When AI Takes Over the Apprenticeship Years: The Silent Loss of Human Expertise

When Machines Take Over the Apprentice Years

There is a moment in every engineering career that cannot be simulated: the first truly unexpected failure that briefly shakes your entire worldview. The dead-end debugging marathon at three in the morning. The baffling "why did that just work?" – and the slow, painstaking understanding that follows. It is precisely from this friction that what we call expertise is born. And it is precisely this friction that powerful AI systems are increasingly removing from everyday professional life.

IEEE Spectrum recently addressed the topic in a widely discussed guest article: author Richard Mitchell, founder of AuraSpark Technologies, draws on his own experience with a nuclear power plant control system to describe how his team deliberately built manual steps into automatable sequences – not in spite of the automation, but because of it. The reason: those who do nothing but supervise eventually stop understanding. This design principle – inefficient by design, on purpose – is more relevant today than ever.

The Data Is Hard to Argue Away

What long circulated as vague unease among vocational training experts is increasingly taking on empirical shape. A Harvard working paper analysing data from approximately 65 million employees across more than 280,000 US companies shows: following the introduction of generative AI, employment among junior staff fell by around nine percent within six quarters – while the number of experienced professionals remained stable or even grew. An analysis by Stanford University based on ADP payroll data confirms the direction: the youngest employees in AI-exposed occupations have lost significant ground since the end of 2022.

Particularly revealing is the distinction drawn by the Stanford researchers: where AI automates, juniors lose out. Where AI merely augments – that is, complements existing skills rather than replacing them – the junior ratio remains stable or even rises. The causality has not yet been conclusively established, and that is important to emphasise. Economists at the New York Fed point to remote work as a contributing factor: companies are reluctant to hire inexperienced talent they cannot onboard in person. Yet both explanations point to the same problem: the channel through which expert knowledge flows from experienced professionals to career starters – the apprenticeship channel – is damaged.

"You cannot become a senior engineer without first having been a junior. Expertise is not downloaded. It is earned."

This line from the IEEE Spectrum article strikes at the heart of the matter. And it confronts companies with a strategic question that reaches far beyond HR departments.

The Automation Paradox: An Old Lesson from Aviation

Safety-critical industries have known this phenomenon for decades under the term Automation Paradox: the more capable the automation, the less practice the human gets – and the more overwhelming the moment when the machine hands back control. That moment typically arrives precisely when the situation is at its most complex.

The prime example is Air France Flight 447, which crashed over the Atlantic in 2009. Iced-over sensors delivered faulty data, the autopilot disengaged correctly and handed control back to the crew. What followed was not a technical failure – it was a competency failure. Pilots who had spent years watching a highly automated aircraft fly itself could no longer manually compensate for an aerodynamic stall at high altitude. The aircraft was intact. The skill that automation had quietly eroded was not.

The response of aviation authorities was instructive: automation was not rolled back – instead, manual flying was institutionalized as a competency in its own right. In 2017, the FAA formally declared in a Safety Alert that manual flying is the foundation of all further piloting skills. Skill decay was recognized as an independent safety risk and factored into training programs.

What this means for companies investing in AI automation

Dr. Maik Bunzel, founder and CEO of mabucon.eu, regularly points to exactly this blind spot in client conversations: "The ROI of automation is often immediately visible – the long-term competency costs are not. When a junior employee never learns why a process works in a certain way, because the AI already handles it, companies are building on a knowledge base that is systematically thinning out."

For organizations adopting AI agents and workflow automation, this raises concrete design questions:

• Which tasks should deliberately remain manual? Not out of nostalgia, but to preserve and develop critical process understanding.
• How is learning from failure enabled in a structured way? AI systems tend to prevent errors – and in doing so, they also remove the opportunity to learn from them. Sandboxed environments and structured case analyses can compensate for this.
• What role does the human employee take on in augmenting vs. automating? The distinction is crucial – not only for job security, but for the quality of the decision-making basis in edge cases.
• How is mentoring scaled in a hybrid human-AI environment? When AI takes over the simple tasks, senior employees need to be planned more deliberately and more structurally as knowledge carriers.

The silent danger: Confident but Wrong

Perhaps the most dangerous scenario is not total dependency on AI systems – that can be named and addressed easily. The real danger is more subtle: a generation of professionals who have learned to supervise and validate AI outputs without ever having internalized the fundamentals needed to assess whether an output is plausible or implausible.

Large language models and agent-based AI systems deliver answers with consistent confidence – regardless of whether they are correct or fundamentally wrong. Confident but wrong is not an anomaly but a structural characteristic of current generative AI systems. Those who lack the professional grounding to recognize these errors will carry them forward into workflows, reports, and decisions – without ever noticing.

This is not an argument against AI automation. It is an argument for conscious design: for systems that not only maximize efficiency, but also cultivate the human competence that makes this efficiency safe and resilient in the long term.

From Efficiency Thinking to Competence Design

The lessons from nuclear power and aviation can be transferred to software engineering, data analysis, legal consulting, controlling, and many other knowledge professions. Wherever AI takes over routine cognitive work, the same structural question arises: How can we ensure that the human expert of the future not only monitors outputs and processes, but truly understands what they are monitoring?

Dr. Maik Bunzel distills this into a simple formula he applies when designing automation projects: "A good AI solution doesn't make people redundant – it makes them better in the moments that truly matter." This presupposes that the moments in which humans still think for themselves are not systematically optimized away.

For companies currently building or evaluating AI agents and automation workflows, this means: the question of ROI must necessarily be supplemented by the question of competence impact. Which capabilities are strengthened by the new architecture? Which ones atrophy? And which safeguards – analogous to manual flying in aviation – are deliberately built in?

Outlook: AI Adoption as a Design Process

The next phase of enterprise automation will not reward those who automate the most. It will reward those who automate the most intelligently – those who understand which competencies must be preserved within the organization, and who design their AI systems accordingly.

The nuclear power plant from the IEEE Spectrum article was never built. But the design principle embedded in its circuits is more relevant today than ever before: Sometimes the inefficient is the smartest choice. Companies that grasp this and implement it in their AI strategy will, five years from now, not only have more efficient processes – they will also have better experts.