What Happens to Apprenticeship When AI Does the Junior Work?

A junior analyst prepares the first model. A new account manager listens while the customer explains the problem. A young engineer traces the failure before the senior engineer arrives. Much of this work is slow, uneven, and increasingly easy to assist with AI.

Owners are right to remove needless repetition. Yet the rough first pass often did something the finished document could not show. It exposed the beginner to weak signals, forced a comparison with an earlier case, and created a mistake close enough to the work for someone experienced to correct it.

When AI produces the draft, summary, diagnosis, or recommended response, the output may improve immediately. What becomes of the person who once learned by struggling through the route?

Brynjolfsson, Li, and Raymond studied the staggered introduction of a generative-AI assistant across 5,179 customer-support agents. Access increased issues resolved per hour by 14 percent on average and by 34 percent among novice and lower-skilled workers, with minimal gains for the most experienced workers. The authors also found suggestive evidence that the system disseminated practices associated with stronger performers and may have supported worker learning.[Generative AI at Work]

This is important counterevidence to the easy fear that AI merely strips experience from work. In this setting, AI appears to have compressed part of the learning curve. Newer agents gained access to patterns and responses that previously took longer to acquire.

Customer support also offers conditions many executive and technical decisions do not: high case volume, digital records, recurring problem types, rapid feedback, and observable outcomes. The study establishes a valuable possibility in a bounded domain. It does not settle what happens when cases are rare, consequences arrive years later, or the correct answer depends on a relationship no system can observe directly.

The World Economic Forum’s 2025 Chief People Officers Outlook reported that participating people leaders placed career stagnation and skill atrophy from AI overreliance among their leading near-term workforce concerns. This is evidence of executive concern, not evidence that atrophy has already occurred across the workforce.[Chief People Officers Outlook 2025: Talent Strategy Amid Global Disruption]

A 2025 survey of 319 knowledge workers gathered 936 first-hand examples of generative-AI use. Higher confidence in AI was associated with less self-reported critical-thinking effort, while critical-thinking activity shifted toward verification, integration, and stewardship. The design was correlational and self-reported, so it cannot establish that AI caused a loss of capability.[The Impact of Generative AI on Critical Thinking: Self-Reported Reductions in Cognitive Effort and Confidence Effects From a Survey of Knowledge Workers]

Taken together, the studies leave a live management question. People may learn faster with a capable assistant, perform with less direct effort, and move their attention toward supervision. Whether that produces stronger future experts depends on the work, the tool, the feedback, and what people are still expected to do without assistance.

Nobody becomes wise by formatting the same report a thousand times. The formative part may be seeing enough cases to notice variation, receiving correction while the memory is fresh, watching an experienced person explain an exception, and carrying consequences as authority expands.

Some junior work is scaffolding. It gives a person a view of how customers describe trouble, how a clean rule meets a messy record, or how a senior leader changes course when one fact no longer fits. Remove the task and the company may also remove the observation point. Automate the output while preserving the case, feedback, and reflection, and the apprenticeship may improve.

The earlier distinction between speed and unaided capability matters here. Assisted performance tells the company what a person and tool can do together. Succession eventually asks what the person can recognize, decide, and recover when the tool is unavailable or outside its reliable range.

Begin with the decisions people will be expected to make three or five years from now. What cases must they have seen? Which cues should become familiar? Which exceptions should stop them? Where should they practice disagreeing with a confident recommendation?

Critical incidents, contrasting cases, decision thresholds, and scenario rehearsal can help when natural exposure becomes thinner. These methods do not replace real responsibility. They can make sure the future leader encounters meaningful variation before the first live case arrives alone.

The Passage applies that logic to a named successor and consequential role: inherited history becomes material for practice, challenge, and progressively independent decisions. The broader workforce problem will require the same discipline at a different scale.

For each AI-assisted workflow, look one step beyond today’s efficiency. Which work disappears? Which exposure disappears with it? Does the employee still inspect the underlying case, see the correction, and understand why the recommendation changed? Is there a point where they must form an unaided view before seeing the machine’s answer?

Some apprenticeships will become better. AI can supply examples, immediate feedback, and access to stronger patterns that were once unevenly distributed. Other pathways may become thin if people approve polished output without building a model of the work beneath it. The evidence does not justify assuming either future everywhere.

That leaves the next question for owners and boards: when AI changes who carries the work, has the company reduced dependence on one expert—or moved dependence into a system that fewer people can judge?