Increased use of generative AI within companies does not appear to automatically lead to increased productivity. According to data reviewed by Persol Research and Consulting, the proportion of full-time Japanese employees who use AI in their work rose from 41.9% in 2025 to 54.3% in 2026, but this spread does not necessarily mean fewer overtime hours or improved organizational results.
In an interview with ITmedia, Yuji Kobayashi, chief researcher at Persol Research and Consulting, said that producing value with AI sometimes requires stepping away from it. The idea does not mean rejecting the tools, but rather preserving the sources of capabilities that models do not produce on their own: contact with reality, judgment about priorities, the creation of meaning, and learning from actual people and environments.
Faster Work Does Not Always Address the Bottleneck
Kobayashi organizes the problems of using AI in organizations into three areas. The first is the “problem of saving time in the wrong direction”: tasks such as gathering information, exchanging emails, or preparing a presentation may become faster, while the bottleneck remains elsewhere. He gives sales as an example; AI may speed up preparation of a proposal, but identifying the customer, understanding their needs, or closing the deal may have a greater impact on the final result.
The second area is the “problem of disappearing available capacity.” The study indicates that 60% of the time saved through AI was reinvested in work, and that more than 70% of this time went to daily tasks. Employees who use the tool most may also take on additional responsibilities, such as internal training related to it, instead of converting the freed-up time into learning or innovation.
The third area is the “problem of disappearing differentiation.” When writing job applications or preparing marketing materials becomes available to everyone through similar tools, the advantage once provided by writing skill alone declines. Kobayashi believes that what a company can accomplish with AI can also be accomplished by competitors and customers, making task acceleration alone less capable of producing an advantage that can be priced.
The Crucial Skill Is Managing the Entire Loop
According to the results of research by Persol Research and Consulting, high performers in an environment that uses AI differ from high performers in an environment that does not. In the traditional case, performance was associated with understanding one’s role, acting quickly, and managing relationships with others. In an AI environment, however, people capable of turning work with the tool into a recurring loop of information gathering, decision-making, execution, review, and learning stood out.
This loop consists of five skills: “gathering” and “decision-making” in the earlier stages, followed by “verification,” “deconstruction,” and “storage” in the later stages. The process begins with gathering data, including firsthand information from the real world that is difficult for AI to access. The user then determines priorities and what should be delegated to the tool, reviews the outputs with others, modifies them or dismantles inappropriate assumptions, and finally shares the experience within the organization so that it becomes collective knowledge.
The importance of these steps lies in the fact that the practical danger is not always a “hallucination” whose error can be easily detected. The greater danger, according to Kobayashi, is producing an answer that appears generally correct and reasonable but is unsuitable for this customer, this timing, or this context. Therefore, the user’s value lies not in accepting the output, but in transforming a “general answer” into an answer suited to the situation.
Multidimensional Thinking and Curiosity Beyond the Screen
The study also identified two intellectual traits among high performers: “solid thinking,” meaning raising the level of abstraction, connecting the subject from multiple angles, and seeing the overall picture; and “animal spirits,” an expression used by economist John Maynard Keynes to describe irrational human motives in economic activity. Kobayashi uses it here to refer to curiosity and a willingness to try what seems interesting.
Multidimensional thinking helps organize the stream of proposals produced by models, while curiosity drives people to try new uses instead of limiting themselves to familiar tasks. However, the study, as presented by the source, did not find that “animal spirits” automatically grow with longer periods of AI use. Rather, Kobayashi points to the possibility of “skill acquisition inhibition,” in which reliance on the tool causes some of the user’s abilities to decline.
What Changes in Practice?
The most important editorial conclusion is that measuring AI’s success by the number of employees who use it or the number of minutes it saves may be incomplete. The more useful question is: Has the time and effort shifted to a point that actually affects revenue, decision quality, or the resolution of a customer’s problem? Training programs should also go beyond prompt writing to include gathering evidence, setting priorities, collective verification, and documenting knowledge.
Kobayashi suggests that developing thinking and curiosity requires activities outside the tool, such as interacting with different people and environments, engaging in side projects or nonprofit and volunteer activities, going deeply into one subject, and reading. This is not a proven formula for every organization, but it reflects the limits of what continuous use of the tool alone can provide. The study was based on a preliminary survey of 20,000 employees, followed by a main study of 3,000 full-time employees. The follow-up study took place between April 22 and May 7, 2026, while the new study was conducted between April 23 and May 7, 2026, with follow-up participants from the October 2025 study and new participants.