Design AI Systems That Actually Strengthen Human Reasoning

Most organizations recognize the importance of having employees who do more than take instructions or apply rules. Critical thinkers ask questions and challenge assumptions, behaviors known to be crucial for organizational agility and innovation. But there’s growing evidence that use of AI can erode this capacity. To help companies address this problem, we build upon existing research in management (including some of our own), cognitive science and human-computer interactions, and the studies of work and technology in organization, along with interviews of executives at companies experimenting with different approaches, to offer a new approach on the design, implementation, and application of AI that that will encourage a diversity of views and critical thinking in the long term. Organizations should focus on two main organizational risks related to the erosion of critical thinking: the loss of expertise and the loss of epistemic pluralism. These risks can potentially harm their ability to stay agile and innovative. Leaders need to recognize that technology is not neutral, and that it might shape our ways of thinking and interacting in negative ways. They should therefore look beyond short-term gains (the promise of efficiency and reduced costs, for example) and instead encourage users to not take technology’s features for granted and experiment in order to shape AI rather than be shaped by it. How AI Shapes Thinking Even before the explosion of generative AI, a number of studies, including Callen Anthony’s 2021 ethnographical study of investment bankers, had already suggested that the phenomenon of “black boxing”, in which workers trusted analytic tools without understanding them, could cause professionals to lose their expertise. More recently, studies have suggested that using AI can compromise people’s ability to question the information they receive, raising concerns about cognitive offloading, or even cognitive surrender. A 2025 MIT Media Lab study investigating the cognitive costs of using an LLM and a 2026 Wharton study on the impact of introducing AI as a third cognitive system both indicate that people using AI are disinclined to question and research AI output. While the long-term impact on cognition remains to be determined, this research does highlight clear signs that AI usage does affect organizational capabilities. To begin with, many researchers worry that AI usage will lead to a decontextualization of expertise in organizations. The argument is that while AI can readily access facts and data, it cannot contextually interpret and apply knowledge meaningfully. Data often contains nuances that can only be recognized by people with intuition, emotion, and lived experience of multiple domains of expertise. We see things in the data because we have direct experience of it and the context in which we understand the data. Because AI is an algorithm and does not experience the data directly, it overlooks the nuances. For example, the fact that an AI can pass a Bar examination does not indicated it understands all nuances of the justice system, its stakeholders, how different cases register differently and why, and how other forms of expertise relate to the law. This decontextualization of legal expertise introduces risk because without an understanding both of the specific case and of a nuanced judicial system, lawyers relying on AI expertise may misinterpret evidence and make mistakes. In fact, one recent study suggests that reliance on AI arguably increases the importance of having people who, based on domain expertise, can validate AI outputs, adapt to new situations, and plan for potential issues. Another concern is that as organizations cut entry-level positions to reduce costs, they risk losing in the near future the pool of middle managers on whose contextual understanding they currently rely on. In addition, AI encourages homogeneous thinking. Several studies have shown that LLMs tend to reduce the diversity of ideas, partly because they tend to fixate on the first idea generated and have a bias towards averaging. This encourages users to ignore surprising or unforeseen information even though it is the unexpected that triggers creative insights. As such, gen AI use leads to the development of monocultures that lack the diverse experiences and perspectives that are essential for organizations’ capacity to innovate. It may also affect companies’ agility as people are more likely to miss signals in the environment inviting change. So how can we control the risks? There’s a rich and well-developed literature on work and technology in organizations that provides potential answers. Specifically, it frames technology as always interpreted and as something contextual that can be adapted by users. How Humans Can Shape AI Current narratives around generative AI tend to present it as a single tool with a chatbot interface that people interrogate to get an answer (see, for example, here and here). They stress that, in order to get better answers, we need to learn how to prompt better and use it more often. There’s nothing wrong with learning how to ask better questions, but it won’t materially reduce the risks. To do that, we need to treat technology not as just a tool, but rather as part of a system that humans can shape. And that means findings ways to design AI-enabled work environments that preserve human agency and leadership. Drawing on the results of experimentation with AI and other technologies by organizations we have studied or worked with, we present three possible ways to do this. Turn AI on its head. The Greek philosopher Socrates used to ask his followers questions in order to lead them to find on their own the answer to a philosophical problem; or in some cases to admit their ignorance. This approach can be used with AI. Consider two examples: The Prompt with Me Challenge: Launched by the large pharmaceutical company Astra Zeneca in August 2025, this 10-day program composed of brief exercises was designed to encourage employees to experiment with AI in order to develop their ability to think critically. , More than 1,500 participants across all the companylearned not only to prompt but, more importantly, how to prompt in ways that that made them learn more and dig deeper. Feedback from participants illustrated how the challenge helped reframe the relationship to AI: “this approach surfaced new angles, challenged my assumptions, and inspired self-reflection for any upcoming conversation” noted one of the participants. In an interview with us, AstraZeneca’s Global Head of Capability Development, Dr Bonnie Cheuk, and Dr Maciej Szymaszek, Head of Strategy & Innovation, Enterprise AI explained that “the challenge … rapidly scaled across the organization, with teams and departments taking part and further customizing it for different workflows, functions, and AI tools.” The original 2025 pilot won a silver Brandon Hall award. Today, a year after it was launched, Prompt with Me has been adopted by more than 50 teams and functions, with thousands of people completing the challenges. The BUILD2GETHER AI–SDG Hackathon: A few months ago a team of researchers from ETH Zürich and NTU Singapore experimented with the use of AI in a 24-hour hackathon (hosted at Sairam Institutions) with close to 2,000 participants across nearly 500 teams. Teams received different types of feedback from AI on proposals submitted for projects that would help the UN achieve its Sustainable Development Goals. Half of them received feedback on how to improve their ideas to better reach and fit the predefined goals. The other half received feedback that identified redundancies between their proposals and projects that were already under way, encouraging them to experiment with alternative ideas. The research team found that when AI challenged participants from the second group to question proposals and help distinguish which elements were novel versus already existing, these teams came up with a wider range of more innovative ideas. Understand the right place and time for using AI. In setting an AI governance framework, companies focus on optimizing where they can implement AI. Instead, they might want to consider where and when AI should not be used. We suggest that leaders: Create AI-free space to think freely: This strategy is inspired by past initiatives to address email overloads. For instance, in 2007, Intel did a pilot with a group of engineers to combat the information overload their employees faced where for one day a week the engineers were encouraged to avoid using email. (They followed other companies like Deloitte or U.S. Cellular, which had also experimented with a no-email day/week.) Similarly, organizations and managers can encourage people to not use gen AI tools during specific time blocks or providing them with “gated interfaces that unlock assistance only after the user deposits their own context,” as suggested by Chengwei Liu and colleagues in a recent HBR article. And another group of researchers reported how a Australian Telecommunication carrier they studied required mid-level managers to engage in “AI-free strategy sessions” before using AI tools, forcing them to first brainstorm strategic plans based on their own judgement and experience. Center human agency: Technology’s ubiquity in current development and growth strategies can sometimes have so much sway over an organization’s decisions and structure that its use becomes unquestioned. Making sure that humans can shape technology starts with giving employees, particularly entry-level ones, a chance to develop expertise and contextual knowledge, by requiring them to work on specific tasks without any AI tools. Think of the junior bankers studied by Callen Anthony. She found that those who learned to understand the process through practicing on their own were better able to do the work. They were more likely to spot mistakes and suggest alternatives. Understanding the process is not crucial only for junior roles. It matters for mid- and senior managers as well. Anthony also found variations among senior bankers: one group considered algorithmic technology as only a tool to produce outputs and thus weakened their expertise while another group of senior bankers made sense of the different technologies used and how they shaped their work. The latter were able to deepen their expertise and meaningfully engage with the junior analysts: asking them to justify their choice of tools and to explain the assumptions and calculations beyond the numbers they presented. At a broader organizational level, studies of technology and work have shown how the effective implementation of new technologies involves redesigning work processes that build on workers’ knowledge. In fact, we see this approach at play in an experts’ report sponsored by the Spanish Ministry of Labor, showcasing the importance of involving workers of enterprises in conversations regarding the implementation of AI. Process in parallel: In some situations, it may be useful to get AI and a team of humans to perform the same task independently in parallel. The team and AI can then follow up to produce a hybrid outcome. For example, With Company, a transformative design consultancy based in Lisbon, has been experimenting with parallel processing in strategic and creative projects. In one early experiment in 2024 with an early-stage venture, two teams worked on the same branding challenge in parallel: one using a more traditional design process, and another relying heavily on AI-supported workflows. The AI-supported team generated close to 1,700 images before arriving at a final proposal, dramatically expanding the exploration space and accelerating divergence. The human-led team produced fewer routes but with stronger coherence, sharper narratives, and greater contextual framing from the beginning. Ultimately, the human-led team’s proposal was selected, but the more general lesson for With Company was that parallel processing revealed differences in the relative strengths of AI and human teams at different stages of the creative process. AI proved useful for transcription, synthesis, rapid exploration and expanding possibilities, while framing, prioritization and judgement still depended heavily on human interpretation and contextual understanding. Since then, With Company has continued experimenting with parallel processing on multiple projects. Try different interfaces. Many gen AI and LLM models are formatted around the chatbot design, which affords a “question and response” interaction in most contexts. Its response is usually a large wall of text that requires the user to read through bulleted points and sift out the answer they’re looking for. However, this form of interface is not well-equipped for all situations. What often happens is that the chatbot interface devolves into a very messy “conversation,” where, for example, a large, summarized text generated by AI is simplified, re-structured, and re-interpreted by an office worker who is overwhelmed by the massive amount of information presented. Gen AI then reinforces that confusion by responding to the reinterpretation in a way that structurally mimics the worker’s response. Research in human-computer interaction has challenged the idea that the only interface to a computer is a keyboard and mouse. Similarly, we can challenge the fact that AI is just a chatbot. One recent paper has proposed moving away from the “frictionless user interface” and replacing it with one that shares raw data and provides competing evidence, forcing the user to make sense of and deliberate on the proposed alternatives. Empirical evidence suggests that intricate interfaces like this work well in a context such as mathematics or coding, where users are already experts. Google, for example, noted that current chat interfaces did not support the complex multi-dimensional and iterative process that mathematical research entails. To fill this gap, it developed the “AI Co-mathematician,” a tool designed with a flexible interface that both helps mathematicians to develop their workstreams and assists in verifying proofs and checking for errors. It can better support the iterative mathematical process by providing an annotation system that visualizes the different steps of the reasoning and makes the reasoning process more visible. Tools like this illustrate that AI need no longer be the generic chatbot that we currently see and use it as. . . . As AI continues to be implemented in organizations, managers must go beyond the hype and adopt a holistic, balanced view of how to use AI in organizations. To support their employees in being active knowledge producers, it is essential for organizations to engage in strategic experimentation with AI in order to identify how to use the technology to support work, preserve expertise and nurture diverse thinking. AI should empower people rather than stand in for them.