#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 employee who became productive too quickly
Imagine an organisation introducing an AI assistant into the work of its junior professionals.
The results appear almost immediately. Reports that once took several days are completed in hours. Employees produce clearer analyses, respond to technical questions more confidently and engage with work that would previously have been reserved for more experienced colleagues. Managers spend less time correcting basic mistakes, and clients receive answers faster.
One employee stands out. Within months, she is producing work that resembles the output of colleagues with several years more experience. Her manager concludes that AI has accelerated her development.
Then the AI assistant becomes temporarily unavailable during an important client engagement.
Without it, the employee struggles to structure the problem. She remembers the recommendation the system would normally generate, but cannot reconstruct the reasoning behind it. She overlooks an unusual feature of the client’s situation and does not recognise that the standard solution could create a different risk.
Her previous performance was real. The work she produced with AI was valuable. Yet the organisation had mistaken an improvement in the performance of the combined human and AI system for an equivalent improvement in the employee’s underlying capability.
This is a hypothetical scenario, but it exposes a question that many organisations may soon need to answer:
Can AI help people become experts, or can it merely help them perform like experts while assistance remains available?
Productivity is not proof of development
The assumption is understandable. When people complete more work, solve more difficult problems and produce higher-quality outputs, it appears reasonable to conclude that they are becoming more capable.
AI can undoubtedly improve performance.
A large field study involving 5,172 customer-support agents found that access to a generative AI assistant increased the number of issues resolved per hour by an average of 15 per cent. The benefits were greater among less experienced and lower-performing workers. The researchers also found evidence suggesting that AI helped diffuse some of the practices used by higher-performing agents. Less experienced workers improved more rapidly, while customers interacted more positively with them. demonstrates that AI can do more than save time. It may help people access useful knowledge that was previously acquired slowly and inconsistently.
That is significant. If AI can make effective practices available at the moment they are needed, employees may no longer have to wait for an experienced colleague to provide guidance. They may encounter more examples, receive more immediate support and reach acceptable performance sooner.
Yet performance and development are not interchangeable.
Performance describes what a person can accomplish under particular conditions. Capability describes what the person has developed the understanding and judgement to accomplish across changing conditions. Expertise requires more than producing the correct answer when assistance is available. It includes recognising which problem is being solved, understanding why an approach works, detecting exceptions and adapting when established patterns no longer apply.
An AI-supported employee may genuinely perform better. The unanswered question is how much of that improved performance remains within the person.
Expertise develops through more than exposure to answers
Essay 3 distinguished information from expertise. Information makes possible answers available. Expertise helps a person determine which answer can be trusted, when it applies and what its consequences may be.
This distinction explains why giving someone expert information does not automatically make that person an expert.
Expertise develops through sustained engagement with a domain. People build knowledge, apply it to varied situations, make predictions, encounter errors, receive feedback and adjust their understanding. Over time, they become better able to recognise meaningful patterns, distinguish important signals from distracting detail and respond when a familiar solution does not fit.
Practice matters, but repetition alone is insufficient. Deliberate practice research emphasises focused activities designed to address weaknesses, supported by feedback and repeated opportunities for improvement. Reviews in professional and clinical education similarly identify feedback, reflection and progressively demanding practice as important conditions for skill development. Duvivier and colleagues’ review of deliberate practice describes this as sustained practice directed towards identified weaknesses rather than the simple accumulation of experience.
The developing professional must do something with the feedback. The person must connect the correction to the underlying problem, revise a mental model and apply the learning when the situation changes.
AI may strengthen this process. It can also interrupt it.
The difference depends on whether AI participates as part of the learning process or removes the need for that process.
The case for AI as a capability accelerator
There are good reasons to believe that AI could help people develop expertise more quickly.
An effective tutor does not merely provide information. It identifies where a learner is struggling, asks questions, offers explanations, adjusts the level of difficulty and provides feedback while the experience is still fresh. In workplaces, managers and experienced professionals often struggle to provide this level of individual attention consistently.
AI could make some of these developmental conditions more widely available.
A junior employee could use AI to explore several ways of framing a problem before selecting one. A trainee could practise difficult client conversations in a simulation and receive immediate feedback. A developing manager could examine a decision from multiple stakeholder perspectives. An engineer could compare alternative explanations for an unexpected result. A professional could request examples that become progressively more complex as understanding improves.
Evidence from education illustrates this potential. In a randomised controlled trial involving 194 university students, a purpose-built AI tutor produced greater short-term learning gains in an undergraduate physics course than an active-learning classroom condition, while students generally spent less time on the material. The researchers designed the tutor around established educational practices, including active engagement, appropriate scaffolding, targeted feedback, self-pacing and controls intended to improve the accuracy of explanations. Kestin and colleagues’ 2025 study in Scientific Reports suggests that AI can support meaningful learning when its behaviour is deliberately designed around how people learn.
The qualification is crucial. The study did not test unrestricted access to a general-purpose chatbot. It tested a structured tutor designed to guide learning.
Other AI-enabled learning systems may offer similar advantages:
- Immediate feedback: People can correct misunderstandings before those misunderstandings become established habits.
- Personalised practice: The difficulty, pace and type of support can be adjusted to the learner’s needs.
- Greater variety: Employees can encounter more cases, contexts and exceptions than everyday work may naturally provide.
- Safe experimentation: People can test decisions and experience simulated consequences without exposing clients, colleagues or the organisation to unnecessary risk.
- Reflection on demand: AI can ask employees to explain their reasoning, compare alternatives and identify what evidence would change their conclusions.
- Earlier access to complexity: Junior employees can engage with demanding work while receiving support appropriate to their current level of development.
These advantages could make AI an important contributor to expertise development. They could also help organisations reduce their dependence on whether a particular manager has the time, inclination and ability to teach.
However, availability is not the same as learning design. An AI system capable of explaining an answer is not necessarily being used in a way that requires the employee to understand it.
When assistance removes the learning
The same technology that creates new opportunities for learning can also make learning easier to bypass.
Consider the difference between two employees using AI to prepare a recommendation.
The first employee defines the problem, develops an initial view and asks the AI to challenge it. The employee compares sources, examines conflicting explanations, tests the recommendation under different conditions and records why the final decision differs from the AI’s first suggestion.
The second employee enters the assignment, requests a completed recommendation and edits the wording.
Both may submit strong work. Their learning experiences are not equivalent.
The first employee is using AI to increase the range and quality of thinking. The second is using it to replace much of the thinking. The visible output may conceal this difference.
Research in education provides a particularly clear warning. In a large field experiment involving nearly 1,000 secondary-school mathematics students, access to a general-purpose GPT-based interface improved performance while the system was available. When access was removed, however, those students performed worse than the control group. A more carefully designed AI tutor that provided safeguards and guided support largely avoided this negative effect. Bastani and colleagues’ 2025 study in Proceedings of the National Academy of Sciences concluded that generative AI can improve practice performance while inhibiting learning when it functions as a substitute for the cognitive work students need to perform.
This was an educational study, not a workplace study. It does not prove that employees will experience the same outcome. It does, however, establish an important distinction: better assisted performance can coexist with weaker performance after assistance is removed.
The lesson for organisations is not that AI should regularly be removed to test whether employees can cope without it. Modern expertise has always included the effective use of tools. The lesson is that organisations cannot infer retained capability from assisted output alone.
If AI performs the retrieval, problem formulation, analysis and evaluation, the employee may gain an answer without developing the mental structures needed to judge that answer later.
Not every struggle deserves to be preserved
Concerns about cognitive offloading can easily lead to the opposite mistake: assuming that people must continue performing every difficult or repetitive task manually because effort is inherently developmental.
That would be a poor principle for work design.
Professionals have always used tools to reduce unnecessary cognitive effort. Written records reduce dependence on memory. Calculators improve computational reliability. Checklists protect against predictable omissions. Databases make information easier to retrieve. These tools do not necessarily diminish expertise. They can free people to concentrate on interpretation, judgement and consequences.
The relevant issue is not whether AI makes work easier. It is which effort it removes.
Some effort is merely friction. Searching through disconnected systems for a document may consume time without developing meaningful capability. Reformatting a report repeatedly may add neither operational nor developmental value. Automating such work can improve both productivity and the quality of employees’ working lives.
Other effort carries learning. Constructing an initial argument exposes gaps in understanding. Retrieving foundational knowledge strengthens the ability to use it. Comparing competing explanations develops discrimination. Making a prediction before receiving an answer allows feedback to correct the learner’s mental model. Taking responsibility for a decision connects reasoning to consequences.
When AI removes friction, it can create space for development. When it removes the thinking through which understanding is formed, it can weaken development.
The objective is therefore not to preserve struggle indiscriminately. It is to preserve, redesign or strengthen the cognitive and experiential activity that produces learning.
Experts must learn to work with an uneven technology
AI does not perform consistently across every task. It can produce excellent results in one area and fail unexpectedly in another that appears superficially similar.
A field experiment involving 758 consultants found that generative AI improved performance on tasks within the technology’s capabilities. Participants completed more tasks, worked faster and produced higher-quality results. On a task outside that capability frontier, however, consultants using AI were less likely to reach the correct conclusion than those who worked without it. The researchers describe this uneven pattern as a “jagged technological frontier”. Dell’Acqua and colleagues’ research in Organization Science shows why expertise cannot be reduced to proficiency in using an AI tool.
Future expertise will include knowing how to collaborate with AI. It will also include recognising when that collaboration is producing an unreliable result.
This creates a developmental challenge. Novices may benefit substantially from AI because it gives them access to knowledge and practices associated with more experienced performers. Yet those same novices may lack the domain knowledge required to recognise when the AI has crossed the boundary between plausible assistance and misleading advice.
The solution cannot be limited to generic AI literacy. Employees need sufficient knowledge of both the technology and the professional domain to calibrate their trust. They must learn what to delegate, what to verify and what should remain subject to accountable human judgement.
AI can help teach these distinctions, but only if organisations deliberately expose employees to the technology’s limitations. If employees encounter only successful AI-supported work, they may become skilled users without becoming discerning collaborators.
The risk of invisible developmental debt
When AI improves output immediately, weaknesses in learning may remain hidden.
An employee can continue producing acceptable work. The manager can continue approving it. Clients may remain satisfied. Nothing appears to be wrong until the employee confronts an unfamiliar situation, assumes greater responsibility or must evaluate an AI-generated recommendation with significant consequences.
By that stage, the missing development may have accumulated over several years.
This can be understood as developmental debt. An organisation receives the current productivity benefit of AI-assisted work while postponing the investment required to build independent understanding and judgement. Like other forms of deferred investment, the cost becomes visible later, when the organisation needs people capable of diagnosing novel problems, supervising others or making decisions beyond the reliable reach of established systems.
Developmental debt is an editorial description, not a new ECCSA model or approved theoretical construct. It identifies a practical risk that leaders should investigate rather than an established measure.
The danger is greatest when leaders use output quality as the main evidence of readiness. An employee may appear prepared for more senior responsibility because the person’s work has improved. If the organisation has not examined how the work was produced, it may not know whether the employee can frame problems, test assumptions, recognise anomalies or explain why a recommendation should be trusted.
AI can conceal a developmental gap at the same time that it helps close a performance gap.
Designing AI to develop expertise
The central issue is not whether organisations should use AI in learning. It is how they should design work so that AI strengthens the mechanisms through which expertise develops.
That begins by distinguishing between assistance intended to produce an output and assistance intended to develop a person. The same AI system can serve either purpose, but the interaction should not be designed or assessed in the same way.
When expertise development matters, leaders can build several conditions into the work.
Require a first position before assistance
Employees can be asked to frame the problem, make an initial prediction or identify possible options before consulting AI. This gives the learner something against which to compare the response and makes misunderstandings visible.
The intention is not to delay work unnecessarily. It is to prevent the AI’s first answer from becoming the employee’s first thought.
Use AI to question, not only to answer
AI can challenge assumptions, generate counterarguments, present alternative scenarios and ask what evidence would alter a conclusion. This shifts its role from answer generator to thinking partner.
A system designed to ask well-sequenced questions may support learning more effectively than one designed primarily to complete the task.
Make reasoning observable
Managers need access to more than the final output. Employees can explain how they framed the problem, where the AI influenced the analysis, which suggestions they rejected and how they verified important claims.
This makes coaching more precise and allows managers to distinguish a communication weakness from a gap in underlying judgement.
Retain human feedback and accountability
AI feedback can be immediate and scalable, but it does not remove the value of experienced human judgement. Managers and specialists understand organisational history, professional standards, relationships and consequences that may not be represented adequately in the system.
Human review is particularly important as employees assume greater responsibility. Expertise develops partly through becoming accountable for decisions, not merely through receiving technically accurate explanations.
Vary the conditions
Practice should include familiar problems, unusual cases, incomplete information and situations in which the AI gives a plausible but flawed recommendation. Employees need opportunities to learn when established patterns apply and when they do not.
Exposure to exceptions helps develop the discrimination that separates competent routine performance from adaptive expertise.
Test retention and transfer
Organisations should examine whether employees can apply their learning to a different situation, explain the underlying principle and recognise when it is inappropriate. This matters more than whether they can reproduce a particular answer without AI.
The relevant question is not, “Can the employee work without technology?” It is, “Has the employee developed knowledge and judgement that improve performance across changing conditions?”
Reduce support progressively
As capability develops, the nature of AI assistance can change. Early support may include explanations, worked examples and structured prompts. Later support can become less directive, requiring employees to make more decisions and defend their reasoning.
The aim is not necessarily to withdraw AI completely. It is to prevent permanent dependence on a level of scaffolding that no longer develops the person.
Measure both performance and development
Most organisations will measure the effect of AI through productivity, quality, speed and cost. These indicators are necessary, but they do not reveal whether expertise is developing.
Leaders need a second set of questions:
- Can employees explain the reasoning behind their recommendations?
- Can they recognise when a familiar solution does not fit the context?
- Do they test AI-generated claims against credible evidence?
- Can they identify uncertainty and state the limits of what is known?
- Can they transfer what they have learned to unfamiliar situations?
- Are they becoming capable of handling progressively greater responsibility?
- Do they know when a decision requires more experienced human judgement?
These questions do not require employees to prove that they can outperform AI. They help the organisation understand what human capability is developing within an AI-enabled system.
This distinction also protects organisations from drawing the wrong conclusion about employees. A person who uses AI extensively may be developing strong expertise if the technology supports deliberate practice, feedback and reflection. Another person may produce equally polished work while learning very little.
The amount of AI use is therefore a poor proxy for development. The quality of the learning interaction matters more.
A different leadership conversation
Can AI create experts?
The evidence does not support a simple yes or no.
AI can give people access to personalised guidance, rapid feedback, varied practice and complex problems earlier in their development. It may diffuse effective practices and reduce the time employees need to reach useful levels of performance. Purpose-built AI tutors show that carefully designed systems can produce meaningful learning gains.
AI can also provide answers before people have formed their own understanding. It can improve performance while assistance is available without creating capability that transfers when the context changes. It can make an employee appear more expert while concealing a weak ability to recognise when the system is wrong.
The technology does not determine which outcome occurs on its own.
Learning design matters. Work design matters. Feedback matters. Reflection matters. Progressive responsibility matters. Human coaching and professional accountability still matter.
AI will not create experts simply because organisations make it available. It can help accelerate expertise when it strengthens the experiences through which knowledge becomes understanding and understanding becomes judgement.
That requires leaders to stop asking only how AI can help employees perform more work. They must also ask what employees are learning while that work is performed.
THINK FUTURE Principle
AI can accelerate expertise when it strengthens the experiences through which understanding and judgement develop. It can conceal weak development when it merely supplies the performance.
This principle is an editorial expression of the essay’s central argument. If AI changes how people access knowledge, practise skills, receive feedback and perform professional work, the leadership question is no longer simply whether learning will improve.
It is:
If AI changes learning, what should organisations redesign?
Coming next
Have We Reached the End of the Career Ladder?
Expertise develops through progressively demanding experience. For generations, graduate programmes, junior roles and early-career work have provided the first stages of that development.
AI may now complete much of the work that once gave inexperienced people their entry into a profession. Organisations may consequently need fewer employees at the bottom of the traditional hierarchy, even while they continue to depend on experienced specialists, managers and leaders in the middle.
In the next essay, we examine a question that reaches beyond the future of entry-level employment:
If the first rung disappears, how do people climb?