Does AI Adoption Increase Productivity?

Source: https://www.bok.or.kr/portal/bbs/B0000347/view.do?nttId=10098529&searchCnd=1&searchKwd=&depth2=201106&depth=201106&pageUnit=10&pageIndex=1&programType=newsData&menuNo=201106&oldMenuNo=201106

It has been almost three years since generative AI like ChatGPT emerged. AI has already become an indispensable tool for many workers. Scenes of delegating report drafts to AI or analyzing data with AI are no longer unfamiliar. As of 2025, over half (51.8%) of employed workers in Korea are using generative AI in their work, and this adoption rate is about eight times faster than the spread of the internet in its early days [Figure 1].

However, while tasks seem to be processed faster with AI, the country’s productivity indicators have shown little change over the past three years since the launch of generative AI [Figure 2]. In this blog post, we will examine whether AI adoption truly reduced working hours, whether that saved time translated into productivity gains, and what is needed to realize AI’s productivity benefits[1].

A graph showing the adoption rate by years since commercialization for generative AI and the internet. The horizontal axis is years since commercialization, the vertical axis is adoption rate (in percent). The internet adoption rate starts at year 0 just after commercialization, reaches slightly above 5% around year 3, and rises in an S-curve over 30 years. In contrast, generative AI already exceeds 50% adoption by year 3. This shows that generative AI is being adopted much faster than the internet was in its early days. For reference, 1995 is set as the year of internet commercialization, and 2022, when ChatGPT was released, is set as the year of generative AI commercialization. Source: ITU. (https://www.bok.or.kr/crosseditor/attachs/images/000059/20260617142143163_HNWSWWKZ.png)

A line graph showing quarterly labor productivity per hour (real GDP divided by total hours worked) from Q1 2010 to Q4 2025, normalized to Q1 2010 = 100. The productivity index rises gradually from 100 in Q1 2010 to about 146 in Q4 2025. The trend line based on Q1 2010–Q3 2022, extended to the present, nearly coincides with the actual productivity curve. This indicates that despite the spread of generative AI, labor productivity per hour has not clearly accelerated beyond its long-term trend. Source: Bank of Korea and national data. (https://www.bok.or.kr/crosseditor/attachs/images/000059/20260617142153651_001ZWJLZ.png)

AI Clearly Saves Time

Workers using generative AI took, on average, 3.8% less time to complete the same tasks. Based on a 40-hour work week, this means saving about 1.5 hours per week [Figure 3]. Assuming all this saved time is reinvested into productive work, the potential productivity improvement is estimated at about 1.0 percentage point[2].

A distribution graph with the horizontal axis showing the work time saving rate (in percent) from AI use, and the vertical axis showing density. The distribution peaks around 0% work time saving rate at about 25% density, with most workers concentrated between -10% and +10%. The tails extend to about -25% on the left and +25% on the right. Overall, the distribution has a high peak near 0% and is slightly skewed to the positive side, with an average work time saving rate of about 3.8%. (https://www.bok.or.kr/crosseditor/attachs/images/000059/20260617142202914_BKT27GWC.png)

By individual characteristics, the more AI was used, the greater the time savings, and workers with shorter tenure benefited more from AI. AI plays an ‘equalizing role[3]’ by partially compensating for lack of experience.

A graph showing estimated coefficients and 90% confidence intervals from a regression analysis with work time saving rate as the dependent variable. Each coefficient represents the average difference in work time saving rate for each characteristic relative to the reference characteristic. Male: +0.2, CI: -0.87 ~ 1.19 Self-employed: +0.9, CI: -0.29 ~ 2.10 Age 15-29: +1.6, CI: 0.01 ~ 3.14 Age 30-39: -0.2, CI: -1.12 ~ 0.76 Age 40-49: +0.8, CI: -0.11 ~ 1.77 College graduate: +1.0, CI: 0.20 ~ 1.80 Graduate school graduate: +0.6, CI: -0.80 ~ 2.03 Managerial: -0.6, CI: -2.01 ~ 0.90 Professional: +1.4, CI: 0.27 ~ 2.51 Service: -1.5, CI: -3.05 ~ 0.03 Sales: -1.7, CI: -3.26 ~ -0.17 Skilled: -1.5, CI: -2.78 ~ -0.15 Machine operator: -0.8, CI: -2.32 ~ 0.70 Laborer: -2.3, CI: -3.94 ~ -0.68 Labor supply elasticity (top 50%): +0.2, CI: -0.51 ~ 0.95 AI usage time (top 50%): +3.3, CI: 2.46 ~ 4.23 Tenure (top 50%): -1.3, CI: -2.14 ~ -0.44 Statistically significant variables: age 15-29, college graduate, professional, sales, skilled, laborer, AI usage time, tenure. Reference characteristics: female, wage worker, age 50-64, high school graduate, clerical, manufacturing. Controls: region, income, assets, working hours. (https://www.bok.or.kr/crosseditor/attachs/images/000059/20260617142213517_R7K38VNU.png)

Saved Time Did Not Translate into Productivity

Although workers saved about 1.5 hours per week thanks to AI, one might expect them to accomplish more work in that time. However, the correlation coefficient between work time saving and increase in work output was essentially zero [Figure 5]. We call this phenomenon the ‘AI productivity disconnect.’ AI increased the speed of individual tasks, but this effect has not translated into overall productivity.

A scatter plot showing the relationship between work time saving rate and change in work output. The horizontal axis is individual work time saving rate (%), the vertical axis is the rate of change in work output (%), with dot size representing population post-stratification weights. Most observations are concentrated around 0-10% work time saving rate and 0% change in work output, with wide vertical dispersion. The regression line is nearly horizontal, and the correlation coefficient between the two variables is about 0.00. This means workers who saved more time did not necessarily increase their work output, illustrating the disconnect between time savings and productivity. (https://www.bok.or.kr/crosseditor/attachs/images/000059/20260617142222838_ZB052GL0.png)

Productivity Increased in Certain Groups

However, a closer look reveals exceptions. Self-employed workers, young adults, and professionals were able to use the time saved by AI to increase their work output. The characteristics of these groups—self-employed individuals whose performance directly translates into income, young adults who quickly adapt to digital technology, and professionals with high work autonomy—appear to drive this difference.

A graph showing estimated coefficients and 90% confidence intervals from a regression analysis with the rate of increase in work output as the dependent variable. Each coefficient represents the additional effect of each characteristic relative to the reference characteristic. Male: -0.2, CI: -0.64 ~ 0.19 Self-employed: +1.0, CI: 0.32 ~ 1.78 Age 15-29: +0.6, CI: 0.00 ~ 1.17 Age 30-39: +0.6, CI: 0.05 ~ 1.08 Age 40-49: -0.11, CI: -0.56 ~ 0.33 College graduate: +0.4, CI: -0.13 ~ 0.83 Graduate school graduate: +0.4, CI: -0.47 ~ 1.19 Managerial: +0.1, CI: -0.59 ~ 0.74 Professional: +0.7, CI: 0.17 ~ 1.18 Service: -0.3, CI: -0.93 ~ 0.43 Sales: -0.9, CI: -1.99 ~ 0.19 Skilled: -0.1, CI: -0.96 ~ 0.80 Machine operator: -0.6, CI: -1.35 ~ 0.08 Laborer: -0.4, CI: -1.37 ~ 0.50 Labor supply elasticity (top 50%): -0.8, CI: -1.34 ~ -0.19 AI usage time (top 50%): +0.5, CI: 0.03 ~ 0.96 Tenure (top 50%): -0.4, CI: -0.97 ~ 0.09 Statistically significant variables: self-employed, age 15-29, age 30-39, professional, labor supply elasticity, AI usage time. Reference characteristics: female, wage worker, age 50-64, high school graduate, clerical, manufacturing. Controls: region, income, assets, working hours. (https://www.bok.or.kr/crosseditor/attachs/images/000059/20260617142232806_5H4HI344.png)

Why Has Productivity Not Increased?

There are four main potential factors. 1 (AI diffusion remains at the task level) Current AI tends to be limited to ‘specific tasks’ rather than ‘entire workflows.’ According to our survey, only 4.4% of tasks saw a significant reduction in work time of 20% or more, indicating that time savings are still limited. 2 (Rigidity of work processes) Without organic adjustments in corporate culture, worker behavior, and work procedures, simply introducing AI makes it difficult to achieve substantial performance improvements. This aligns with our earlier finding that relatively clear productivity gains were observed in groups with high work autonomy, such as the self-employed and professionals. 3 (Bottlenecks in the production process) Even if a large part of the workflow is streamlined, if a bottleneck exists at a specific stage, the overall work can still be delayed. For example, no matter how quickly data analysis and report writing are done using AI, if the approval process is slow, the productivity gains are diminished. 4 (Misaligned incentive structures) If there is little reward for additional performance, workers may have little incentive to reinvest spare time into productive activities. This is also consistent with the phenomenon of productivity gains observed in groups with a strong link between performance and reward, such as the self-employed and professionals.

However, there is no need to be discouraged. The current productivity disconnect may simply be a typical transitional phase observed in the early stages of adopting general-purpose technologies, similar to the J-Curve[4] or the Solow Paradox[5].

How Can We Increase Productivity?

First, for ‘standardized tasks’ with clear evaluation criteria (e.g., report summarization, data organization), the workflow itself should be redesigned around AI to channel saved time into more valuable activities. On the other hand, for ‘open-ended tasks’ where human judgment and creativity are key (e.g., new business development, R&D), AI should be used as an assistant while further developing human capabilities. In particular, it is essential to redesign learning pathways so that new and junior employees do not miss important skill-building opportunities when they delegate basic tasks to AI. While AI’s potential is clear, the role of people and institutions will be crucial in converting that potential into real performance. We will continue to monitor and analyze various leading indicators that explain the productivity transition and strive to provide policy recommendations.

• [1]For details, refer to BOK Issue Note No. 2026-12, “Does AI Adoption Increase Productivity? Analysis of the First Three Years” (https://www.bok.or.kr/portal/bbs/P0002353/view.do?nttId=10098322&searchCnd=1&searchKwd=&depth2=201156&depth3=200433&depth=200433&pageUnit=10&pageIndex=1&programType=newsData&menuNo=200433&oldMenuNo=200433).
• [2]This estimate is based on the strong assumption that all saved time is reinvested into productive activities and a production function approach. It should be interpreted as an upper bound of the actual productivity increase effect.
• [3]This is consistent with existing studies showing that generative AI reduces productivity gaps between workers with different experience levels (Brynjolfsson et al., 2025; Dell’Acqua et al., 2026; Cui et al., 2024; Hofmann et al., 2024).
• [4]This refers to a phenomenon where productivity temporarily drops when a new technology is first introduced due to adaptation, but then rises sharply in a ‘J’ shape once the technology is fully mastered.
• [5]Professor Robert Solow famously said in 1987, “You can see the computer age everywhere but in the productivity statistics,” pointing out the paradoxical situation where technological progress does not immediately translate into productivity gains.