On the twentieth of each month, a product planner at a mid-sized company in Seoul sees the same line on her card statement: 49,000 won for the AI assistant she pays for herself. Her employer has no licence for it. She drafts meeting notes with it, rewrites vendor emails and builds the first version of every slide deck, and by her own count each of these jobs now takes half the time it did a year ago. She has stopped noticing the saving. The deadlines have not moved, her manager has not asked what she does with the spare minutes, and most days she spends them on the next email.

The scenes that open and close this article are illustrations; the people in them are invented.

Her habit now has a measurement. On 17 November the Bank of Korea’s employment research team published an issue note, “The Spread of AI and Its Productivity Effects”, built on a survey of 16,342 employed people aged 18 to 64. The fieldwork ran from 14 August to 23 September. It repeats a survey from a year earlier, which makes it one of the few central-bank datasets that can show how fast the use of generative AI at work moves from one year to the next. The numbers rose sharply. The link between the time saved and more or better output stayed close to zero.

What rose

Of the respondents, 66.5% said they use generative AI at work, up from 51.8% in 2025. The share using it regularly went from 17.1% to 30.6%, and average weekly use from 5.9 to 8.0 hours, according to Seoul Economic Daily’s account of the note. Seoul Shinmun reports that this is the highest rate among the countries the bank compared, ahead of the United States at 43.0%.

Users reported that AI cut their working time by 9.2% on average, against 3.8% a year earlier. On a 40-hour week that is about 3.7 hours saved, where last year it was 1.5. The share of tasks on which the reduction exceeded 20% rose from 4.7% to 17.4%. These figures come from what workers say about their own work; the survey does not time them.

Money has followed use. Among people who use AI at work, 28.4% pay for a subscription themselves and 28.2% have their employer pay. The self-paying users spend 49,000 won, about 36 dollars, a month, and the bank extrapolates roughly 2.8 trillion won a year across all workers. The savings differ by who pays. Workers who paid for AI or whose company paid for it reported cutting working time by 11.4% and 11.6%, against 8.3% for people on free tools. The bank warns that gaps in access to paid tools, and in the ability to use them, could widen differences between workers. Differences by education, which it did not find last year, are statistically significant this year.

What did not follow

The bank asked whether the time saved became more work or better work. The correlation between the rate of time saved and the increase in work volume is 0.08, and with the improvement in quality 0.09. Both are close to zero. The employment team’s head, Oh Sam-il, said last year’s survey showed almost no relationship between saved time and output, and that a weak one has begun to appear this year.

Where the time went is partly known. According to the survey, 33.8% of the saved time went to concentrating on core tasks and 29.3% to additional work, so 63.1% went back into work. Seoul Shinmun, citing the report, lists the rest: study and self-development 13.2%, leisure at the workplace 10.0%, personal errands 6.3% and leaving early 2.3%. A worker who spends the hour on a course, or on a coffee, is saving time without producing more, and the survey cannot say whether that is waste or a reasonable use of an hour.

The bank also finds that AI has not reached the centre of the job. The higher the share of core tasks in a worker’s duties, the less likely that worker was to have used AI in the same week. Where AI does reach core work, the gain is larger: the rate of quality improvement is 14.0 percentage points higher for workers with a high share of core tasks. The bank’s 61.1 trillion won estimate of the potential value of the saved time assumes that all of it goes back into production, and the bank says it cannot be compared with the 2.8 trillion won spent on subscriptions because the two rest on different calculations and coverage.

Why the saving grew

The bank offers a reading of why the savings more than doubled in one year. It found no statistically significant difference in time saved between workers who began using AI last year and those who began this year. Oh read this as less a matter of workers growing more skilled and more the technology improving. Active habits still counted: workers who wrote prompts and verified the output saw their time reduction rise by 3.6 percentage points more than those who did not.

Oh described the missing link as complementary assets. The technology cannot deliver its potential, he said, until reward systems and organisational structures change with it. The report’s recommendations say the same: “Policies should support workers in developing the skills to identify ways to use AI for their jobs and verify its outputs. Companies, meanwhile, should integrate AI into core tasks and improve their work flow and organizational structures so that time saved from AI use can be put to more valuable tasks.”

A year earlier

The 2025 issue note, “The Rapid Spread of AI and Its Productivity Effects: Based on a Household Survey”, estimated that the 3.8% saving implied a potential productivity gain of 1.0% for Korea, close to the 1.1% it gave for the United States. Korea used AI more but saved less time per user, 3.8% against 5.4% in the United States. The note also found that workers with shorter careers saved more time, which it read as AI narrowing differences in skill.

Other central banks are measuring the same gap from the other side. The Bank of England’s November Monetary Policy Report says the contribution of AI to UK growth so far reflects the building and installation of the technology rather than broader gains from its adoption, which is consistent with an impact on productivity that has been “relatively small so far”. Korea’s survey adds a view from inside the workplace: workers report faster work, while the link to more output is weak.

The limits matter. The survey is self-reported, and the bank’s own estimate of potential value rests on an assumption about how workers use saved time. Correlations of 0.08 and 0.09 describe a relationship across individuals in one year, not proof that no productivity gain exists, and this account rests on press coverage of the note and on the bank’s 2025 note, not on the full 2026 text.

The statement on the twentieth

The planner’s card statement is, in a small way, part of the bank’s data. She is one of the 28.4% who pay out of pocket, and the 49,000 won leaves her account whether or not anyone at her company changes how work is divided. The saved hours are real to her. Whether they become anything her employer can count depends on a decision made in a meeting she is not invited to, about what a planner’s day is for.

References

  • Bank of England (2026). Monetary Policy Report, November 2026. bankofengland.co.uk
  • Bank of Korea (2025). The Rapid Spread of AI and Its Productivity Effects: Based on a Household Survey (BOK Issue Note No. 2025-22; Korean title “AI의 빠른 확산과 생산성 효과: 가계조사를 바탕으로”). bok.or.kr
  • Kim, H. (2026). AI Use at Work Jumps in Korea, Saving 3.7 Hours a Week. Seoul Economic Daily. en.sedaily.com
  • Lee, H. (2026). Korean workers embrace AI, but productivity gains remain elusive: BOK. The Korea Times. koreatimes.co.kr
  • Lee, S. (2026). AI 확산, 일은 빨라졌는데 성과는 안 늘었다. Seoul Shinmun. seoul.co.kr