Who will become the expert when AI does the entry-level work?

AI can already help a junior professional produce work that looks remarkably mature.

It can generate code, draft a design, prepare an impact assessment, propose test cases and explain unfamiliar concepts in seconds. For organisations under pressure to improve productivity, the attraction is obvious: people can perform useful work sooner and experienced colleagues spend less time on routine guidance.

But this creates a leadership question that deserves more attention.

The risk is not simply that AI replaces some junior tasks. The deeper risk is that organisations remove the experiences through which professionals develop judgment, while continuing to assume that expertise will somehow appear later.

The hidden value of routine work

Looking back at my early career, some of the most developmental work did not appear strategic at the time.

Code reviews developed a troubleshooting mindset. They taught me to look beyond whether code worked and ask how it might fail, how it would behave under different conditions and how another engineer would maintain it.

Manual testing created a different kind of understanding. Following a process step by step exposed the underlying logic, boundary conditions and dependencies in a way that merely reading the expected result could not.

Creating design documents forced me to understand a customer’s network architecture more deeply. A diagram or document could not be completed responsibly without understanding interfaces, assumptions and the consequences of a design choice.

Even finding the relevant technical requirements for a design was formative. The task required more than locating a document. I had to decide which requirements applied, how they interacted and whether they changed the design.

Formal training can explain the concepts behind these activities. It cannot fully reproduce the judgment built by performing them repeatedly, receiving feedback and confronting exceptions.

Productivity and learning are different outcomes

Research shows that AI can shorten the experience curve. A large field study involving 5,172 customer-support agents found that access to a generative-AI assistant increased productivity by 15% on average. Less experienced and lower-skilled workers gained the most, and agents with two months of tenure and AI support performed as well as or better than unsupported agents with more than six months of experience.

The study also found encouraging signs of durable learning. During periods when the AI system was unavailable, workers who had been exposed to its recommendations retained some of their productivity improvement. AI appeared to transfer elements of expert practice rather than merely supply answers.

That is the optimistic case: AI can make high-quality guidance available at the moment of work and help people learn faster.

Other evidence shows why leaders should remain careful. In a field experiment involving nearly 1,000 high-school mathematics students, access to a general-purpose AI assistant improved performance while the tool was available. When access was removed, however, those students performed 17% worse than students who had never received it. A more carefully designed AI tutor, with safeguards intended to support learning, largely mitigated the damage.

The educational setting differs from professional work, but the underlying distinction is highly relevant. Producing a correct answer with AI does not prove that the person has developed the capability to produce, evaluate or defend that answer independently.

A 2025 study of 319 knowledge workers provides another warning. Across 936 examples of AI-assisted work, higher confidence in AI was associated with less reported critical thinking. Higher confidence in one’s own ability was associated with more. The study was based on self-reports and does not establish that AI caused the reduction, but it suggests that critical thinking changes when people rely on generated output. More effort moves toward verification, integration and oversight.

Taken together, the evidence does not support a simple verdict that AI either develops or deskills people. It can do both. The result depends on how people engage with it and how leaders design the surrounding work.

The capability illusion

I have seen junior professionals use AI to create code or prepare a network-impact report much faster than before. The output may be well structured and technically convincing. But when the person lacks sufficient understanding of the design, they can struggle to explain why the solution is safe, which components are affected, what assumptions were made or how the system will behave in an exception.

This creates what I call a capability illusion: the output appears senior, while the judgment behind it remains junior.

The illusion is difficult to detect if managers assess only the submitted artifact. A polished document, comprehensive test suite or functioning piece of code may hide the fact that its owner cannot challenge the reasoning, adapt it to a new context or take responsibility when conditions change.

That gap matters because AI remains fallible and context-dependent. People cannot responsibly verify an answer if they do not understand the domain well enough to recognise what may be missing. AI fluency therefore cannot replace domain knowledge. Its value increases when it is built on top of domain understanding.

Redesign apprenticeship instead of removing it

Organisations should not preserve routine work merely because previous generations learned that way. AI gives us an opportunity to create a faster and more purposeful route to expertise.

The goal should be accelerated apprenticeship: let AI reduce low-value effort while deliberately preserving diagnosis, explanation, challenge, feedback and accountability.

Consider testing. AI can generate and execute many tests automatically. Requiring a graduate to perform every repetitive test manually would waste the technology. But designing the test strategy should remain the learner’s work. They should decide what must be tested, identify failure modes, prioritise risk, explain the chosen coverage and evaluate whether the generated tests are sufficient.

The same principle applies to design and coding. AI may generate the first output, but development should assess whether the professional understands the architecture, can identify weak assumptions, can explain trade-offs and can defend the final decision.

This changes the purpose of assessment. If AI will be available in everyday work, banning it from all assignments may test an artificial environment. Allowing its output to pass without examination is equally weak. A better approach is to permit AI-assisted production and assess the person’s understanding through explanation and defence.

Exams, reviews and development conversations should increasingly ask:

• Why did you choose this approach?
• What alternatives did you reject, and why?
• Which assumptions are most fragile?
• How did you verify the AI output?
• What would change if one part of the environment changed?
• What failure would be hardest to detect?

These questions reveal whether AI has accelerated learning or merely concealed its absence.

A five-part model for AI-era development

1. Attempt

Ask the learner to frame the problem, create an initial diagnosis or outline a strategy before using AI. The purpose is not to produce a complete answer unaided. It is to make their current mental model visible.

2. Augment

Use AI to expand the option set, simulate scenarios, surface relevant knowledge and demonstrate patterns associated with stronger performance. The learner should compare the generated response with the initial approach rather than simply replace it.

3. Challenge

Require the learner to find assumptions, omissions and possible errors. This builds the habit of treating AI output as material to be evaluated rather than authority to be accepted.

4. Defend

Ask the learner to explain the final recommendation in their own words and take ownership of its consequences. For important work, oral defence, peer review or scenario-based questioning may reveal understanding more effectively than another written submission.

5. Perform independently

Periodically remove AI assistance or introduce an unfamiliar exception. The objective is not to recreate a pre-AI workplace. It is to determine whether the person has retained enough knowledge to recognise when the tool is wrong and continue when it is unavailable.

This model also helps leaders distinguish between different levels of AI autonomy. A novice may use AI extensively but require deeper review and explanation. As the person demonstrates domain knowledge and sound judgment, the organisation can expand both their authority and the autonomy of the systems they supervise.

What leaders must measure

Most AI programmes measure adoption, usage and immediate productivity. Capability development requires additional evidence.

Leaders should examine whether people can explain AI-assisted work, detect deliberately introduced errors, transfer knowledge to a new situation, design an approach before generating an answer and perform critical tasks when assistance is removed.

They should also monitor who receives access to difficult cases and expert feedback. If AI handles every routine case and senior professionals retain every complex one, juniors may lose both the foundational repetition and the exceptions that build judgment. Development therefore has to be designed into work allocation rather than left to chance.

The leadership responsibility extends beyond training courses. It includes task design, review practices, assessment, mentoring, decision rights and the measures used to recognise performance.

The next generation of experts

AI can help people rise the competency curve much faster. It can make expert patterns visible, provide immediate feedback and enable early-career professionals to attempt work that was previously beyond their reach.

But speed of output should not be mistaken for depth of capability.

Tomorrow’s experts will still need domain knowledge, systems understanding and the confidence to challenge a plausible answer. They will need to explain decisions, defend trade-offs and remain accountable when AI contributes most of the initial work.

The organisations that develop those experts will not be the ones that preserve every old task. They will be the ones that understand what each task used to teach—and deliberately build that learning into a new form of work.

The practical examples and interpretations in this article reflect Saket Kumar’s personal perspective. They do not represent the views of any employer or customer.