At a recent dinner with enterprise leaders, the conversation settled on an uncomfortable question. Where will the next generation of leaders come from? Youth unemployment in the UK sits at 16%, with close to a million 16 to 24-year-olds not in education, employment or training. At the same time, companies are accelerating the adoption of AI into the roles that once served as the first rung of a career. While this means fewer entry-level jobs, it also removes the work that trained many of those at the table to lead.

Entry-level work can be tedious, but this is where pattern recognition is built. Repeating tasks exposes the edge cases, while allowing mistakes when the stakes are low. Automating these tasks takes away the slow accumulation of knowledge on which many managerial structures rely. A short-term gain may become a long-term liability.

We may be mistaking familiarity for effectiveness. Those who climbed a traditional career ladder tend to assume it is the best way to develop judgement, because it is the path they experienced. In reality, it is a slow and uneven system that relies on time, repetition, and a degree of luck. Some people emerge as strong decision-makers despite the process, not because of it. If AI removes the need for repetitive junior work, it does not have to remove the development of future leaders. But it does require a redesign.

Applying agents to human workflows captures the inefficiencies of those workflows. In the same way that classroom training is ineffective at tackling specific challenges, agents don’t work well at replicating the messy world of human communication. The redesign starts with how people learn.

The new apprenticeship model is a shift from doing to deciding. Juniors will no longer spend years preparing reports or gathering data. Instead, they interpret outputs, challenge assumptions, and decide what matters. That exposure accelerates learning because it puts them in contact with the core of the business earlier. Leadership is built through making decisions under uncertainty, not by completing tasks in sequence. With AI handling execution, the constraint shifts from access to information to the ability to judge it.

Training today often takes place away from the office. The lessons are generic and largely forgotten by the time workers get back to their desks. AI agents providing continual assistance can explain, prompt, and critique work as it is being done. The knowledge worker learns from a system that has seen every similar case before.

A typical day may start with a review of work that an agent performed overnight. The worker evaluates the output before the agent challenges the assessment. Then the human makes the decision, and the agent records and adapts. The agent does, while the human decides.

The next step is to review edge cases where the outcomes are not as expected. Adjustments are made to how decisions are framed and evaluated, and company-specific logic is introduced to the system. This creates visibility into decision-making across other teams. This improves consistency and exposes juniors to how decisions are made at higher levels.

Juniors should progress faster than before with agentic assistance. Their learning is continuous, contextual, and based on feedback, rather than the time spent on an online course. Precision replaces the previous hit-and-miss approach. This is how you rebuild the lost apprenticeship layer.

The pushback is that if people become dependent on AI then they will not learn. Early exposure to decision-making may breed false confidence. Agents are only as good as the logic they are trained on. In a system that only rewards output, these risks are real.

Management should be about making effective decisions, rather than validating processes and approving outputs. As a result, agentic support must explain reasoning rather than provide answers. In turn, this must be challenged. In this way, humans improve an agent’s judgement as well as their own.

As an example, a junior estimator who once spent months assembling cost plans is now given a complete AI-generated estimate on day one. Instead of building it, they test assumptions, adjust risks, and decide whether to bid. In weeks, they are exposed to decisions that previously took years to encounter.

AI is removing the work that once trained future leaders, but it does not remove the need for judgement. The old model relied on time, repetition, and chance to produce decision-makers. The new model replaces that with earlier exposure, faster feedback, and systems that guide how decisions are made.

There is a stark choice. Either agents hollow out the talent pipeline, or they are used to compress learning for the next generation. The companies that treat this as a redesign of how people develop, rather than how work gets done, will produce better leaders, faster.