#THINKFUTURE
Series: Building the Future on a Cracked Foundation
About this series
Artificial intelligence has become one of the defining leadership topics of our time. Organisations are investing billions in technologies that promise greater productivity, lower costs and faster decision-making. Those investments matter, but this series begins with a different question:
What if the greatest challenge of AI is not how it changes work, but how it changes the way organisations develop capable people?
In this THINK FUTURE series, Building the Future on a Cracked Foundation, we explore that question through the lenses of research, organisational psychology and practical leadership. Each essay challenges a common assumption and invites leaders to think differently about capability, work and organisational success.
The presentation nobody could challenge
Imagine a management team reviewing a strategic proposal prepared by one of the organisation’s younger professionals.
The presentation is impressive. It contains a detailed analysis of the market, a summary of relevant trends, comparisons with competitors and several recommendations. The reasoning appears coherent, the language is confident and the presentation anticipates many of the questions the executives might ask.
The employee prepared it with the assistance of generative AI.
At first, this seems like an example of precisely what organisations hope AI will achieve. A relatively inexperienced employee has produced high-quality work in a fraction of the time that might previously have been required. The employee appears more capable, the organisation moves faster and senior managers have a strong foundation from which to make a decision.
Then one executive asks why a particular recommendation should be trusted.
The employee repeats the argument in the presentation but cannot explain the assumptions behind it. When asked what evidence might contradict the recommendation, the employee is unsure. A second executive points out that one comparison does not fit the organisation’s regulatory environment. The employee had not noticed the difference.
The presentation contained relevant information and produced the appearance of expertise. What it did not reveal was whether the person presenting it understood the boundaries, weaknesses and consequences of the advice.
The scenario is hypothetical, but it raises a question that will become increasingly important as AI enters professional work:
If people can produce expert-looking answers, how will leaders know whether expertise has genuinely developed?
The assumption behind easier access
Organisations have long invested in improving access to information. Knowledge repositories, search engines, digital learning platforms and decision-support tools all reflect a reasonable belief: people make better decisions when they can obtain better information.
Generative AI appears to take this progression much further. It does not merely retrieve documents. It can summarise them, compare alternatives, explain unfamiliar concepts, generate recommendations and present its conclusions in polished professional language.
This creates a powerful but potentially misleading inference. If an employee has immediate access to the information and reasoning that an expert might use, it can appear that the gap between the employee and the expert has narrowed substantially.
In terms of assisted output, it may have narrowed. In terms of expertise, the answer is less certain.
Expertise is not simply the possession of more information. An expert must interpret information in context, recognise which details matter, identify when familiar patterns do not apply and judge the limitations of the available evidence. Experts also need to recognise the boundaries of their own knowledge and understand when a decision requires further investigation, consultation or caution.
AI can give a person access to possible answers. It cannot automatically give that person the experience required to know which answer can be trusted.
Information is an input, not a capability
Information is essential to expertise, but it is not equivalent to expertise.
A person can read extensively about negotiation without being able to recognise the moment when a discussion is approaching an impasse. A manager can memorise models of organisational change without knowing which resistance signals indicate a deeper loss of trust. A graduate engineer can access every formula required for a calculation while lacking the practical judgment to recognise that the result is implausible.
The difference becomes particularly visible when circumstances depart from what is normal.
Novices often focus on the information immediately in front of them. Experienced professionals are more likely to interpret that information as part of a broader pattern formed through repeated exposure to relevant situations. Their knowledge is not merely larger. It is organised differently and connected to cues, consequences and possible courses of action.
Research into expert performance has repeatedly examined these differences between novices and experts. Expertise involves domain-specific knowledge, but it also draws on pattern recognition, practical and tacit knowledge, feedback, sustained practice and the ability to make discriminations that may remain invisible to someone with less experience. The Cambridge Handbook of Expertise and Expert Performance describes expertise as situated within particular domains and contexts rather than as the simple accumulation of general information. Its treatment of professional judgment also shows why much of the knowledge supporting expertise may be difficult to articulate explicitly.
This does not mean that expertise is mysterious or that experience always produces good judgement. Experienced people can become overconfident, repeat outdated practices or develop poor habits. Practice itself is not enough. Expertise requires opportunities to act, receive credible feedback, correct errors and gradually engage with more demanding situations.
Information contributes to this process, but it cannot substitute for the whole process.
The knowledge people struggle to explain
Organisations tend to value knowledge that can be documented. Processes can be mapped, guidance can be written and lessons can be stored in searchable systems. These activities matter because they make valuable knowledge more accessible and reduce unnecessary dependence on particular individuals.
Yet not all professional knowledge can be fully converted into instructions.
An experienced interviewer may sense that a candidate’s answer requires further exploration. A senior project manager may recognise that a technically accurate progress report is concealing a troubled project. An experienced coach may notice that the issue being discussed is not the issue preventing progress. A specialist may recognise that a familiar solution is inappropriate because one contextual detail has changed.
The expert may be able to explain the judgement afterwards, but the initial recognition often depends on patterns developed through experience. Research on practical intelligence describes tacit knowledge as knowledge about how to act effectively in real situations that is often acquired through experience rather than direct instruction. Studies comparing novices and experts have found differences in both the amount and organisation of this practical knowledge.
AI may make codified information easier to obtain, but ease of access does not eliminate the need for contextual interpretation. In some situations, it may increase that need because the employee now receives more content, more alternatives and more confidently expressed recommendations than before.
The organisation’s challenge is no longer simply whether employees can find information. It is whether they can distinguish what is relevant, question what appears credible and recognise what is missing.
When performance conceals understanding
Generative AI creates an additional difficulty because the quality of an output may no longer reveal the capability of the person producing it.
Historically, a manager could learn something about an employee’s reasoning by reviewing that person’s work. Weak assumptions, poor structure or incomplete analysis helped identify where guidance and development were required. A polished answer usually suggested that some level of competence had been developed, even if further verification was necessary.
AI weakens that connection.
An inexperienced employee can produce a fluent answer, persuasive analysis or professional report with limited understanding of the subject. This can be beneficial when the purpose is to improve communication or help the person begin an unfamiliar task. It becomes dangerous when the quality of the artefact is treated as evidence of the employee’s independent capability.
The distinction is between AI-assisted performance and human capability.
A person may perform well because the combination of the human and the technology is effective. That is a legitimate and valuable form of performance. Organisations do not need employees to work without tools merely to prove that they can.
However, the organisation still needs to know what the human contributes to the combined system. Can the employee frame the problem correctly? Can the person detect an unreliable assumption, recognise an exception, question the AI’s recommendation and take responsibility for the consequences?
These capabilities become more important, not less important, when the technology produces convincing outputs.
AI’s uneven frontier
The difficulty is compounded by the uneven nature of AI performance.
AI can perform remarkably well on one task and fail unexpectedly on another that appears similar. Its limitations are not always obvious to the person using it. This makes reliance difficult to calibrate, particularly for employees who lack enough domain knowledge to recognise when the technology has crossed from a strong area into a weak one.
Research involving consultants described this as a “jagged technological frontier.” In a field experiment, generative AI improved speed and quality on tasks within its capabilities. On a task outside that frontier, consultants using AI were less likely to reach the correct solution than participants who worked without it. The technology could therefore operate as both a performance enhancer and a source of error, depending on whether users recognised the boundary of its competence. The research was subsequently published in Organization Science.
The important implication is not that people should distrust AI by default. It is that effective human–AI collaboration requires calibrated trust.
People need to know when to rely on the technology, when to verify it, and when to reject its advice. This is partly a matter of AI literacy, but it is also a matter of domain expertise. A person who does not understand the field may be poorly positioned to identify whether the AI’s answer is sensible, incomplete, or dangerously wrong.
The paradox is difficult to ignore:
The people who benefit most from AI’s access to expertise may sometimes be the least able to recognise when its expert-looking output cannot be trusted.
The risk of borrowed confidence
Fluent language creates an impression of understanding. A carefully structured answer, supported by plausible explanations, can make both the creator and the reader more confident in the work.
The confidence may belong to the output rather than to the person.
This creates the possibility of borrowed confidence. Employees may feel more capable because the combined human and AI system produces impressive work, even when their ability to evaluate that work has not developed to the same degree.
Metacognition becomes central here. Metacognition includes the ability to monitor one’s thinking, assess the reliability of one’s understanding and recognise when further evidence or assistance is required. It allows professionals to ask not only, “What do I know?” but also, “How confident should I be that I know it?”
A 2025 study of 319 knowledge workers examined how they perceived critical thinking while using generative AI. Participants described verifying information, assessing outputs and adapting responses as important elements of their work. The study also found that greater confidence in AI’s ability to complete a task was associated with less reported critical-thinking effort. Because this was a survey of workers’ perceptions rather than a direct test of retained expertise, its conclusions should be treated carefully. It nevertheless illustrates why confidence in the tool can influence how actively people scrutinise its output.
Research on automation bias provides an older and broader warning. A systematic review found automation bias to be a recurring effect across fields, even though automated decision support often improved overall performance. Users can omit necessary actions when a system fails to prompt them or follow incorrect automated recommendations despite contradictory information.
This evidence does not justify the conclusion that people inevitably surrender their judgment to technology. It shows something more useful: improved system-level performance can coexist with new forms of human error.
The more convincing the system becomes, the more important it is for people to understand what would justify disagreeing with it.
Assistance is not the same as learning
Article 2 distinguished between producing better work with support and developing capability that endures when support is removed. That distinction becomes even more important when organisations attempt to assess expertise.
Evidence from education shows how easily assisted performance can be mistaken for learning. In a large field experiment, secondary-school students using a general-purpose generative AI interface performed better during supported mathematics practice but performed worse when the assistance was removed. Students using a specially designed AI tutor, intended to guide rather than simply answer, largely avoided the negative effect. The findings cannot be transferred directly from school mathematics to professional work, but they demonstrate that the design of assistance matters.
AI can accelerate learning when it provides explanations, asks questions, gives feedback and helps learners compare possible approaches. It can also bypass learning when it removes the need to retrieve knowledge, construct an argument, struggle with ambiguity or learn from an error.
Cognitive offloading is not inherently harmful. Professionals have always used external tools to reduce unnecessary mental effort. Calculators, checklists, databases and written records allow people to devote attention to higher-value questions. The issue is not whether thinking is offloaded, but which thinking is offloaded and what the person must still be capable of doing.
A useful tool reduces effort that no longer adds value. A poorly designed dependency may remove the effort through which understanding and judgment were meant to develop.
Expertise in an AI-enabled organisation
The answer is not to require people to reproduce everything AI can do. Expertise should not be defined as the ability to work exactly as professionals worked before generative AI.
AI changes what experts can accomplish. It can help experienced people consider more alternatives, locate relevant evidence, test assumptions and communicate complex ideas. It may also allow less experienced employees to engage with more advanced work earlier in their careers.
The opportunity is therefore larger than protecting existing expertise. Organisations can use AI to develop new forms of expertise in which human judgment and technological capability reinforce one another.
That requires leaders to distinguish among three questions:
- What can the employee produce with AI? This measures the performance of the combined human and technological system.
- What does the employee understand? This examines whether the person can explain the reasoning, assumptions, evidence, and limitations behind the output.
- What can the employee judge? This tests whether the person can recognise an exception, question the recommendation, compare consequences and decide when further expertise is needed.
These questions are related, but they are not interchangeable. An employee may perform well with AI while still requiring development in understanding or judgment. Another may possess strong domain expertise but need support in learning how to use AI effectively.
Organisations need both kinds of insight.
Redesigning how expertise is assessed
If leaders continue assessing capability mainly through completed outputs, they may struggle to distinguish human expertise from technological assistance.
This does not mean employees should be prohibited from using AI during assessment or work. In many roles, the relevant capability will include using AI well. The assessment must therefore examine both independent reasoning and effective collaboration with technology.
Leaders can begin by changing what they review.
Instead of evaluating only the final recommendation, they can ask the employee to explain how the problem was framed, what assumptions shaped the analysis and which evidence would alter the conclusion. Instead of asking whether AI was used, they can examine how its output was verified, challenged and improved. Instead of treating confidence as competence, they can test whether the employee recognises uncertainty and the limits of the available evidence.
Developmental work can also be designed to make thinking visible. Employees can compare an AI-generated answer with alternative evidence, identify weaknesses intentionally placed in a scenario, explain why a recommendation would fail under different conditions or defend a decision before experienced colleagues.
Such practices do more than detect mistakes. They create the feedback, reflection and accountability through which expertise grows.
The purpose is not to prove that humans are better than machines. It is to ensure that the human contribution to the combined system remains visible, assessable and capable of development.
A different leadership conversation
The first essay in this series asked what capability an organisation might stop developing when AI changes work. The second argued that work produces both immediate results and the people required for future performance.
This third essay adds a further distinction. Better access to information can improve performance, but it does not automatically produce expertise. A high-quality output can demonstrate effective use of AI without demonstrating that the employee understands the answer or can judge its reliability.
Leaders must therefore resist two opposite mistakes.
The first is to dismiss AI-assisted work as somehow less legitimate. Human beings have always extended their capabilities through tools, other people and accumulated knowledge. AI can help professionals perform better and may create powerful new learning opportunities.
The second is to assume that improved output proves that human capability has improved by the same amount. that the technology cannot safely make on their behalf.
The future organisation will not succeed because every employee stores more information internally than an AI system can access. It will succeed because its people can frame the right problem, evaluate evidence, recognise uncertainty, challenge plausible answers and accept accountability for consequential decisions.
Those capabilities cannot be inferred from fluency alone.
THINK FUTURE Principle
Information makes answers accessible. Expertise determines which answers can be trusted, when they apply and what their consequences may be.
AI can extend what people know and what they can accomplish. The leadership responsibility is to ensure that access to better answers strengthens rather than conceals the development of human judgement.
As AI makes expert-looking work available to almost everyone, leaders should ask:
What evidence would show that our people are not merely producing better answers, but becoming more capable of judging when those answers are wrong?
Coming next
Can AI Create Experts?
If access to information does not automatically create expertise, could artificial intelligence nevertheless help people develop expertise more quickly?
Emerging research suggests that it can, but not under every condition. AI may provide rapid feedback, personalised guidance, simulated practice and earlier access to complex work. It may also allow people to bypass the thinking, struggle and reflection through which judgment develops.
The difference may depend less on the technology itself than on how organisations design the learning around it.
In the next essay, we examine a question with profound implications for professional development:
Can AI help people become experts, or does it merely help them perform like experts while the assistance remains available?