The business case for generative AI is becoming harder to dismiss. AI can help employees work faster, improve the quality of certain tasks, and give less-experienced workers access to capabilities that once took years to develop. But a deeper question sits underneath the productivity numbers: what happens to human capability when machines begin doing the work through which people develop expertise?
That question may matter more than how many jobs AI replaces. A company can automate a task without losing anything important, particularly when the task is repetitive or administrative. However, some work creates value in two ways: it produces an immediate result and, at the same time, teaches the person doing it how to become better. When that work is automated, the organization captures the first benefit while potentially losing the second.
Consider a young analyst asked whether her company should enter a new market. A few years ago, she might have spent several days researching competitors, speaking with customers, questioning the data, testing different hypotheses, building scenarios, and defending her recommendation to senior executives. She would probably have made mistakes along the way. That process was inefficient in the narrow sense, but it was also how she developed judgment. Today, an AI system can produce a sophisticated market analysis in minutes. The company saves time, and the output may even be better. What the productivity dashboard cannot show is whether the analyst became more capable or simply became better at asking a machine to do the thinking.
The productivity-capability paradox
There is strong evidence that AI can increase productivity. A study of 5,179 customer-support agents found that access to a generative-AI assistant increased productivity by 14% on average, with a 34% improvement among novice and lower-skilled workers. The researchers also found evidence that AI helped less-experienced employees adopt practices associated with more experienced colleagues. In this sense, AI can democratize expertise by allowing people to benefit from knowledge they have not yet accumulated themselves.
That finding raises an important question for executives: If AI allows a junior employee to perform like someone with five years of experience, how does that employee acquire the judgment normally developed during those five years? The answer is not necessarily that AI makes people less capable. It is that productivity and capability are different outcomes. A person can perform better with a powerful tool without developing an equivalent level of independent expertise.
A 2026 field experiment involving 128 knowledge workers found that generative AI consistently improved efficiency across several forms of knowledge work. Yet the effect on quality varied depending on the nature of the task. AI improved performance on some activities while reducing quality on certain knowledge-acquisition tasks-the kinds of activities through which workers develop understanding. Organizations do not build expertise simply by completing tasks. They build expertise by allowing people to struggle with difficult problems, encounter unexpected information, test assumptions, make mistakes, and gradually develop mental models that help them recognize patterns.
Automating the experience curve
Consider a junior consultant investigating why a client's sales have suddenly collapsed. She develops a hypothesis, tests it against the data, discovers that it doesn't fit, and starts again. She talks to customers, challenges the assumptions in the original analysis, and eventually discovers something that wasn't obvious at the beginning. Those hours may look expensive on a timesheet, but they are also teaching her how to investigate ambiguity and distinguish a plausible explanation from a correct one.
Now imagine an AI system generating five plausible explanations in thirty seconds. The consulting firm saves hours, which is exactly what the technology is supposed to do. But if the junior consultant increasingly relies on those explanations instead of developing her own investigative process, a different question emerges: Who becomes the experienced consultant five years from now? This is not an argument for forcing employees to do work that machines can do better. It is an argument for recognizing that some forms of work have a developmental function. If companies eliminate every difficult task from the early stages of an employee's career, they may inadvertently eliminate the experiences that create the senior employees they will need later.
AI can compress the experience curve, and that can be enormously valuable. However, companies need to understand the difference between compressing learning and eliminating learning. A 2026 randomized experiment involving 1,174 adults found that generative AI substantially narrowed performance differences between participants with different levels of education. When AI was available, the performance gap between higher- and lower-education groups fell dramatically. But when participants subsequently had to work without the technology, a significant portion of the underlying performance gap returned. The researchers also found that sustained effort mattered for retaining gains. Using AI successfully and developing independent capability are two different objectives, and organizations need to manage both. For leaders navigating this terrain, an AI Learning Path for CEOs can provide structured guidance on balancing these competing demands.
Judgment becomes the scarce resource
As AI makes information more abundant, information becomes less scarce. As AI makes analysis cheaper, analysis becomes less scarce. As AI makes writing, presentations, research summaries, and other forms of knowledge production easier, the ability to produce those things becomes less differentiating. What becomes more valuable is knowing what to believe. That requires domain knowledge, context, skepticism, and critical thinking. It means recognizing when the data doesn't make sense, identifying an assumption that everyone else has overlooked, understanding which variables actually matter, and knowing when an apparently sophisticated answer is fundamentally wrong.
A 2025 study of 319 knowledge workers, based on 936 real-world examples of AI use, found that "greater confidence in AI was associated with less critical-thinking effort." The researchers also found that AI changes the nature of critical thinking: instead of spending as much effort gathering information and producing an initial solution, workers increasingly have to verify AI-generated information, integrate its output, and oversee the resulting work. That shift can be beneficial but it also creates a new dependency. The better AI becomes at producing convincing answers, the more important it becomes for humans to have enough knowledge to challenge those answers.
Someone can be excellent at prompting an AI system and still have no idea whether the answer makes sense. The valuable combination is increasingly AI fluency, domain expertise, and critical thinking. Research from Harvard Business School and Boston Consulting Group provides another reason for caution. In a field experiment involving 758 knowledge workers performing realistic consulting tasks, researchers found that AI improved performance substantially on tasks within its effective capability frontier. But when participants encountered a complex managerial task outside that frontier, AI assistance actually reduced the likelihood of producing a correct solution. The researchers described this as a "jagged technological frontier": AI can be extremely capable at one task and surprisingly unreliable at another that appears similar to a human observer.
Measuring more than productivity
Most organizations are focused on the metrics that AI can improve: hours saved, tasks automated, costs reduced, revenue generated, and output per employee. Those measures matter, but they don't tell executives whether the organization is becoming more capable. CEOs should add a different set of questions to their AI strategy: Are employees becoming better decision-makers, or simply faster operators? Are junior employees still developing genuine expertise? Which experiences are being eliminated because they appear inefficient, and which of those experiences may actually be where future leaders learn?
These are difficult questions because the answers won't appear immediately in quarterly results. The cost of losing a developmental experience today may not become visible for five or ten years, when the organization discovers that it has fewer people capable of making difficult decisions independently. The strongest companies will not simply train employees to use AI. They will train them to interrogate it. They will teach people to verify rather than accept, to question rather than copy, and to understand the reasoning behind a recommendation rather than simply repeat it. They will use AI to remove meaningless work while protecting the experiences that build expertise. This may require CEOs to rethink what they mean by efficiency. Not every form of friction is waste. Sometimes friction is where learning happens.
Why this matters for executives and strategy
The differentiator will not be whether a company has an AI assistant. It will be whether its people have the judgment to know when that assistant is wrong. The biggest risk of the AI revolution is not that companies will suddenly become dumb. It is that companies could become faster at producing answers while gradually losing the ability to judge whether those answers deserve to be believed. That is not primarily a technology problem. It is a leadership problem. When execution becomes cheap, judgment becomes valuable. When answers become abundant, critical thinking becomes scarce. The CEOs who understand that distinction may build more than productive companies-they may build companies that can still think. For those ready to act on this distinction, exploring resources on AI for Executives & Strategy offers a starting point for developing the judgment frameworks their organizations will need.
Your membership also unlocks:
