4 lessons from behavioral economics that could help improve human-AI interaction

1. Fighting brain fry with transparent processes

Because AI tools can make it easier to get more work done, employees could find themselves being asked to do too much and could experience AI-related burnout. Moreover, when evaluating the results of AI-enabled or AI-generated work, decision-makers may need to consider what to revise, trust, or discard before moving forward. The accumulation of those extra efforts might produce a form of cognitive exhaustion that Harvard researchers call “brain fry,”² which occurs when workers are attempting to work beyond their own cognitive limits on an ongoing basis. Over time, brain fry can contribute to degraded motivation, weaker collaboration between colleagues, and higher rates of stress and burnout.³

Behavioral research offers a possible design response that could help limit one underlying cause of brain fry for AI users: the perceived need to review multiple tools, outputs, and alternatives before feeling confident enough to move forward. Studies find that people have more confidence in recommendations they perceive as being the result of purposeful, productive efforts, a phenomenon Harvard researchers have dubbed the “labor illusion.⁴”

Therefore, to keep employees from feeling the need to keep checking, comparing, and second-guessing AI outputs, organizations could increase transparency in AI tools so that employees don’t feel obligated to independently reconstruct how an output was produced. “Deep think” labels or badges confirming that AI tools and outputs are based on leading data sources, confidence cues that the information provided was double-checked for source accuracy, and structured reviewer notes that include sourcing and underlying assumptions could give employees clearer signals about when to trust or further scrutinize an AI output.

If AI systems make their reasoning process more visible—showing sources, flagging assumptions, or signaling when deeper analysis has been performed—humans can spend less effort trying to determine how much scrutiny is required and more effort applying judgment where it matters.

3. Protecting experience with beneficial friction

Deloitte research has identified human capabilities—which include traits such as curiosity, divergent thinking, and resilience—as increasingly important to successful team performance in an AI-powered future,⁹ yet those traits might be atrophying under heavy AI reliance. And these capabilities around judgment and creative thinking could become increasingly important as AI takes on more routine work and leaves only trickier exceptions for humans to manage.¹⁰

One potential design response is to build “beneficial friction” into AI workflows through deliberate pauses that keep humans actively engaged. For example, employees might be asked to articulate their reasoning before receiving an AI output, demonstrate understanding before acting on it, or complete quality assurance checks and peer reviews for key AI-generated work.¹¹ These “speed bumps” can turn AI use into a moment of practice by requiring workers to reason, explain, question, and validate rather than simply accept an answer. In this way, friction can help strengthen the skills AI could otherwise erode while still capturing AI efficiencies.

4. Preventing AI groupthink with independent judgment

A subtler risk can emerge when AI becomes the primary input to strategic decision-making. Convergence, or what researchers have dubbed the “agentic convergence trap,” is the tendency for AI-assisted strategy to drift toward similarity as a result of shared AI tools, shared training data, and shared output patterns.¹² The risk is compounded by sycophancy, a feature of many AI models that reflexively validate rather than challenge the assumptions of the user, inadvertently discouraging scrutiny. The result can be undifferentiated strategies, diminished internal debate, and overconfident decision-making—familiar organizational issues that could be perpetuated or even exacerbated by AI.

One structural response that could combat these issues is requiring employees to articulate their own estimate or recommendation before querying an AI tool. This approach draws on the logic of the Delphi method, a structured technique for using diverse expert perspectives to inform strategic thinking. In this method, individual views are gathered independently before collective discussion begins, specifically to prevent the distortion of peer pressure and anchoring.¹³Applied to AI-driven business, the goal is to help ensure that AI supports human judgment at the points it can provide real value, introducing strategic diversity among organizations and teams.

Designing AI interactions that keep humans thinking

As leaders focus more on designing human-AI interactions that can help enable organizations’ success, measuring time saved, tasks completed, and volume generated are likely to remain important business metrics, but it might also be necessary to track AI’s potential impacts on human cognition and to help prevent possible declines in reasoning quality. For instance, project retrospectives could examine whether AI-assisted teams considered alternatives, surfaced risks, and applied independent judgment. The organizations most likely to sustain advantage from AI adoption may be those that think as carefully about what using AI does to their people as they do about what using AI does for their business.