The AI Productivity Paradox: A Central Bank Measured It

A central bank just put a number on the thing every office worker already suspected. Generative AI is saving people time. That time is not turning into output.

The Bank of Korea surveyed 5,512 workers and found that those who use generative AI at work cut their working hours by 3.8%, roughly 1.5 hours a week. Then it checked whether that saved time showed up as more output. The correlation was near zero. Not small. Near zero. People are getting their work done faster and producing about the same amount as before.

This is the AI productivity paradox, and it is the most important economic story about AI that almost nobody is pricing in. Hundreds of billions of dollars have gone into building these models. The measured productivity payoff so far, even under the most generous assumptions, tops out around 1%.

Does AI actually improve productivity?

Yes, but only barely. Generative AI improves task-level efficiency, saving the average worker about 1.5 hours a week, according to the Bank of Korea. But those time savings show near-zero correlation with actual output, and the best-case productivity gain is roughly 1.0%. The work gets faster; the measured output barely moves.

That split is the whole paradox. AI writes the first draft, the boilerplate, the summary, and it does it faster, so the efficiency is real and easy to feel. Productivity, the output that makes a company or an economy richer and justifies the capital, has not shown up. The same June 2026 issue note calls the link between time saved and output growth “essentially zero.”

What the Bank of Korea actually found

Two pieces of research carry this story, both from the Bank of Korea’s research department, both worth reading at the source.

The first, BOK Issue Note 2025-22, Rapid Adoption of Artificial Intelligence and Its Productivity Effects: The Case of Korea (Suh, Oh, and Kim, November 2025), is built on a survey of 5,512 employed Koreans. The adoption numbers alone are striking. 63.5% of those surveyed use generative AI in some form, and 51.8% use it for work, which the bank notes is close to double the work-use rate in the United States. Korea is not a laggard here. It is one of the most AI-saturated workforces on earth, which is exactly what makes it a clean test case. If the productivity dividend exists anywhere, it should be visible in Korea first.

The survey put the time saving at 3.8% of a 40-hour week. Then it estimated the productivity gain you would get if all of that time were redirected into productive output: about 1.0%. That is the ceiling, the generous case, the world where nothing leaks.

AI's productivity gain in Korea: best case versus what was realized

Item	Best case if time converted	Realized in the data
Productivity gain	1.0%	0.0%

The second paper is where the floor falls out. BOK Issue Note 2026-12, Does AI Adoption Improve Productivity? Effects Over the First Three Years (Suh, Oh, and Yoon, June 2026), checked whether the time savings actually became output. They did not. The note states plainly that the time savings “do not translate into actual output growth,” with essentially zero correlation between the two. The academic version of this work, a February 2026 preprint, Generative AI and the Reallocation of Time (Suh and Oh), is blunter still: workers capture the benefit “primarily as on-the-job leisure rather than increased output.” The hour you save does not go back into the machine. It goes to you.

This is not new: the Solow paradox returns

In 1987, the economist Robert Solow wrote a line that has outlived almost everything else from that decade of productivity debate: “You can see the computer age everywhere but in the productivity statistics.” Companies were buying computers by the millions. Output per hour was flat. The gap between the visible technology and the invisible payoff got a name, the Solow paradox, and it lasted the better part of a decade before the productivity numbers finally moved in the late 1990s.

The lesson from that episode is not that the computers were useless. It is that a general-purpose technology pays off only after the work around it is rebuilt. Factories had to be relaid out. Offices had to rewire how decisions got made. The hardware showed up years before the org chart caught up, and productivity tracked the org chart, not the hardware.

AI looks like it is running the same play. The tool is in everyone’s hands. The workflows, incentives, and reporting structures that would turn a faster draft into a bigger result are mostly untouched. If the historical pattern holds, the productivity statistics move once the work is redesigned, not when the next model ships. That could be a year. It was closer to a decade last time.

Why time saved doesn’t become output

Three forces sit between the hour you save and the output your employer measures. None of them is exotic. All of them are doing real damage to the AI productivity case right now.

The Jevons loop. When something gets cheaper, you use more of it. The 19th-century economist William Stanley Jevons noticed that more efficient steam engines burned more coal, not less, because cheaper power meant more uses for it. The office version: AI makes writing a report faster, so people write more reports, longer reports, more often. The time cost of producing the report fell, so the quantity of reports rose to fill the gap. The reviewing, formatting, and circulating of all that extra material then eats the saved time. Faster production created more to consume, and the net change in useful output rounds to nothing.

The rework tax. A meaningful share of what AI produces has to be checked and fixed, and the checking is not free. Workday research published in January 2026, a survey of 3,200 AI-using employees fielded by Hanover Research, found that nearly 40% of the time workers save with AI is offset again by rework, correcting and rewriting low-quality output. Save six hours, hand back nearly two and a half to fixing what the model got wrong. The headline time saving and the net time saving are different numbers, and the gap is the slop.

No reason to convert. Even when the time is genuinely freed, almost nobody is paid to turn it into more output. This is the quiet one, and it might be the biggest. If you finish your work an hour early and your pay is the same whether you produce more or not, the rational move is to keep the hour. The Bank of Korea’s finding that workers bank the savings as on-the-job leisure is not a moral failure. It is what the incentive structure asks for. BCG’s June 2026 report AI at Work: Why Strategy Matters More Than Tools put a number on the management half of this: about 66% of regular AI users report getting limited or no guidance on what to do with the time they save. The time is freed and then left on the floor.

It is not just Korea

The easy objection is that this is one country, one survey, one methodology. It is not. The independent evidence lines up uncomfortably well, and it is worth being precise about what each source is.

Source	What it is	Finding
Bank of Korea (2025–2026)	Central-bank survey, 5,512 workers	3.8% time saved, near-zero correlation with output, ~1.0% potential gain
NBER Working Paper 34984 (2026)	Academic survey of ~6,000 corporate executives, four countries	About 90% report no productivity impact from AI over the past three years
BCG AI at Work (June 2026)	Consultancy workforce survey	Workers save hours weekly; about 66% get little or no guidance on the saved time
Workday (January 2026)	Vendor workforce study, 3,200 employees	Nearly 40% of AI time savings lost to rework

The NBER paper deserves a word on status, since it is the heaviest hitter. Artificial Intelligence, Productivity, and the Workforce: Evidence from Corporate Executives (2026) is a working paper, not peer-reviewed, but it surveys roughly 6,000 senior executives across the United States, the United Kingdom, Germany, and Australia. About 69% of their firms use AI. Roughly 90% report no detectable impact on productivity or employment over the prior three years. These are the people who authorized the spending, reporting on their own firms, and they are not seeing it either.

For a sober academic read on why, the University of California’s Seven Myths about AI and Productivity (California Management Review, October 2025) is the best single overview. Its through-line matches the central banks: the headline lab-experiment gains are real but do not survive contact with how organizations actually run.

What would actually unlock productivity

The Bank of Korea does not stop at the diagnosis, and its prescription is the most useful part for anyone deciding what to do next. Turning efficiency into productivity, the bank argues, takes three things that have nothing to do with the models themselves: organizational redesign, reallocating people toward the work AI cannot do, and performance-based incentives that actually reward converting saved time into output.

Read that list against the three leaks above and it maps almost perfectly. Redesign attacks the Jevons loop by deciding what work should stop, not just speed up. Reallocation moves the freed hours toward higher-value tasks instead of leaving them idle. Incentives give people a reason to convert. The technology is not the bottleneck. The management is.

There is a smaller signal worth holding onto. The gains are not zero for everyone. The Bank of Korea found larger effects for less-experienced workers, an equalizing pattern, and the corroborating research consistently finds real output gains among the self-employed, professionals, and younger workers, the people whose pay is tied directly to what they produce. Where the incentive to convert exists, the conversion happens. That is the tell. The paradox is not a property of AI. It is a property of how most jobs are paid.

For anyone weighing the spend, this reframes the question. The cost side of AI is increasingly knowable: the AI model tracker and the cost-per-task calculator cover what the underlying models actually cost to run. The return side is the open question, and the Korean data says the return is currently being paid out in shorter workdays, not bigger output. That is a real benefit to workers. It is just not the benefit the capital was underwriting.

It also helps explain a pattern visible on the map of who uses AI most. South Korea is one of the countries that uses AI far more than its income predicts, which is precisely why it is the right place to measure the payoff first. The usage is there. The output dividend, for now, is not.

Frequently asked questions

Does AI save time at work?

Yes. The Bank of Korea found generative AI cuts working time by about 3.8%, roughly 1.5 hours a week per worker who uses it. The time saving is real and consistent across multiple studies. The open question is what happens to the saved time, not whether it exists.

Why don't AI time savings turn into more output?

Three reasons. Faster output invites more output, so people produce more reports rather than more value (a Jevons effect). Rework on AI output eats a chunk of the savings. And most workers have no pay incentive to convert saved time into extra production, so they bank it as shorter effective workdays.

Is the AI productivity paradox permanent?

Probably not. It mirrors the 1980s Solow paradox, when computers were everywhere except in the productivity statistics until the late 1990s, after work was redesigned around them. The Bank of Korea argues the gap closes once organizations restructure work and tie pay to output, not when models get better.

Which workers actually see productivity gains from AI?

The ones whose pay is tied to output. Research consistently finds real gains among the self-employed, professionals, and younger or less-experienced workers, while salaried employees in rigid structures tend to bank the saved time as leisure rather than extra output.

Sources

• Suh, D., Oh, S., and Kim, M. (2025). Rapid Adoption of Artificial Intelligence and Its Productivity Effects: The Case of Korea. BOK Issue Note 2025-22, Bank of Korea. bok.or.kr
• Suh, D., Oh, S., and Yoon, J. (2026). Does AI Adoption Improve Productivity? Effects Over the First Three Years. BOK Issue Note 2026-12, Bank of Korea. bok.or.kr
• Suh, D., and Oh, S. (2026). Generative AI and the Reallocation of Time: Productivity, Leisure, and Fulfilling Work (preprint). arXiv:2602.12695. arxiv.org
• National Bureau of Economic Research (2026). Artificial Intelligence, Productivity, and the Workforce: Evidence from Corporate Executives (working paper, not peer-reviewed). NBER Working Paper 34984. nber.org
• California Management Review (2025). Seven Myths about AI and Productivity: What the Evidence Really Says. UC Berkeley Haas. cmr.berkeley.edu
• Boston Consulting Group (2026). AI at Work: Why Strategy Matters More Than Tools (fourth edition, June 2026). bcg.com
• Workday (2026). Beyond Productivity: Measuring the Real Value of AI (research fielded by Hanover Research, November 2025). newsroom.workday.com
• Solow, R. (1987). We’d Better Watch Out. New York Times Book Review, July 12, 1987 (origin of the productivity-paradox quotation).