It has been nearly four years since ChatGPT launched. AI has already become an irreplaceable tool for many workers. It is no longer unusual to ask AI to draft a report or to use AI for data analysis. In fact, as of 2025, more than half of employed persons in Korea (51.8%) are using generative AI at work, and this pace is as much as eight times faster than the diffusion of the Internet in the past[Figure 1].
Although AI adoption certainly seems to have sped up work, Korea's productivity indicators have shown little change over the three years since generative AI was introduced[Figure 2].In this blog post, we examine whether AI adoption has indeed reduced working time, whether the time saved has translated into productivity gains, and what needs to be done to realize AI's productivity effects.
AI Does Save Time
Workers who use generative AI took, on average, 3.8% less time to perform the same tasks. Based on a 40-hour workweek, this amounts to saving about 1.5 hours per week[Figure 3]. If all of this time saved is assumed to be devoted to production, the potential productivity gain is estimated at around 1.0 percentage point
Looking across individual characteristics, the time-saving effect was larger among heavier AI users, and workers with less experience benefited more from AI's help. In this sense, AI appears to play an “equalizing role” that partly compensates for a lack of experience.
Time Savings Do Not Translate into Productivity Gains
One might expect that workers saving 1.5 hours a week thanks to AI would accomplish more during that time. However, the correlation between time savings and the increase in output was essentially zero[Figure 5]. We refer to this phenomenon as the "AI productivity disconnect." AI has accelerated individual tasks, but this effect has not translated into overall productivity.
Productivity Increased in Specific Groups
Looking more closely, however, there were exceptions. Self-employed workers, young workers, and professionals were using the time saved by AI to increase work volume. This difference appears to reflect the characteristics of self-employed workers, whose performance translates directly into income; the strengths of young workers, who adapt quickly to digital technologies; and the high degree of work autonomy among professionals.
Why Did Productivity Gains Not Appear?
There are four broad factors that may explain this.
① AI diffusion remains confined to the task level. Today's AI tends to be applied to “specific tasks” rather than “entire jobs.” In our survey, only 4.4% of tasks saw a substantial time reduction of 20% or more, indicating that the time-saving effect remains limited so far.
② Rigidity in workflows. Introducing AI alone, without organically adjusting corporate culture, worker behavior, and business processes, is unlikely to deliver real performance gains. This is also why relatively clear productivity increases were observed among self-employed workers and professionals, who enjoy greater autonomy over how they work, as noted above.
③ Bottlenecks in the production process. Even if a large share of a workflow becomes more efficient, the process as a whole can still be delayed if a bottleneck exists at a particular stage. For example, no matter how quickly AI handles data analysis and report writing, the productivity gain is diminished if the approval process itself remains slow.
④ Misaligned incentives. When rewards for additional output are weak, workers have little incentive to reinvest freed-up time into production. This, too, is consistent with the fact that productivity gains emerged in groups such as the self-employed and professionals, where the link between performance and reward is strong.
That said, there is no need for discouragement. The productivity disconnect observed today may simply be a typical transitional pattern seen in the early stages of adopting a general-purpose technology — similar to the J-curve or the Solow paradox.
What does it take to turn AI into productivity gains?
For “standardized tasks” with clear evaluation criteria for outputs — such as report summarization or data organization — workflows should be redesigned around AI so that the time saved is channeled into higher-value activities. For “open tasks” in which human judgment and creativity are central — such as new business development or R&D — AI should instead be used as an assistant, while continuing to build up people's own capabilities. In particular, as new and junior employees hand off basic tasks to AI, it will also be essential to redesign learning paths so that they do not lose important opportunities to build skills.
AI's potential is clear, but converting that potential into real performance gains will also depend on the role of incentives and institutions. We will continue to examine a range of data, track and analyze leading indicators that help explain this productivity transition, and work to offer relevant policy recommendations.
- [1] For further details, please refer to BOK Issue Note No. 2026-12, "Does AI Adoption Improve Productivity? Effects over the First Three Years."
- [2] This estimate is based on the strong assumption that all saved time is reinvested in production activities and on a production-function approach; it should therefore be interpreted as an upper bound on the actual productivity gain.
- [3] This is consistent with existing studies showing that generative AI reduces productivity gaps among workers by career experience (Brynjolfsson et al., 2025; Dell'Acqua et al., 2026; Cui et al., 2024; Hofmann et al., 2024).
- [4] This refers to the pattern in which productivity temporarily dips as workers adjust to a newly introduced technology, before rising sharply once the technology is fully mastered — tracing a shape resembling the letter “J.”
- [5] Professor Robert Solow observed in 1987 that the computer age could be seen everywhere except in the productivity statistics, highlighting the paradox that technological progress does not immediately translate into higher productivity.