becomes more efficient while the learner's thinking becomes less visible.

This is not an argument for withholding useful tools. It is an argument for redesigning how reasoning, feedback and progressive responsibility occur when the old task no longer provides them automatically.

Five capabilities are especially vulnerable

When AI changes developmental work, five capabilities deserve explicit protection.

1. Pattern recognition

Professionals learn patterns through repeated exposure to cases, not only through descriptions of the pattern. They notice combinations, variations and early signals that become difficult to articulate. If AI processes the underlying material and presents only the classification, the novice sees fewer of the cases from which pattern recognition develops.

2. Contextual interpretation

A procedure can state what usually applies. Professional judgment must decide what the present situation means. This requires knowledge of local constraints, stakeholder histories, informal authority and recent events. A novice develops contextual interpretation by comparing formal guidance with lived cases and discussing why an exception matters.

3. Error sensitivity

Experienced professionals often notice that something feels inconsistent before they can fully explain why. This sensitivity develops through exposure to errors, near misses, corrections and consequences. If AI removes the initial analysis and supervisors see only polished output, novices may encounter fewer visible mistakes and fewer explanations of why they matter.

4. Consequential awareness

Junior tasks teach how one person's work affects another. Preparing data reveals what downstream analysts require. Drafting a project plan exposes dependencies. Reviewing a client case reveals how wording changes expectations. Removing the task can remove contact with the chain of consequences unless the workflow makes that chain explicit.

5. Professional judgment

Judgment integrates the other capabilities under uncertainty. It determines what matters, which standard applies, when evidence is sufficient and what action can be defended. It cannot be acquired from explanations alone because part of judgment is calibration through decisions, feedback and consequence-bearing practice.

These capabilities should not all remain human-only activities. AI can support their development by presenting varied cases, exposing counterexamples and generating challenges. The design question is whether the professional still performs enough meaningful interpretation to develop and retain them.

What the evidence does and does not establish

The long-term effects of generative AI on professional formation remain uncertain. The technology is changing faster than occupational training systems, and longitudinal evidence across professions is still limited. It would therefore be premature to claim that AI is already causing universal deskilling or the collapse of apprenticeship.

Earlier automation research nevertheless identifies a durable risk. Bainbridge observed that when automated systems perform routine operations, human operators may be left to intervene in rare and difficult situations without receiving the continuous practice needed to maintain those skills.⁴ The irony is that automation can make the remaining human contribution more demanding while reducing the opportunities through which readiness develops.

Generative AI also offers evidence of a more constructive pathway. Brynjolfsson, Li and Raymond studied the introduction of an AI assistant among 5,172 customer-support agents. Access increased productivity by 15 per cent on average, with larger gains among less-experienced and lower-skilled workers. The researchers also found evidence consistent with worker learning and diffusion of the practices used by stronger performers.⁵ In this setting, AI did not merely remove work. It helped make expert knowledge more accessible during performance.

Organisation-level evidence also complicates the deskilling narrative. A study of German firms found that sustained AI adoption was associated with a 14 per cent increase in new apprenticeship contracts among firms already providing training, although continuing-training resources shifted towards higher-skilled employees.⁶ This does not establish that AI improves apprenticeship quality, and the German dual system may not generalise to other labour markets. It does show that AI adoption can increase the organisational demand for structured formation rather than simply eliminate it.

The responsible conclusion is conditional:

AI will not automatically erode professional formation, but productivity gains do not automatically preserve it. The outcome depends on how work, supervision and learning are redesigned together.

Redesign formation rather than preserving obsolete work

If a task is inefficient but developmentally valuable, the objective should not be to preserve the inefficiency. It should be to preserve or improve the learning mechanism.

Several design patterns can help.

Supervised AI-assisted practice allows the learner to use AI while making their own framing, interpretation and revisions visible to a supervisor. The review focuses on decisions rather than prose.

Attempt before assistance requires an initial diagnosis, structure or recommendation before AI provides its version. The purpose is not assessment theatre. It creates a trace of the learner's thinking that can be compared and coached.

Contrastive practice asks the learner to compare their answer, the AI output and an expert response. They identify where the answers diverge and explain which approach should be trusted under which conditions.

Edge-case simulation uses AI to generate unusual, conflicting or high-risk scenarios that routine work may no longer expose. A qualified professional must validate the scenarios and guide the debrief.

Explanation and defence requires the learner to justify the assumptions, evidence and consequences of an AI-assisted output. If they cannot explain why a recommendation is appropriate, completion should not be treated as competence.

Progressive responsibility moves the learner from observation to bounded decisions and then to greater consequence as capability is demonstrated. AI support may change at each stage rather than remaining permanently maximal.

The governing principle is to automate the burden while deliberately relocating the learning.

Employers and professions must own the formation architecture

Organisations cannot treat professional development as an accidental by-product of whatever work remains after automation. If they remove the tasks through which people once developed, they inherit a design responsibility: where will the required capability now come from?

Every significant automation initiative should therefore include a capability impact assessment. Alongside cost, speed, quality and risk, the organisation should identify what the task previously taught, which roles depended on that learning and how capability will be demonstrated in the redesigned system.

Managers also need new forms of visibility. Reviewing a final AI-assisted deliverable may reveal little about the learner's reasoning. Supervision may need annotated decisions, oral defence, observed practice, simulations or comparison across cases. The objective is not surveillance. It is to make development visible enough to support meaningful feedback.

Professional bodies face a related challenge. Experience requirements often count time served or tasks completed because these measures once functioned as rough proxies for exposure and responsibility. When AI can complete much of the task, completion becomes weaker evidence of capability. Professions may need more direct evidence of interpretation, verification, transfer and judgment.

The individual professional still has a responsibility

Employers should redesign formation, but professionals cannot assume that every organisation will do so quickly or well. Each person needs to understand which capabilities their work is no longer developing automatically.

This is where the professional-formation argument connects to Your Career Has a Curriculum and the Self-Curriculum. The individual should maintain a capability map, seek assignments that provide genuine developmental exposure and periodically practise selected tasks without assistance. They should ask for feedback on reasoning, not only on the finished output, and pursue edge cases that stretch judgment beyond routine production.

The objective is not to prove independence from AI. It is to remain capable of using AI without becoming unable to recognise when its support is inadequate.

The Professional Formation Audit

Before automating or removing a task that has traditionally been performed by developing professionals, ask five questions:

1. What did this work teach? Identify the patterns, concepts, relationships, standards and consequences encountered through doing it.
2. Which parts were burden, and which parts were formative? Separate repetitive effort from interpretation, feedback, variation and responsibility.
3. Where will the learning now occur? Name the practice, simulation, comparison, coaching or assignment that will replace the developmental function.
4. How will capability become visible? Determine how supervisors will observe reasoning, test transfer and identify overreliance on assistance.
5. When will responsibility increase? Define the evidence required before the learner handles more uncertain or consequential work.

If an organisation can describe the productivity gain but cannot answer these questions, its automation case is incomplete.

Removing the task does not remove the need for capability

The work that develops professionals has never been perfectly designed. Apprenticeship can reproduce outdated practices, allocate burden unfairly and depend too heavily on the quality of individual supervisors. AI gives organisations an opportunity to improve these systems rather than merely preserve them.

But improvement requires seeing formation as part of the work architecture. A faster document, cleaner analysis or automated recommendation does not explain how the next generation will learn to recognise a flawed assumption, interpret an exception or accept responsibility for a difficult decision.

Return to the junior project professional. There is no reason they must spend two hours formatting a risk register. There is every reason they must examine the project, confront uncertainty, learn from experienced judgment and become progressively responsible for what the register represents.

The task can change. The developmental requirement remains.

If organisations automate the work but do not redesign the learning, they may gain immediate productivity while consuming the capability on which future performance depends.