If AI Helps New Employees Ramp Up Faster, Could It Weaken the Path to Expertise?

After AI Enters the Workflow · Season Two: “From Personal Tool to Organisational Capability” · Article 8

A new employee once needed months to learn how to answer difficult customer questions. Now an AI assistant draws on prior conversations and proposes replies. In the first week, a novice can produce work close to the style of experienced colleagues. Departmental productivity rises and training time falls.

A year later, managers observe another result. When AI cannot answer, sources conflict or a customer presents an exception, some employees cannot analyse the problem independently. They learned to select and edit suggestions without passing through the foundational work from which those suggestions were originally formed.

AI may shorten an experience curve. It may also remove the learning stages inside that curve. Organisations must distinguish earlier production of an acceptable result from earlier acquisition of independent judgement.

Existing evidence shows that novices can gain more

A study of 5,179 customer-support agents found that a generative AI assistant increased issues resolved per hour by 14 per cent on average. The increase was 34 per cent for novice and lower-skilled workers, while experienced and highly skilled workers saw little effect. The researchers suggested that the system could disseminate the practices of stronger workers and reported suggestive evidence of learning. NBER, “Generative AI at Work”

This is an important value of AI as a knowledge-diffusion tool. It can provide examples and advice at the moment of work, reducing the time a new employee must wait for expert help.

The study took place in a specific support environment. It does not demonstrate long-term learning in every occupation. Better performance with a suggestion does not necessarily mean equal capability when the suggestion is absent.

Expertise is more than knowing the answer

An expert completes standard work, but more importantly recognises when the standard does not apply. The expert notices missing facts, conflicting rules, stopping conditions and the people affected by error.

These abilities often form while performing lower-level work. Junior lawyers read many documents, programmers debug their own faulty code, and analysts reconcile raw data. Repetition can be inefficient, but it also exposes people to how evidence is produced, how errors recur and where systems break.

When AI supplies the finished artefact, the organisation may preserve output while removing diagnosis. A novice learns to judge whether an answer looks right without learning how to reconstruct it from source material.

Not every foundational task has educational value

Protecting a development pathway does not require romanticising mechanical labour. Copying formats, moving data manually and repeating searches may not build judgement at all. They can consume time better used for learning.

The key is to identify which stages form capability. Does the step expose the employee to important evidence? Does it train anomaly detection? Does it develop causal or rule-based understanding? Does it provide timely feedback? If not, automation may improve both efficiency and learning.

If the answer is yes, the organisation should not merely remove the stage. It should create a substitute: require novices to explain AI suggestions, compare alternatives, work through deliberately retained exceptions and complete some tasks independently in a low-risk setting.

Dependence changes the future supply of experts

A department may initially place a small group of experts above many AI-assisted novices. That increases short-term throughput. But if novices rarely exercise independent judgement, who becomes the next generation of supervisors?

This is a temporal problem of organisational capability. A “thin waist” can emerge: a few experts at the top and many system-dependent operators below, with too few roles through which intermediate expertise develops.

The OECD’s AI and Skills reports that skills shortages constrain adoption while AI increases demand for higher-level skills, digital capability, data interpretation, management and problem-solving. It stresses training linked to work context and lifelong learning. OECD, “AI and skills”

The OECD Skills Outlook 2025 also describes both complementarity for novice workers and possible reductions in entry-level tasks and opportunities. Training, consultation and workplace design influence which outcome develops. OECD Skills Outlook 2025

Design AI as a coach, not only an answer supplier

If capability formation is an objective, the interaction should change.

For lower-risk work, an employee can record an initial judgement before seeing the AI suggestion, then identify differences and evidence. Material recommendations should link to sources so the novice can see how the answer was formed. AI may generate counterexamples and questions, while a qualified mentor confirms important feedback.

Periodic independent exercises are not a prohibition on AI. They measure whether employees can work during system failure, unusual exceptions and high-risk cases. Assessment should include problem definition, evidence checking and escalation—not output volume alone.

The tacit knowledge of experienced staff should also become cases, failure modes and evaluation criteria. Otherwise, AI imitates historical outputs while the organisation still cannot explain why its experts made those choices.

Redesign junior roles rather than simply removing them

When AI handles routine work, junior roles can shift from pure execution towards supervised analysis: checking citations, investigating anomalies, maintaining test sets, comparing customer outcomes and recording new exceptions. These tasks contribute to the system while developing the judgement the organisation will later need.

OECD work on an AI-ready public workforce emphasises training tailored to work context, practical application, facilitation and a sustained learning environment. OECD, “Building an AI-ready public workforce”

One generic AI course cannot replace experience if the daily opportunities for learning have been automated away.

The cost of learning cannot be transferred entirely to workers

If an organisation raises AI-assisted output targets while expecting employees to learn tools, verify errors and preserve expertise in their own time, part of the apparent efficiency gain is a transfer of cost. Staff need formal learning time, mentor support and a safe environment for experimentation.

Managers should also examine who receives those opportunities. Experienced employees may retain access to complex cases while novices are confined to editing AI drafts. Contractors and part-time staff may use the tool without entering the meetings where judgement is formed. Unequal access to developmental work can widen internal capability gaps.

Training measures should move beyond course completion. Can the employee explain a recommendation, detect an error, handle an exception independently and escalate correctly? Learning is an operating cost of the AI system, not a private obligation that workers must absorb to remain employable.

An organisation that budgets for model access but not for this learning has understated the true cost of its AI capability.

That requires treating learning capacity as an outcome to be measured. Beyond speed and error rates, the organisation should observe whether newer staff can explain conclusions, detect deliberately seeded anomalies, perform critical steps without AI and assume more complex judgement over time. If immediate output rises while these capacities decline, the institution is consuming expertise accumulated in the past without replenishing it. Short-term efficiency cannot justify weakening the people who will have to review future systems and exceptions.

Conclusion: remove valueless labour, not the formation of judgement

Helping novices reach the current quality standard faster is a genuine benefit. The issue is not whether they physically perform every step. It is whether the organisation can still develop people who discover new problems, handle exceptions and supervise the system.

The rule should be:

Automate repetitive labour that does not produce learning. Whenever a stage forms professional judgement, replace it with cases, explanation, independent practice, mentor feedback and progressively greater responsibility.

Success should not be measured only by first-week output. It should include whether the employee can make an independent judgement a year later when AI is uncertain or wrong. Training costs saved today should not become a capability debt expressed as a shortage of experts tomorrow.

Primary sources and further reading

• NBER: Generative AI at Work
• OECD: AI and skills
• OECD Skills Outlook 2025
• OECD: Building an AI-ready public workforce

Continue reading: Explore the After AI Enters the Workflow series.