AI Users Say They Save Hours. Average Weekly Work Hours Have Not Fallen

The average US private-sector workweek was 34.4 hours in September 2026. In November 2022, the month ChatGPT launched, it was 34.5. In between, the share of workers using generative AI on the job climbed to 39.2%, and they now report AI time savings worth about 2.2% of all US work hours. The saved time is real enough to measure. It just never turned up as shorter weeks.

The short answer

No, AI has not shortened the workweek. US average weekly hours were 34.5 when ChatGPT launched and 34.4 in September 2026. Workers say AI saves them about 2.2% of their hours, roughly 45 minutes a week. That time has gone into more work, new oversight tasks and slack inside the day, not into going home earlier.

What AI users say about AI time savings

Start with the claim, stated as carefully as the people who measure it state it. The best US series on AI time savings comes from the Real-Time Population Survey run by economists Alexander Bick, Adam Blandin and David Deming. The St. Louis Fed publishes it on FRED.

The question is specific. Workers who used generative AI for their job last week are asked how many additional hours they would have needed to finish the same work without it. Answers come in brackets from under 1 hour to more than 4 hours.

The answer has risen steadily. Hours saved as a share of all work hours went from 1.4% in Q4 2024 to 2.2% in Q2 2026. Over a similar stretch, the share of employed adults using generative AI for work in the past week rose from 28.2% to 39.2%, according to a FRED Blog post from August 2026.

The tracker counts every worker, not just users

That 2.2% is a workforce figure. Non-users count as zero, so the saving among actual users is several times larger. This matters for the comparison that follows, because average weekly hours are also a workforce figure. Like is being compared with like.

The FRED Blog post is candid about the method. It says the measurements are self-reported and "inherently approximate." I take that seriously, and I still think the series is the best US number available. It is nationally representative, it asks a concrete counterfactual question, and it has been run the same way for two years.

Korea measured almost the same thing

A second national survey points the same way. Donghyun Suh and Samil Oh of the Bank of Korea surveyed 5,512 Korean workers between May and June 2025. They found 51.8% used generative AI for work.

Among workers who used it in the prior week, working time fell by an average of 3.8%. Spread across the whole workforce, that becomes a 1.4% reduction in active hours. About half of users reported savings close to zero, so the average is carried by a minority who save a lot.

Two countries, two methods, one rough answer. Generative AI is returning somewhere between 1% and 2% of total work time. That is not trivial. Across a US private workforce putting in 34.4 hours a week, it is a measurable quantity of time, and it should be visible somewhere.

BLS average weekly hours since ChatGPT launched

The obvious place to look is the BLS average weekly hours series. It comes from the monthly payroll survey of employers, and it is the most-watched hours number in the US economy.

The September 2026 Employment Situation release states it plainly: "The average workweek for all employees on private nonfarm payrolls remained at 34.4 hours in September." For production and nonsupervisory employees, the figure was 33.8 hours.

Now set that against the history on FRED's AWHAETP series. The workweek was 34.3 hours in December 2019. It spiked to 35.0 in January 2021, when the pandemic had removed many shorter-hour jobs from the average. It was 34.5 in November 2022. Through all of 2025 it sat between 34.1 and 34.3. It has been 34.4 since August 2026.

The most exposed sectors did not shorten either

An economy-wide average could bury a real change in the sectors where AI use is heaviest. So check those sectors directly. Professional and business services averaged 36.6 hours in November 2022 and 36.7 hours in September 2026. The information sector, which includes software publishing and media, was 36.9 hours at both dates.

The information sector did wander in between. It reached 37.6 hours in September 2025 before settling back. That is a swing in both directions, not a trend toward shorter weeks. If generative AI were cutting the workweek anywhere, these two sectors are where it would show first, and it has not.

What the American Time Use Survey shows about work hours

Payroll hours have one weakness worth naming early. BLS defines them as hours paid, not hours worked. A salaried analyst who finishes by 3pm and a salaried analyst who works until 9pm can both appear in the payroll survey as a standard week.

The American Time Use Survey closes that gap. It asks people to reconstruct the previous day in a diary, so it records time actually spent working. It is the closest public measure to what the AI time-savings claim is about.

The result is the same flat line. In the 2022 survey, full-time employed people worked 8.4 hours on an average weekday they worked. In the 2024 survey the figure was again 8.4 hours. In the 2025 survey, released on 25 June 2026, it was 8.5 hours. Across all days they worked, full-time workers averaged 8.1 hours in both 2024 and 2025.

The first two rows measure hours and do not mention AI. The third row mentions AI and does not measure hours. No public US dataset does both for the same person, which is why this comparison has to be made across sources.

The 2025 diary data has a hole in it

The 2025 numbers deserve a caveat in the same sentence as the figure. All ATUS operations were suspended from 1 October to 12 November 2025 during the federal government shutdown. BLS states that it is "not possible to quantify the effect of the shutdown on the ATUS estimates."

The sample was also smaller than usual, at about 6,100 people against roughly 7,700 in 2024. So the 8.5 figure is weaker evidence than the 8.4 figures before it. It still lands where the payroll data lands, and the two sources are built in completely different ways.

The arithmetic of 45 missing minutes a week

Put the two numbers on the same basis. If generative AI saves 2.2% of all work hours, and the average private-sector week is 34.4 hours, the saving is about 0.75 hours. That is roughly 45 minutes per worker per week.

If that time had come off the schedule, the average workweek would now be around 33.7 hours. Instead it is 34.4. The payroll series moves in steps of 0.1 hours, which is 6 minutes, so a 45-minute shift would be about seven steps. It is not a rounding error that a noisy month could hide.

My position is that this gap is the most useful single fact in the AI productivity debate, and it gets almost no attention. The saving is real by the most careful available measure. The workweek is unchanged by the two most careful available measures. Both are true, so the time went somewhere else.

This is a narrower claim than the one made in the analysis of whether AI can pay for a four-day week. That piece asks whether the saving is large enough to fund a policy. This one asks whether any of it has reached the clock at all. The answer to the second question is no, and it explains the first.

It also fits the wider macro picture. Output per hour can rise while hours hold still, which is the pattern behind the gap between AI adoption and measured GDP growth. Flat hours do not disprove the savings. They tell you which channel the savings went through.

Where AI time savings go instead of shorter weeks

There are four places saved time can land: more output, new work, slack inside the day, or a shorter week. The evidence now points clearly at the first three. Each has at least one serious study behind it, and none of the studies finds the fourth.

Into more work, often unasked

Researchers Aruna Ranganathan and Xingqi Maggie Ye followed a US technology company of about 200 employees for eight months in 2025. They reported the early findings in Harvard Business Review in February 2026. Their summary is blunt: AI tools "didn't reduce work, they consistently intensified it."

They saw three changes. Employees "worked at a faster pace, took on a broader scope of tasks, and extended work into more hours of the day," often without being asked. The company had not mandated AI use. Workers did more because the tools made doing more feel possible and rewarding.

This is one company and the authors call it in-progress research, so it describes a mechanism rather than a national rate. It is still the clearest account of why a saved hour does not stay saved. The person holding it fills it, usually with work.

Into new tasks the tools create

The strongest evidence comes from Denmark, where survey answers can be linked to administrative records of pay and hours. Anders Humlum and Emilie Vestergaard's paper, revised in March 2026 under the title "Still Waters, Rapid Currents", estimates "precise null effects on earnings and recorded hours." It rules out effects larger than 2% two years after ChatGPT launched.

The authors also explain where the change went. Employers "absorb AI through task reorganization," including new tasks in content generation, AI oversight and AI integration. Reviewing model output, writing prompts and checking work for AI use are all new jobs inside old jobs. Each one spends part of the time the tool saved.

Software teams know this pattern well. Faster code generation moved the queue to review, which is the argument made in the analysis of code review as the new bottleneck. The hours did not disappear. They changed desks.

Into slack, and into the weekend

The Korean paper offers the least comfortable finding for anyone selling productivity. The correlation between a worker's reported time savings and their reported output change was 0.008, which is effectively zero. The authors conclude that "workers capture efficiency gains primarily as on-the-job leisure, rather than increasing their output." Among users, the share of time spent on on-the-job leisure rose by 1.3 percentage points.

Some of the time also appears to be moving later in the week. Workplace analytics vendor ActivTrak reported in March 2026 on 443 million hours of activity from 163,638 employees. In a subset of 10,584 users tracked 180 days before and after adopting AI, email activity rose 104% and chat and messaging rose 145%. AI users lost 23 minutes of focused time a day.

Across its whole 2023 to 2025 dataset, ActivTrak also found weekend work rising by more than 40%. Treat that as directional. ActivTrak sells monitoring software, its customers are not a random sample of employers, and a before-and-after design cannot separate AI from everything else that changed in the same months.

Where this argument is weakest

The case above rests on comparing different surveys with different designs. That has real limits, and some of them cut against the conclusion.

Paid hours and averages can hide a real shift

The payroll series counts hours paid per job, and the workforce has changed composition since 2022. If AI users shortened their weeks while other workers lengthened theirs, the average could stay flat. The sector data argues against this, since the most exposed sectors are flat too, but it does not rule it out.

Users are also a minority. With about 39% of workers using AI in a given week, a large change concentrated in a small group of heavy users would be diluted in any national average. The diary data is the better test of that, and its 2025 wave is the weakest year in the series.

Nobody knows what hours would have been without AI

This is the stronger objection. Flat hours only count against shorter weeks if hours would otherwise have stayed flat. A tight labour market, return-to-office mandates or weak hiring could have pushed hours up, with AI quietly pulling them back to level.

I do not think that story is likely. The workweek has held between 34.1 and 34.5 hours for most of the period since 2022, without a visible upward pressure for AI to offset. But it is not testable with public data, and anyone claiming certainty either way is going past the evidence.

Finally, every time-saving figure here is self-reported. If workers overstate savings, the gap with hours shrinks, and the puzzle partly dissolves. That is a real possibility. It would also mean the productivity case for AI rests on a smaller number than its promoters claim, which is not a better outcome for them.

What AI time savings mean for your team's hours

I write for the person who runs delivery, because the destination of saved time is a management decision whether or not anyone makes it on purpose. Default behaviour, as every study above shows, is that the time flows to whatever is loudest. Usually that is more messages, more review and more scope.

That default is the wrong one for most teams. Silent intensification produces the burnout pattern described in the analysis of AI pace and engineering burnout. It also makes your productivity gains invisible, because they never show up as anything you can count.

The alternative is to decide in advance where saved time should go. Pick one destination per team: more output, higher quality, shorter weeks, or deliberate development time. Then measure that destination, not the tool usage. If saved hours are meant to become output, your review criteria have to reward output, a shift covered in the piece on changing performance review criteria for AI.

I would not promise staff shorter hours on the strength of a self-reported number. I would promise them that saved time will not automatically become more meetings. That is a commitment you can actually keep.

Frequently asked questions

Where to start this week

Ask each team lead one question: if your team saved four hours last week using AI, where did those hours go? Write the answers down. Most will not know, and that blank is the thing to fix.

Then choose one destination for the next quarter and tell the team what it is. If you choose output, define the unit you will count. If you choose time back, put it on the calendar as a fixed block that meetings cannot claim. Either choice beats the default, which is giving the time to your inbox.

Related reading