The evidence argues against the easy story

The strongest workplace evidence does not support a simple claim that AI inevitably makes less-experienced employees less capable. Brynjolfsson, Li, and Raymond's 2025 study in the Quarterly Journal of Economics followed 5,172 customer-support agents using a generative-AI assistant. Productivity increased by 15% on average, with larger gains among less-experienced and lower-skilled employees. Workers with two months of tenure and AI assistance performed as well as or better than untreated workers with more than six months of tenure, across several outcomes. [2] AI accelerated the experience curve in that setting.

The distinction is important because the agents remained inside the work. They saw the recommendations, decided how to use them, and continued to encounter real customer problems, so AI could teach. But a different workflow is also possible. A young professional can receive a polished research summary, a plausible financial model, and fluent recommendations. None of it requires the ability to distinguish a strong answer from one that merely sounds strong.

Fluency can arrive before discrimination.

Judgment has to become a designed outcome

The proposition underneath this can be stated in one line. AI compresses the distance between novice and competent output. It does not compress the distance between novice and competent judgment.

Which raises a question managers can act on, and it is a question about quantity before it is a question about quality.

Judgment exposure is the number and consequence of situations in which a professional has to form a view, defend it, be corrected, and eventually own the result. Traditional apprenticeship generated it incidentally. Comparing forty sources produced a view about which evidence deserved weight. Rebuilding a model produced a view about which assumptions mattered.

The exposure was a by-product of the labor, which is why nobody had to design it. Remove the labor and the exposure goes with it unless somebody puts it back. That is the actual mechanism of the risk.

Four questions measure it for any individual.

• How often does this person have to form an independent view rather than assemble one?
• How often is that view challenged by someone who knows more?
• How often do they see the consequence of a decision they influenced?
• How often do they handle a case with no established answer?

A professional producing excellent output with low scores on all four is not developing, however good the work looks. That is harder for a manager to see than output quality, and it eventually determines who is capable of leading the organization.

Previous professional-development models relied heavily on repetition. A young analyst compared many sources and gradually learned which evidence deserved weight. An engineer rebuilt calculations and saw which assumptions changed the outcome. A finance professional reconciled data until inconsistencies became recognizable. Much of that manual effort can disappear without loss.

The question is what replaces the repeated practice of forming a view, having it challenged, correcting it, and eventually taking responsibility for it. AI-native apprenticeship should be built around progressive judgment rather than around preserving production tasks that no longer need doing. The employee should use AI from the beginning; what changes over time is the level of intellectual and consequential responsibility they carry.

The Judgement Ladder

The learning path we use, a judgment ladder, moves the professional progressively closer to ownership of the decision, with AI present at every step.

The employee does not graduate away from AI. They graduate toward stronger judgment while using it.

The manager's role changes too

If AI performs more first-pass production, senior professionals cannot assume that learning will occur automatically in the background, so managers have to expose their reasoning more deliberately. A research task can require the professional to rank sources rather than simply produce a summary. An AI-generated draft can become an exercise in defending the argument and identifying where it fails. A model can shift from construction to scenario design and assumption challenge. Meeting preparation can be replaced by greater participation in the meeting and a structured discussion afterward about what changed.

The productivity benefit remains while the learning experience changes. This can produce a better apprenticeship than the traditional model, because judgment becomes explicit rather than an accidental by-product of long hours of production.

Measure development differently

AI-native employees may reach competent output quality unusually quickly, which makes output itself a weaker proxy for development. A manager therefore needs to know whether the employee can explain the assumptions underneath an answer and distinguish reliable from weak evidence. They also need to know whether the person recognizes a plausible but incomplete output, handles an exception, and takes responsibility when there is no obvious answer. Those are harder measures than documents produced or hours completed, and they are closer to the capabilities the organization eventually needs from its managers, experts, and leaders.

The firms that adapt best will not protect inefficient junior work. They will allow AI to remove it while deliberately increasing the quality and frequency of the judgment experiences that replace it.

An AI-native workforce will not need less apprenticeship. It will need apprenticeship designed for a world in which the first answer is already easy to produce.

‍