AI Is Automating the Work That Creates Experts

When companies decide what to automate, they usually measure time, cost, error rates and output. They rarely measure what a task teaches the person doing it. That omission matters because many tasks that look repetitive or inefficient are also part of the invisible apprenticeship system through which people become capable professionals.

A junior analyst building a first model, a developer tracing a basic bug, a support agent handling a common complaint and a lawyer preparing an initial research memo are all producing work. They are also learning how information fits together, which details matter, where mistakes appear and when a familiar pattern no longer applies. If AI removes the task without replacing the learning process, the organisation may gain speed today while weakening the expertise it will need later.

The workforce question is therefore larger than whether AI will reduce entry-level hiring. Leaders need to ask what happens to the practice loops that once turned inexperienced employees into people capable of judging, correcting and supervising intelligent systems.

Expertise Is Built Through Correction

Expertise is often described as knowledge, but knowledge alone is not enough. Experienced people have encountered enough cases to recognise weak signals, notice unusual combinations and understand when a standard answer is unsafe. That judgement develops through repeated cycles of attempting work, receiving feedback, correcting errors and seeing the consequences of decisions.

Many junior tasks are valuable precisely because they create those cycles. Drafting the first version of an analysis forces someone to decide what is relevant. Cleaning data reveals how imperfect real information can be. Handling straightforward customer cases teaches the normal pattern, which later makes an abnormal case easier to recognise. Reviewing an expert’s finished output does not create the same depth because the difficult choices have already been made and hidden.

AI can compress these cycles. A first draft arrives before the employee has framed the problem. A code assistant proposes the solution before the developer has traced the failure. A research tool summarises the evidence before the analyst has had to compare sources. The output may be better and faster, but the employee receives fewer opportunities to build an internal model of why it is correct.

AI Can Accelerate Learning, but the Design Matters

The risk is not inevitable. Some of the strongest evidence on generative AI at work shows that it can help less experienced employees learn faster when it is used as guidance rather than as a substitute for participation.

An NBER study of 5,179 customer support agents found that access to a generative AI assistant increased productivity by 14 percent on average, with a 34 percent improvement among novice and lower-skilled workers. The researchers found suggestive evidence that the system helped disseminate the practices of stronger agents and move newer workers down the experience curve more quickly.

The design of that work matters. Agents received suggestions during real customer conversations, but they still had to interpret the situation, choose how to respond and observe the outcome. The learning loop remained intact. AI added timely access to expert patterns without completely removing the employee from the decision.

That is different from a workflow in which the system completes the task and the employee performs a quick approval at the end. When people remain active in diagnosis, choice and consequence, AI can function as a coach. When they are reduced to accepting polished output, the technology may increase short-term production while doing much less for long-term capability.

Automation Is Reaching the Tasks That Build Human Capital

It is easy to assume that AI will remove only low-skill, repetitive work. Current evidence suggests a more complicated pattern.

Anthropic’s January 2026 Economic Index found that the tasks covered by Claude tended to require more education than the economy-wide average. In an experimental estimate, the researchers found that removing those AI-covered tasks from occupations would produce a first-order deskilling effect on jobs because many of the covered tasks represented higher-education components of the role. Anthropic was careful to say this was not a prediction, since occupations may evolve and new tasks may appear, but the finding should change how companies think about automation.

Routine and developmental are not opposites. Preparing an analysis can become routine for an experienced employee while remaining essential practice for a junior one. Drafting standard contracts, debugging familiar errors, preparing forecasts and reviewing customer history may be automatable, but they also expose employees to the structure of the profession.

This creates a strategic problem. The easiest task to automate may also be the task through which a new employee learns the context needed to handle the difficult task later.

The Verification Paradox

Many organisations respond by saying that people will move from production to verification. AI will generate the work, while humans will check accuracy, apply context and remain accountable. This sounds reasonable, but it contains a paradox: verification itself requires the expertise that the automated work used to help create.

A Microsoft Research study (MSFT ) involving 319 knowledge workers and 936 examples of generative AI use found that greater confidence in AI was associated with less critical-thinking effort. The study also found that AI shifted critical thinking toward verification, integration and task stewardship. Those activities are valuable, but a person cannot verify well without enough domain knowledge to recognise what may be missing.

An inexperienced employee can check whether an answer is clearly inconsistent or unsupported. It is much harder to notice that the analysis used the wrong frame, ignored an operational constraint or reached a plausible conclusion for the wrong reason. Polished AI output can hide the absence of understanding because it looks complete before the employee has developed the judgement to interrogate it.

Research on the “jagged technological frontier” makes the problem more concrete. In an experiment with 758 consultants, AI users completed tasks within the technology’s capability frontier 25.1 percent faster and produced results of more than 40 percent higher quality. On a task outside that frontier, however, AI users were 19 percentage points less likely to reach the correct solution than participants working without AI.

The critical capability is therefore not simply using AI. It is recognising when the system is operating outside the boundary of reliable performance. That boundary is rarely obvious, and learning to see it requires experience with both successful cases and failures.

The Hidden Succession Risk

The World Economic Forum’s Future of Jobs Report 2025 found that employers expect 39 percent of workers’ core skills to change by 2030. It also found that 63 percent of employers already view skills gaps as a major barrier to transformation, while 77 percent plan to upskill employees and 41 percent expect to reduce workforce numbers as AI automates tasks.

These strategies may conflict if organisations remove too much of the work through which skills are formed. A company can reduce junior hiring, rely on senior employees to oversee AI and still report strong productivity for several years. The weakness appears later, when experienced people leave, demand expands or the organisation needs employees capable of handling situations the system has not seen.

Senior experts can also become a new operational constraint. If AI increases the volume of work while only a small group can verify difficult outputs, those people spend more time reviewing and correcting rather than teaching. The company gains production capacity but concentrates judgement in fewer hands.

No organisation can automate apprenticeship and then assume it will continue buying experience from the market. If many companies follow the same strategy, the external pool of experienced talent eventually becomes thinner and more expensive.

Measure the Developmental Value of Work

Before automating a task, companies should evaluate it on two dimensions. The first is its output value: how important, costly and time-sensitive is the result? The second is its developmental value: what capability does performing the task build?

Tasks with low output value and low developmental value are obvious candidates for full automation. Reformatting routine documents or transferring information between systems may teach little after basic competence has been reached.

Tasks with high output value but low developmental value can often be automated with appropriate control. The organisation still needs governance because the result matters, but preserving the manual process may contribute little to employee growth.

The most neglected category contains tasks with modest immediate output value but high developmental value. A junior employee’s first draft may not be commercially important because a senior person will revise it, yet producing it creates a valuable comparison between initial reasoning and expert feedback. Removing that step saves time but also removes practice.

Tasks with both high output value and high developmental value should usually be augmented rather than fully delegated. Employees can use AI for research, options and critique, but they should retain meaningful responsibility for framing the problem, explaining the decision and learning from the outcome.

Build the Apprenticeship Layer Deliberately

Once developmental tasks have been identified, companies can redesign them rather than preserving inefficient workflows unchanged. The goal is not to make employees repeat every manual process from the past. It is to recreate the experiences that build judgement in a faster, more intentional form.

One option is to preserve first-pass thinking. Before seeing an AI recommendation, the employee records an initial diagnosis, forecast or decision. The system can then provide an alternative, and the employee must explain where the two differ and which approach they would use.

Organisations can also build libraries of AI failures, edge cases and disputed decisions. Reviewing only successful outputs teaches employees to trust the system. Studying where it failed helps them recognise the conditions that require escalation, additional evidence or human intervention.

Managers should evaluate whether learning transfers beyond the tool. Can the employee explain the reasoning behind the final output? Can they identify the assumptions that would make it wrong? Can they adapt when the context changes or the AI is unavailable? These questions reveal more about readiness than the number of prompts used or tasks completed.

Most importantly, junior employees still need real responsibility. Simulations and structured practice help, but expertise also depends on consequences, feedback and contact with real complexity. AI should make those experiences safer and more frequent, not remove them completely.

The Workforce Advantage Will Come From Faster Expertise

The strongest organisations will not be those that automate the greatest number of junior tasks. They will be those that remove low-value effort while accelerating the development of people who can use, challenge and supervise AI responsibly.

This requires a broader definition of return on automation. Time saved and output produced are important, but leaders should also measure whether the new workflow strengthens or weakens the organisation’s future supply of judgement. A process that reduces cost while damaging the expertise pipeline may be efficient only on a short time horizon.

Before approving the next automation project, leaders should ask one additional question: what did people learn by doing this work, and where will that learning happen after the work is automated?

The answer will determine whether AI simply produces more output or helps produce the next generation of experts.