The faster AI evolves, the more organizations need to learn to create "friction"
Leaders are passionate about AI because it makes knowledge reuse extremely easy: generating drafts instantly, writing code on the fly, and completing analyses in real time. But research shows that making knowledge easy to reuse hides a hidden cost.
We study how organizations learn. In our recently published research in Management Science, we adopted a formal analytical model (a mathematical framework that analyzes how individuals in organizations choose between "independent exploration" and "reusing others' achievements") to examine what happens when "good enough" answers can be obtained almost at zero cost. Model predictions show that as the degree of knowledge reuse increases, independent exploration behaviors decrease, and team members gradually converge on a small number of identical methods. Simply put, productivity rises, but innovation stagnates quietly without being noticed.
We have observed similar patterns in empirical research. For example, in a 2015 paper published by Kevin Boudreau and Karim Lakhani in Research Policy, they found that when computational biologists had unobstructed access to each other's intermediate solutions, people devoted more energy to improving existing schemes rather than conducting independent experiments. As a result, the exploration space became narrower and narrower, which is highly consistent with the "productivity trap" we predicted.
We have also experienced this scenario in our daily work. To keep up with the development curve of AI, our institution set up weekly regular meetings to share the latest tools and application cases. The attendance rate was very high at the beginning, but as time went by, the team put forward fewer and fewer new discoveries, fewer and fewer people participated in discussions, and the meeting content gradually turned into introductions to the latest AI tools. The reason is that without these popular science sessions, other people would not have enough background knowledge to understand those novel applications. The result is that a small number of people continue to explore, while the majority wait to benefit from these exploration results. This mechanism does promote knowledge sharing, but it weakens people's enthusiasm for investing in high-cost exploration work, which is the source of new insights.
It follows that generative AI has changed the economic logic of learning. If a seemingly reasonable strategic memo can be generated in a few minutes, fewer and fewer people will be willing to spend several days doing relatively difficult work that can produce original insights, such as communicating with customers, cross-validating data, and repeatedly testing hypotheses. When "good enough" answers become free, fewer people will be willing to pay the cost to find "better answers". Eventually, enterprises may be full of people who can quickly call various AI outputs, but lack people who can judge whether these results are reliable, adjust them according to actual conditions, or even overturn their conclusions.
Avoid the Productivity Trap
Solving this problem requires a fundamental shift in the way we view effort. In our model, adding some friction to knowledge sharing at the individual level can instead build what scholars call "absorptive capacity" and increase the total amount of knowledge at the collective level. Absorptive capacity refers to the ability to evaluate, adjust and improve ideas, rather than just copy them. When people have to make certain efforts to use the achievements of others, they will carry out more independent inspections and adjustments. This input not only helps them learn from others, but also generates new knowledge for others to draw on. As a result, the long-term stock of knowledge in the entire system is higher, bringing higher long-term average performance.
This finding prompted us to pilot a set of "friction protocols" in MBA and doctoral seminar classes. We set a simple rule: to participate in the sharing session, one must first prove that they have tried independently. We welcome success stories, but we value those "failed prompts" more, that is, attempts that hit a dead end, and situations where AI hallucinations occur. It is easy to copy others' success from the Internet, but only through real exploration can one record a course of failure.
Although the introduction of "strategic friction" seems to slow down the team at the moment, it actually makes everyone better. More participants come to the scene with "partial solutions" to the same problem, and the team can then combine these partial solutions into a more complete practice. By forcing participants to "struggle" with the tool first, they build absorptive capacity. In this way, when a breakthrough achievement is shared, participants can not only understand why it works, but also know how to adjust and apply it. What we share is not just prompts, but also expands the total knowledge of the entire team.
Cultivate Absorptive Capacity
Our research is based on a formal model rather than a field experiment: the model proves that learning from others requires a certain level of prior understanding, so that long-term knowledge production will be improved, which means that more different practices will be retained, fewer people will completely abandon exploration, and organizations are less likely to fall into a small number of "good enough" solutions. The core mechanism revealed by our research — absorptive capacity — is built on decades of research, pointing out that existing knowledge can provide a conceptual framework to help individuals encode and apply new information. Just as children learn language by building new vocabulary on top of words they have already mastered, the process by which professionals accumulate expertise is the same.
In practice, absorptive capacity is reflected in the difference between "getting an answer" and "being able to use that answer to get things done".
A manager with absorptive capacity will not just accept the competitor analysis report generated by AI, but will be able to find the gaps in it, verify its analysis with observed market conditions, and adjust suggestions according to their own scenario. Managers who lack absorptive capacity will forward the AI output as it is without any modification.
When we piloted the "friction protocol" in seminar courses, we tracked two indicators before and after implementation. The first is the diversity of problems and application scenarios that emerge in the sharing session — which is a proxy indicator for measuring independent exploration; the second is the depth of adjustment of the shared content (did the participants return a copied prompt, or a set of workflows adjusted for specific purposes?). After requiring evidence of independent attempts, the changes were very significant. Before implementing the "friction protocol", students rarely raised original questions, and almost no one shared immature partial solutions. After introducing this rule, each student contributed an average of two different questions and about one partial solution for each question. Since everyone made independent attempts in advance, the discussion process became more active, and the shared content reflected real re-creation: participants brought modified workflows instead of simply copied prompts. Everyone was able to quickly build on each other's phased results, and often combined complete solution paths during the discussion. This phenomenon is exactly the result predicted by our model: friction does not reduce learning; on the contrary, it increases both the quantity and quality of shared knowledge.
For managers who want to track the absorptive capacity of their teams, there are several observable behaviors that can be used as verification indicators: Can employees detect errors, omissions or unreasonable assumptions in AI outputs and explain the reasons? Is the final deliverable substantially different from the original AI output, with new facts, constraints and counterexamples added? Do employees contribute original observations, such as customer feedback, internal system data or competitor dynamics, rather than just generic prompts? And can they transfer the experience gained in one scenario to new problems instead of starting from scratch every time? When these behaviors continue to increase, it indicates that the team's absorptive capacity is strengthening.
Unfortunately, the way many enterprises deploy AI today is precisely eliminating the necessity for people to understand things. Leaders often focus on increasing adoption rates and ease of use, but inadvertently weaken the organization's absorptive capacity. Therefore, if they want to protect the long-term innovation momentum of the organization, leaders need to rethink the logic of simply pursuing "the easier to use, the better", and start to consciously redesign "strategic friction" back into the organizational system.
Introduce Strategic Friction
Our goal is not to make work painful, nor to add bureaucratic obstacles, but to ensure that knowledge reuse is built on capabilities. The key is to let people invest enough effort in the process so that they can keep learning. Our research shows that there is an optimal level of absorptive capacity in organizations. Too little friction will weaken exploration; too much friction will slow everything down. Moreover, not every task is worth adding friction. For routine work that requires less judgment, such as organizing meeting minutes, speed is the key. Strategic friction should be reserved for tasks where organizations really need employees to think and learn, rather than for work whose goal is simply output.
In the enterprise context, this means shifting from an "oracle model", where users ask questions and the system gives answers, to a "reciprocal model", where users first contribute their understanding to extract wisdom. There are many ways to achieve this transition, ranging from simple strategic adjustments to more ambitious system designs. Leaders can start with the following aspects.
Set rules to require records of independent attempts. For example, before using AI to write market analysis, analysts must first submit their own preliminary judgments, which do not need to be complete, but must explain what information they have mastered and what problems they are stuck on. This kind of friction is the easiest to implement because it does not require modifying the system. This is exactly the approach we adopted in academic seminars: if you cannot prove that you have made a real attempt, you cannot participate in the sharing session.
Design AI to require user input first before generating complete output. For example, when an analyst requests to generate a competitor analysis, AI will not immediately give a report, but first ask the other party to provide three specific observations about the competitor's recent pricing strategy. Then AI will integrate this new information into the analysis results. This approach forces analysts to conduct investigations before obtaining answers, which not only cultivates absorptive capacity, but also collects front-line intelligence, preventing AI from continuously reusing outdated information. Of course, users still input information through prompts, so this kind of friction is relatively mild.
Establish a threshold interface that "unlocks help only after information is entered first". For example, before an analyst asks the internal AI to generate a risk assessment report, the system interface will require him to upload a baseline assumption and mark three key variables or constraints. Only after these manual inputs are completed will the "Generate" button be unlocked. AI will continue to deepen the analysis based on the information provided by the analyst, instead of directly replacing the analyst. This ensures that every output is a collaboration between human and AI, rather than a simple task delegation. This is the most demanding kind of friction, which requires enterprises to develop customized AI tools.
Matters at the strategic level can be started immediately by team leaders. Solutions at the interface level require certain design investment, but can bring double benefits: they not only improve the absorptive capacity of users, but also continuously collect front-line business information, preventing the AI model from falling into a closed loop that only relies on its own outputs. Practices at the system level are long-term investments, but this is exactly the direction enterprises should move towards — designing AI tools that make people smarter, not just make work faster.
How to Identify Builders
Efforts to prevent organizations from stagnating can even start before employees log into the system. Many enterprises still recruit talents for the purpose of pursuing "correct answers". But in an AI-saturated environment, this ability is difficult to form a competitive advantage. What really matters is the "correct process". In interviews and performance evaluations, managers should observe how candidates and employees use AI tools: Do they treat AI outputs directly as finished products, or do they ask follow-up questions, verify assumptions, and add their own observations?
Recent research has made these patterns visible. A field study conducted by Hila Lifshitz and her colleagues from Harvard University, the Massachusetts Institute of Technology and the University of Warwick analyzed nearly 5,000 human-AI interactions from 244 Boston Consulting Group consultants. The research focused on three different human-AI collaboration styles, distinguished mainly by two questions: who decides what to do, and who decides how to do it?
"Centaurs" retain human dominance in both dimensions, and AI only provides targeted assistance. "Cyborgs" let humans decide "what to do", but give AI more room for participation in "how to do it", and work collaboratively with AI through continuous critical dialogue. "Self-automators" leave both to AI, and fall into the state of "no-skilling" as called by researchers — neither cultivating domain expertise nor improving AI application capabilities. In our framework, "Centaurs" and "Cyborgs" are builders; while "Self-automators" are the "free riders" warned by our model.
These findings point out that the difference between builders and free riders is mainly reflected in the following two behaviors.
Selective Delegation. Builders understand the boundaries of what AI is good at and what it is not. They delegate tasks such as drafting documents, brainstorming and pattern comparison to AI, and reserve work that requires high-level judgment for themselves. From the prompt records, we can see that their questions are specific and targeted, rather than outsourcing all the work.
Interrogation Depth. Builders challenge AI when using it. They ask AI to explain the reasoning process, point out contradictions, and feed their own data to AI. They also use AI in reverse: first give their own analysis, and then ask AI to test the assumptions in it. Free riders, on the other hand, tend to accept everything as it is. The direction of information flow speaks for itself: builders continuously input background information and contextual knowledge to AI, while free riders only extract ready-made answers from AI.
Evaluating employees' AI usage not only helps to cultivate stronger absorptive capacity and healthier human-AI collaboration habits, but also helps organizations identify and invest in people who truly create knowledge, rather than just consume knowledge.
AI can immediately improve the operating speed of an organization. But the challenge for leaders is how to prevent this speed from eroding the organization's absorptive capacity. By introducing carefully calibrated "strategic friction" in tools, workflows and talent recruitment, organizations can be helped to not only continuously output answers, but also continuously generate insights.
Chengwei Liu, Jerker Denrell, Jerry Luukkonen, Nick Chater | Article
Chengwei Liu is Associate Professor of Strategy and Behavioral Science at Imperial College Business School, and co-author with Balázs Kovács of The Smart Contrarian: Thinking Differently About Thinking Differently (Harvard Business Review Press, 2026). Jerker Denrell is Professor of Behavioral Science at Warwick Business School. Jerry Luukkonen is Lead Consultant at Dectech, a behavioral science consulting firm. Nick Chater is Professor of Behavioral Science at Warwick Business School, Fellow of the British Academy, and co-founder of Dectech.
Qiqing Liao | Editor
This article is from the WeChat official account "HBR Chinese" (ID: hbrchinese), the author is HBR-China, and 36Kr is authorized to release it.