What is cognitive economics?

In Walliser's account, economic analysis examines the information actors hold, the beliefs they form, and the decisions that follow. People learn and revise their reasoning through interaction. The interpretation of a situation matters alongside the information available. This gives cognitive processes an explicit place in the analysis of economic activity.

The expression cognitive economy is sometimes used more broadly for an economy in which knowledge contributes heavily to value creation. That is a characterization of production rather than the same research program as cognitive economics. Bringing the two ideas together is useful when studying work with AI, provided their meanings remain distinct. A change in how knowledge is produced raises questions about how people interpret and use it.

The knowledge economy provides an earlier reference point. In 1996, the OECD described economies increasingly based on the production, distribution, and use of knowledge, with learning central to their operation. Generative AI belongs within that history. It adds systems that contribute directly to outputs previously prepared by professionals. The relevant changes occur in particular tasks and workflows before they can be interpreted at economy-wide scale.

What does AI change before a decision is made?

An AI system can draft wording, combine information, or generate options before a professional selects a course of action. Depending on the task, assistance changes how operations are divided between the person and the system. The effect needs examination at that level. A useful result in one activity provides a reason to investigate another application, while the evaluation still needs to fit its own task.

Brynjolfsson, Li, and Raymond (2025) studied the staggered introduction of an assistant among 5,172 customer support agents. Access increased issues resolved per hour by 15% on average, with substantial differences across workers. Less experienced agents benefited more from the assistance. The authors also report evidence consistent with learning, including improvements persisting during tool outages.

The study describes one company and a particular support arrangement. It suggests how assistance can diffuse useful professional practices, with effects varying by experience and starting skill. For another organization, the local question concerns which tasks and people would benefit, and under what conditions. A business case needs observations from that proposed use rather than transferring the average directly.

Performance depends on what is delegated

Dell'Acqua and colleagues tested GPT-4 assistance in a preregistered experiment involving 758 consultants. First circulated as a working paper in 2023, the study was published in Organization Science in 2026. On tasks within the tested system's capability frontier, assisted participants completed more tasks, worked faster, and produced higher-quality outputs. On a task selected to fall outside that frontier, assistance reduced the likelihood of a correct solution.

The authors describe a jagged technological frontier. Tasks that appear similarly difficult to professionals can differ in whether AI assistance helps or harms performance. The frontier concerns the systems and tasks evaluated, and it changes as capabilities and work situations evolve. Professionals therefore need a basis for deciding which parts of a workflow to delegate and which require a different approach.

Time saved in drafting also needs to be considered alongside verification and correction. Producing a first proposal and using it in a consequential decision involve different standards. An evaluation can track both stages. That gives management a more informative account than a target to increase AI use across the workforce.

More information changes the demands on judgment

A fluent response may rest on an unsuitable assumption or omit a constraint known to the profession. Evaluating it requires knowledge of the situation and the consequences of using the output. The checking process needs to fit those consequences. A low-stakes draft and a proposal affecting a customer or major investment warrant different kinds of review.

Lee and colleagues (2025) surveyed 319 knowledge workers who described 936 examples of generative AI use. Greater confidence in AI was associated with less reported critical thinking. Respondents also described shifts toward verification and integration of outputs. These are self-reported practices and associations, useful for investigating how people allocate effort rather than establishing a causal decline in intellectual ability.

The claim that AI “frees up time to think” needs examination in actual work. Saved time might go into more thorough checking, other tasks, or higher expected output. Those arrangements offer different opportunities to learn. Management needs to specify the professional contribution expected after a first answer becomes available. Otherwise, adoption measures leave the resulting work largely unexplained.

How should organizations redesign knowledge work?

Introducing an assistant changes questions about how information becomes action. Someone frames the request and supplies context. Someone examines the proposal and decides whether it can be used. Disagreement needs a route to resolution when an output conflicts with professional experience. These are work-design decisions alongside the technical choice of a tool.

OperationQuestionWork arrangement to specify
Frame the requestWhat problem needs solving, and what context does the system need?The intended result and relevant boundaries before using the tool.
Review a proposalHow will accuracy and suitability for the situation be assessed?Verification sources and profession-specific quality criteria.
Decide on useWho can accept the proposal, and who resolves a disagreement?Decision authority and a route for further review.
Learn from correctionsWhat changed between the initial output and the accepted result, and why?Annotated cases and revisions to practice based on observed errors.

Quality criteria help specify what a useful proposal looks like before assistance is arranged. Depending on the task, the criteria may concern factual accuracy, suitability for a particular decision, or compliance with professional requirements. Employees need a way to explain why they accepted or revised the output. That account also supports selection of an appropriate verification method.

A disagreement process needs authority and time. If a generated proposal conflicts with professional experience, the organization needs to identify who examines the evidence and who decides. A rule permitting challenge has limited practical value when nobody has time to use it. An automatic expectation to follow the output creates another set of consequences for responsibility and learning.

These questions connect to AI and management. Accepting and using an output is a professional act with an owner and consequences. Training can prepare people for that act when the work arrangements support it. Its content therefore starts with the relevant situations rather than a generic aim to make everyone an advanced user.

Adaptability needs opportunities to learn

The word “cognitive” can draw attention toward individual abilities while obscuring the organization's circulation of knowledge. The OECD's account already distinguished access to information from the know-how required to select and use it. AI makes that distinction practical in new ways. A team needs opportunities to compare interpretations and retain what it learns from corrections.

One possible design reserves time for professionals to explain choices on selected work cases. Comparing the generated proposal with the accepted result can reveal a missing skill or a rule needing revision. Annotated cases provide material for subsequent decisions and learning. Their usefulness is evaluated through what changes in later practice rather than the number of cases collected.

Workplace AI training can include judgment and verification alongside tool operation. It needs to connect with available time and assigned responsibilities. Measures such as output quality, rework, and the quality of decisions prepared with assistance help assess the arrangement. Rising use without better work is a reason to revisit the design.

Frequently asked questions

Is cognitive economics about minimizing mental effort?

Here, it concerns the role of cognition in economic phenomena, following Walliser's research perspective. Minimizing mental effort is a different use of related language. Questions about attention and the effort of changing practices are addressed separately in the cognitive cost of change. Keeping the meanings distinct avoids moving between different explanations without notice.

Does this concern only office-based work?

Knowledge also circulates in technical and production work. Interpreting a signal, transferring know-how, or deciding how to respond to an incident involves cognitive activity in material conditions. The relevant analysis needs to examine the particular occupation and work setting. AI assistance is one possible influence within that activity.

Have we entered a new economic era?

That expression offers an interpretation of ongoing developments. The cited studies establish changes in particular work settings and tasks. A broader economic diagnosis requires evidence at a different level. A company can begin with the tasks changing in its own operations and the outcomes it can observe.

What should a leader measure first?

Start with the result the workflow needs to deliver and the operations being delegated. Compare quality, time including review and correction, and the consequences of errors. Usage information then helps explain exposure to the tool. It becomes part of an account of the work rather than a substitute for the outcome.

About the author

Maxime Rabéchault is the founder and president of UNREST, a change enablement and organizational development consultancy grounded in applied behavioral science. Trained at Université Paris Dauphine-PSL, the Conservatoire national des arts et métiers (CNAM), University College London (UCL), and Sciences Po Paris, he also teaches as an adjunct at IAE Paris-Sorbonne. His work connects observable professional practices with the organizational conditions that make transformation possible.

References

Walliser, B. (2000). L’Économie cognitive. Odile Jacob. https://www.odilejacob.fr/catalogue/sciences-humaines/economie-et-finance/economie-cognitive_9782738107695.php

OECD (1996). The Knowledge-Based Economy. OCDE/GD(96)102. https://one.oecd.org/document/OCDE/GD(96)102/En/pdf

Brynjolfsson, E., Li, D., & Raymond, L. (2025). Generative AI at Work. The Quarterly Journal of Economics, 140(2), 889–942. https://doi.org/10.1093/qje/qjae044

Dell’Acqua, F. et al. (2026). Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality. Organization Science, 37(2), 403–423. First circulated as a working paper in 2023. https://doi.org/10.1287/orsc.2025.21838

Lee, H.-P. et al. (2025). The Impact of Generative AI on Critical Thinking: Self-Reported Reductions in Cognitive Effort and Confidence Effects From a Survey of Knowledge Workers. CHI ’25, article 1121, 1–22. https://doi.org/10.1145/3706598.3713778