The Problem Is Not the Tool. It Is the Habit.
When workers consistently accept AI-generated outputs without questioning them, two organizational risks accumulate over time.
The first is the loss of contextual expertise. AI can retrieve facts and generate plausible text, but it does not experience the data it processes. It cannot apply the kind of situated judgment that comes from years of working within a specific domain, organization, or set of relationships. A lawyer who relies on AI to interpret a case without deeply understanding the judicial system’s nuances is not just delegating a task—they are potentially misreading the situation.
The second risk is the erosion of epistemic pluralism. Large language models tend to average across their training data and fixate on early-generated ideas. Studies suggest this systematically reduces the diversity of outputs. When many people in an organization use the same tools in the same way, the result can be a kind of intellectual monoculture—one that is poorly equipped to detect weak signals, adapt to change, or generate genuinely novel ideas.
Both risks compound quietly. They do not announce themselves in a single failed project. They show up gradually, in the quality of decisions, the depth of analysis, and the organization’s capacity to respond when conditions shift.
Three Approaches Worth Examining
Research and organizational experimentation point toward concrete strategies. None of them require abandoning AI. All of them require treating AI as something humans can shape—not just a tool to be adopted as-is.
Turn AI Into a Question, Not an Answer
The default interaction with a chatbot is: ask a question, receive an answer, move on. A more productive framing inverts this. The goal is to use AI in ways that surface assumptions, challenge initial thinking, and force deeper engagement.
AstraZeneca’s “Prompt with Me” program, launched in 2025, offers a documented example. The 10-day initiative was designed not to teach employees to extract better answers from AI, but to use prompting as a mechanism for critical reflection. Participants reported that the approach challenged their assumptions and surfaced new angles they had not initially considered. Within a year, the program had been adopted across more than 50 teams and functions.
A separate experiment at a 24-hour hackathon involving close to 2,000 participants tested a related idea. Teams that received AI feedback designed to identify redundancies and push them toward alternative ideas produced a wider range of more innovative proposals than teams that received straightforward improvement suggestions. The framing of AI feedback—not just its content—shaped the quality of human thinking.
Define Where AI Should Not Be Used
Most AI governance frameworks focus on where AI can be deployed. A more complete framework also defines where it should not be.
This is not a novel concept. Organizations have experimented with email-free time blocks to combat information overload. The same logic applies to generative AI. Structured AI-free periods—whether for strategy sessions, early-stage problem framing, or entry-level skill development—create space for the kind of unassisted thinking that builds genuine expertise.
One Australian telecommunications carrier studied by researchers required mid-level managers to complete AI-free strategy sessions before engaging AI tools. The sequence mattered: human judgment first, AI assistance second.
The point about entry-level roles deserves particular attention. Research on investment banking found that junior analysts who worked through processes manually—before relying on algorithmic tools—were better at spotting errors and questioning assumptions. As organizations reduce entry-level positions to cut costs, they risk eliminating the developmental pipeline that produces the contextual expertise their senior teams currently rely on.
Process in Parallel
Rather than choosing between human judgment and AI output, some organizations are experimenting with running both simultaneously and comparing the results.
A design consultancy based in Lisbon tested this approach on a branding project. One team used a traditional design process; another relied heavily on AI-supported workflows. The AI-supported team generated close to 1,700 images, expanding the exploration space rapidly. The human-led team produced fewer options but with stronger narrative coherence and contextual framing. The human-led proposal was ultimately selected—but the more durable insight was structural: AI and human teams demonstrated different strengths at different stages of the creative process.
Parallel processing is not efficient in the short term. It is a method for understanding where AI adds genuine value and where human judgment remains irreplaceable.
Reconsider the Interface
The chatbot format is not a neutral design choice. It shapes how people interact with AI, what they ask, and how they process responses. A wall of bulleted text invites passive consumption, not active deliberation.
Research in human-computer interaction has long challenged the assumption that a single interface type suits all tasks. Some recent proposals advocate for interfaces that present raw data and competing evidence rather than synthesized conclusions—designs that require users to make sense of alternatives rather than simply accept a recommendation.
Google’s “AI Co-mathematician” illustrates what a task-specific interface can look like. Designed for mathematical research, it includes an annotation system that makes the reasoning process visible and supports iterative verification. It is built around the actual workflow of its users, not around the generic chatbot model.
The implication for organizations is practical: the interface through which employees interact with AI is a design decision, and it has consequences for how much critical thinking the interaction demands.
What Leaders Should Actually Do
The research does not argue against AI adoption. It argues against passive adoption—the kind where efficiency metrics are tracked and cognitive costs are not.
A few concrete implications follow:
- Measure what you might be losing, not only what you are gaining. Expertise erosion and idea homogenization are harder to quantify than time saved, but they are not invisible.
- Design AI-enabled workflows deliberately, with explicit decisions about where human judgment must precede AI input.
- Invest in entry-level development even when AI can perform those tasks. The contextual knowledge built at junior levels does not disappear from the organization—it migrates upward over time.
- Experiment with interface design rather than defaulting to the chatbot format for every use case.
The organizations that will use AI most effectively over the long term are likely not those that adopt it most aggressively, but those that adopt it most thoughtfully—preserving the human capabilities that AI cannot replicate and that remain the actual source of organizational resilience.
Comments (0) No comments yet
Want to join this discussion? Login or Register.
No comments yet. Be the first to share your thoughts!