AI Makes Novices Faster, but It Does Not Make Them Experts

Imagine an employee who, only a month ago, could not handle a difficult customer question without help. They searched for information, struggled with wording and repeatedly called a manager over. Now they use generative AI and produce a confident, well-structured response within minutes.

From the outside, the company appears to have made a technological leap. Productivity is up. The new hire is closing the gap with the strongest people on the team. The manager no longer has to rewrite every message.

Then the problems surface. One answer confuses the terms of a contract. Another promises something the company cannot deliver. A third is impeccable except for the part that matters most: the recommendation does not apply to this particular situation.

AI has made the employee faster. It has not made them more experienced.

That is one of the central paradoxes of adopting AI in a business. A model can quickly give someone the visible signals of competence: professional language, familiar arguments and a clean decision structure. It cannot transfer responsibility, pattern recognition or the judgment to know when a standard answer should not be used.

Who benefits most from AI?

One of the clearest answers comes from Generative AI at Work, a study of 5,179 customer-support agents, some of whom received access to a generative AI assistant.

Productivity rose by 14% on average. That headline figure concealed a substantial difference between workers. The improvement for novice and lower-performing agents reached roughly 34%, while the most experienced agents benefited far less.

The researchers suggested that the system was spreading the practices of high performers through the organisation. It learned from previous conversations and offered newer agents wording and approaches that an experienced colleague might have produced without assistance.

In effect, AI became a mechanism for transferring the parts of expertise that could be formalised. A new employee no longer had to spend months collecting useful response patterns. The system supplied them in the middle of the conversation.

A similar pattern appeared in software development. Across three field experiments involving 4,867 developers, access to an AI coding assistant increased completed tasks by an average of 26.08%. Less experienced developers adopted the tool more often and experienced larger productivity gains.

For a company, this is a compelling result. AI can shorten onboarding, reduce the cost of routine work and make some of the strongest employees' methods available across the team.

It does not follow that a model turns a novice into an expert. It allows a novice to reproduce part of expert behaviour sooner.

Borrowed competence

Real expertise is not simply a collection of correct actions. It includes the ability to recognise when the usual action is no longer appropriate.

An experienced salesperson notices when a price objection is really fear of making a decision. A manager sees that a formally successful project depends on one exhausted employee. A founder recognises that an investor's offer is attractive only until the control provisions are taken into account.

This kind of knowledge is difficult to convert into an instruction. It grows from context, earlier mistakes, weak signals and the consequences of decisions for which someone once had to take responsibility.

AI can produce a persuasive answer, but it does not bear the consequences of using it. It does not know the company as the founder does. It cannot feel the relationships inside a team, and it cannot see the full history of a negotiation unless that history has been provided explicitly.

This creates a dangerous illusion. Inexperience used to be visible: a new employee hesitated, asked questions and wrote unevenly. Now the same person's mistake can look like a professional assessment. AI does not merely accelerate the answer. It can make that answer more convincing whether it is right or wrong.

The jagged frontier

An experiment with 758 Boston Consulting Group consultants captured this problem particularly well. Its results were published in Organization Science in 2026.

The researchers described a “jagged technological frontier.” Generative AI does not have a simple boundary with easy tasks on one side and difficult tasks on the other. A model may perform impressively on creative and analytical work, then fail on another task that appears similar or even simpler to a human.

For tasks inside the frontier, participants using AI completed 12.2% more work and were 25.1% faster on average. The quality of their solutions also improved.

For a complex management task outside that frontier, however, participants with AI were 19% less likely to reach the correct answer than those working without it. The consultants sometimes followed a convincing suggestion and failed to notice that it was leading them in the wrong direction.

The central problem is that the frontier is invisible. A model does not show a red warning when it becomes unreliable. It continues to answer in the same calm, confident voice.

This is where experience retains its value. An expert is not someone who always knows the answer. An expert is more likely to recognise the moment when a ready-made answer should not be trusted.

When one person with AI works like a team

The field experiment The Cybernetic Teammate involved 776 professionals at Procter & Gamble. The researchers compared individuals and teams working with and without AI.

Individuals using AI could produce results comparable to teams working without it. The technology also helped people cross functional boundaries: technical specialists developed stronger commercial proposals, while commercial specialists handled the technical part of the task more effectively.

For a founder, the opportunity is significant. A small company can perform work that previously required several specialists. Employees can reach into adjacent disciplines, and the business becomes less constrained by rigid functional divisions.

Entering another professional territory, however, is not the same as operating safely within it. AI may help a marketer write code or a developer prepare a commercial offer. The reliability of that code and the realism of that offer still need to be assessed by someone who understands the consequences of failure.

The further AI extends an employee's reach, the more important independent review becomes.

When convenience replaces thought

For The Impact of Generative AI on Critical Thinking, presented at CHI 2025, researchers surveyed 319 knowledge workers and analysed 936 examples of generative AI use at work.

Higher confidence in AI was associated with less self-reported effort spent critically evaluating the result. Greater confidence in one's own ability to perform the task, by contrast, was associated with more critical engagement.

The limitation matters: this was a self-report study, so it does not prove that using AI directly weakens critical thinking. It does reveal a familiar organisational risk. The more convenient the tool and the more persuasive its answers, the easier it becomes to stop examining the foundations of a decision.

New employees are especially exposed. They do not yet have the internal criteria required to distinguish a strong answer from a merely plausible one. They may miss an error not because they are careless, but because they do not yet know where such errors tend to hide.

How to adopt AI without creating an illusion of expertise

Training employees to work with AI should involve more than collecting effective prompts. A team needs to understand task categories, levels of risk and the depth of review each result requires.

When an error is easy to detect and reverse, AI can be given more room. A draft email, a document summary or several headline options do not require the same control as a financial model, a contract term or advice that could affect people's health and safety.

It also helps to separate producing an answer from accepting it. An employee may use AI to prepare a result, but the quality criteria should exist independently of the model. Otherwise, the system is effectively being asked to mark its own work.

Expert feedback must remain visible. If a senior specialist silently corrects every error, the system may become more productive without the team becoming more capable. A junior employee needs to understand why the answer changed, which signal they missed and under what conditions the original recommendation might have been acceptable.

Finally, companies need to measure more than speed. Discussions about AI and employee productivity easily collapse into counts of resolved tickets, completed texts or finished tasks. Higher output can coexist with errors that only become visible later. Responsible AI adoption therefore needs quality measures: rework, returns, corrections and the consequences of the decisions being accelerated.

AI for business creates value when a company understands exactly what it is accelerating and who remains accountable for the outcome.

The expert's new role

As AI spreads, the expert's value moves away from producing every individual answer and towards designing the system in which decisions are made.

The expert frames the problem, defines standards, notices exceptions and decides where automation must stop. The manager creates an environment in which speed is not mistaken for quality. A mentor helps a founder see consequences that cannot be reconstructed from instructions and public data alone.

Speaking to an experienced person is therefore not an alternative to using AI. A founder can use a model to generate options, then test the decision with someone who has already faced a comparable situation. Through United Mentors, founders can speak with experienced entrepreneurs and operators who understand not only how a sound recommendation should be phrased, but what a poor decision can cost in practice.

AI makes knowledge more accessible. A new employee can become useful sooner, and a small team can take on work that recently required an entire department.

Knowing an answer and being able to take responsibility for it remain different capabilities.

AI shortens the journey from a blank page to a plausible solution. Expertise is needed on the next stretch, where somebody has to decide whether that solution can be trusted with the company's money, people and future.