Strategic Management Explorer

What Perfumers Can Teach Us About Working With Generative AI

By Tomoko Yokoi, Daniella Laureiro-Martinez, Federico Magni, and Stefano Brusoni

Most discussions of generative AI at work start with task allocation: What should the AI do, and what should remain human? That question matters—but new research suggests it misses an equally important challenge.

Humans and generative AI do not simply have different capabilities. They represent knowledge differently. People draw on tacit, embodied, contextual experience. GenAI operates through codified information and statistical representations. Even if each side performs its assigned task well, their outputs still have to be translated and integrated into one coherent solution.

That is the core insight of “Organizing across cognitive asymmetry in human–AI collaboration: A study of perfume creation,” a 2026 Strategic Management Journal article by Tomoko Yokoi, Daniella Laureiro-Martinez, Federico Magni, and Stefano Brusoni.

The Research

The researchers conducted a qualitative study inside a multinational fragrance house that introduced GenAI into professional perfume creation. The setting is unusually revealing because perfumery depends heavily on smell, memory, aesthetic judgment, and embodied expertise—knowledge that is difficult to put into words or data.

The authors describe the difference between human tacit knowledge and AI’s codified knowledge as “cognitive asymmetry.” That asymmetry creates a representational gap: AI can generate useful possibilities without possessing the sensory and contextual understanding that gives those possibilities meaning to an expert perfumer.

The study identifies an organizing process called “representational integration” through which professionals bridge that gap.

Practice 1: Allocate Tasks According to Cognitive Advantage

The first practice is not simply “give repetitive work to AI.” The observed division of labor reflected the different ways humans and GenAI contribute to creative problem-solving.

Humans framed the problem and evaluated potential solutions. GenAI generated alternatives. This division allowed the system to use AI’s generative capacity without pretending that the model possessed the same experiential judgment as the human expert.

In perfume creation, a client brief can be abstract, metaphorical, and sensory. Human experts interpret what the brief means in context. GenAI can then expand the solution space by producing alternatives, but humans remain central to deciding whether those alternatives make sense.

The broader management lesson is that task allocation should reflect cognitive strengths, not just relative speed or cost.

Practice 2: Translate Tacit Knowledge Into a Form AI Can Use

The second challenge is knowledge conversion.

Human expertise often exists in forms that are difficult to articulate: a sensory impression, an intuitive judgment, or experience accumulated across years of practice. AI cannot directly access that embodied knowledge. People therefore need to convert part of it into codified representations that the model can process.

The authors call this movement from embodied expertise into explicit representation “disembodiment.” The output then needs to move in the other direction. Humans interpret the model’s codified suggestions through their own sensory and contextual knowledge—a process the paper describes as “re-embodiment.”

This translation work is central to collaboration. Without it, the human and AI may each produce something useful but fail to create a shared basis for coordination.

Practice 3: Steer the AI as the Human’s Understanding Changes

Creative work is rarely a single prompt followed by a final answer. The study shows that GenAI outputs can change the human expert’s own understanding of the problem.

As new alternatives appear, professionals revise their mental representation of what they are trying to create. They then steer the AI again, generating another cycle of alternatives and interpretation.

The result is an iterative process rather than a static handoff. Human and AI contributions continually reshape the problem being solved.

This is especially important for knowledge-intensive work where the goal itself becomes clearer through exploration.

Why Integration Matters More Than Simple Automation

Much of the human–AI literature focuses on determining which agent is best at which task. Yokoi and colleagues shift attention to what happens after those specialized tasks are performed.

Specialization creates an integration problem. The more different the human and AI contributions are, the more deliberate coordination may be required to turn them into a coherent outcome.

The authors’ central point is not that human tacit knowledge makes AI unhelpful. It is the opposite: GenAI can be highly generative, but organizations capture more value when employees can connect its outputs to experience, context, and embodied judgment.

What Managers Should Do Differently

  • Design workflows for integration, not just task substitution: A workflow that says “AI generates, human reviews” may be too simplistic for complex knowledge work. Managers should identify where representations need to be translated between human experience and AI-readable forms.
  • Build skills for expressing tacit expertise: Prompting is only part of the capability. Employees also need to learn how to make assumptions, sensory judgments, contextual constraints, and problem frames explicit enough for AI to work with them.
  • Preserve human evaluation and sensemaking: The paper emphasizes that human contextual understanding remains indispensable. In the observed process, humans framed problems and evaluated solutions rather than outsourcing those roles to the model.
  • Expect iteration: When AI outputs change how people understand a problem, effective collaboration requires repeated steering rather than a one-time request. Organizations should design time and processes for that back-and-forth.

The Takeaway

The value of generative AI in expert work does not come simply from finding tasks a model can perform. It comes from organizing the relationship between two very different forms of cognition.

Yokoi, Laureiro-Martinez, Magni, and Brusoni’s study suggests that successful human–AI collaboration requires three things: allocate work according to cognitive advantage, translate between tacit and codified knowledge, and continually steer the AI as human understanding evolves.

The strategic question is therefore not only, “What can AI do for our experts?” It is, “How will we organize the translation and integration work that turns AI output and human expertise into something neither could produce in isolation?”

Topics

Technology
Innovation
Organizations
Human Resources
Strategy Process
Strategy Practice
Behavioral Strategy

Big Six Questions in Strategic Management

About the Authors

Tomoko Yokoi is a researcher at the Chair of Technology and Innovation Management at ETH Zurich and at IMD Business School’s Global Center for Digital Business Transformation. Daniella Laureiro-Martinez is a professor at the Chair of Technology and Innovation Management in ETH Zurich’s Department of Management, Technology and Economics. Federico Magni is an Assistant Professor in the Division of Leadership, Management and Organisation at Nanyang Business School, Nanyang Technological University. Stefano Brusoni is Professor of Technology and Innovation Management in ETH Zurich’s Department of Management, Technology and Economics.

 

Published Date
17 November 2026

Reference

Yokoi, T., Laureiro‐Martinez, D., Magni, F., & Brusoni, S. (2026). Organizing across cognitive asymmetry in human–AI collaboration: A study of perfume creation. Strategic Management Journal.

Contributed By
Tomoko Yokoi, Daniella Laureiro-Martinez, Federico Magni, and Stefano Brusoni

Article Type
Article Summary/Abstract

NEWSLETTER

Sign up to receive updates on the latest research, events, and SMS news.

Related posts

RECENT POSTS