What should happen to the time AI saves?

About a third of workers who used AI at work in the last week told the US Census Bureau it had saved them one to two hours. The survey that produced that figure stops there. It asks how much time AI recovered and never asks what happened to it next. Two studies that did try to follow the hour, one tracking 25,000 Danish workers for two years and one spending eight months inside a single technology company, arrive at the same shape of answer from opposite directions: the hour gets reorganised into new work or absorbed into a longer day. Neither found anyone treating where it goes as a decision.

What the Census figure measures, and the question it does not ask#

The US Census Bureau's Household Trends and Outlook Pulse Survey, fielded in March 2026, found that 56 per cent of workers had used AI for at least one of eleven listed job tasks. Among that group, 31 per cent said it saved them one to two hours, 25 per cent said less than an hour, 15 per cent said three to four hours and a further 15 per cent said more than four. Ten per cent reported no time saved at all, and 4 per cent said AI cost them additional time. The question behind the figures is how many extra hours a worker estimates they would have needed without the tool, which is a real and useful number. It is not a measurement of what happened to the hours the tool freed up, and the survey does not attempt one.

Reorganised, not banked#

Humlum and Vestergaard linked adoption surveys to administrative labour records for roughly 25,000 workers across 7,000 Danish workplaces in eleven occupations exposed to generative AI, and followed them for two years after ChatGPT's release. Graded entry. Earnings and hours worked came back essentially unchanged, precise enough to rule out an effect larger than about 2 per cent in either direction. If the time AI saved were simply being banked as shorter days or fatter pay, a change of that size across that many workers over two years is close to the smallest thing this kind of data could show. It did not appear.

What did appear, in the same data, was substantial task reorganisation: new work overseeing and integrating what the AI produced, replacing some of what the job used to consist of. The hours stayed the same. What filled them changed underneath.

Absorbed into a longer day#

Ye and Ranganathan, at UC Berkeley's Haas School, spent eight months inside a 200-person US technology company: real-time observation, attendance at meetings, and more than forty semi-structured interviews across functional groups, published through Harvard Business Review in February 2026. Graded entry. Rather than reducing work, they found generative AI intensifying it. 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 to. Three mechanisms did the work: job scope quietly expanded, tasks bled into lunch and evenings, and people ran AI processes in parallel while doing something else, so the tool's output competed for attention rather than freeing it.

This is a single firm, studied qualitatively, and the authors describe the work as in progress: no percentages, no control group, no claim to generalise beyond what forty interviews and eight months of observation can support. What it adds to the Danish result is a mechanism at the level of a single person's day for why the aggregate hours never move. Time saved on one task is not returned. It is spent starting the next one sooner.

The option neither study found#

Put the two together and a task's freed time goes to new oversight duties, to a longer working day, or both. What is absent from both studies, and from every other source read for this page, is anyone treating the freed repetition as something to keep on purpose. That connects directly to the argument this site makes elsewhere about capability debt and the missed reps: the minutes a tool removes from a task are frequently the minutes that used to build or maintain the judgement to do that task without it. A default that reabsorbs those minutes into throughput or into longer hours is the missed reps mechanism again, now arriving through the calendar rather than through the org chart. This reading is this page's own, not a finding of either study, and neither Humlum and Vestergaard nor Ye and Ranganathan test it.

Making the destination a decision#

• Name where the time is currently going before assuming it was saved. Ask whether a freed hour became new oversight work, a longer day, or genuinely returned time, because the two studied defaults are both easy to mistake for the third.
• Track hours and scope, not only output. The Danish null result on hours was only visible because hours were measured directly; a team watching output alone would have seen productivity rise and missed the reorganisation happening underneath it.
• Where the removed minutes were the practice, decide on purpose whether to keep some of them. Not every task's repetitions build a skill worth protecting. Where they do, this is the same design choice as desirable difficulty and deliberate practice: friction kept because of what it produces, not left in by accident.
• Expect the default to reassert itself. Neither study found a case of freed time being deliberately redirected. Absent a specific decision, throughput and hours creep appear to be what happens on their own.

What this evidence does not establish#

• Neither study measures capability. The Danish data covers earnings, hours and task content; the Berkeley ethnography covers pace, scope and hours worked. Nothing here measures whether the people involved got better or worse at the underlying work, which is a claim the illusion of competence already gives reason to distrust from self-report alone.
• The Danish result is Danish, and occupation-specific. Eleven exposed occupations in one country's labour-market institutions is not a claim about every job everywhere.
• The Berkeley finding is one firm, in progress, with no published figures. It establishes a mechanism worth taking seriously, not a rate.
• The Census figures are self-reported estimates of a counterfactual. A worker's guess at how much longer a task would have taken without AI is not an audited measurement, and the survey was not designed to check it against anything.
• Nobody has tested the alternative this page argues for. Whether deliberately preserving some of the removed repetition actually protects judgement better than the two observed defaults is this research's own position, not a measured result.

Key sources

• U.S. Census Bureau (2026). AI Use at Work. Household Trends and Outlook Pulse Survey, March 2026 fielding.
• Humlum, A. and Vestergaard, E. (2025, revised 2026). Still Waters, Rapid Currents: Early Labor Market Transformation under Generative AI. NBER Working Paper 33777. Graded entry.
• Ye, X. M. and Ranganathan, A. (2026). AI Doesn't Reduce Work, It Intensifies It. Harvard Business Review, 9 February 2026. Research described by UC Berkeley Haas as in progress. Graded entry.

Related SuperSkills research#

On what removing repetition costs, capability debt and the missed reps. On keeping friction by design, desirable difficulty and deliberate practice. On whether output growth means anything, does AI actually make people more productive. On who ends up doing the new oversight work, who owns verification and who supervises work they cannot do.

About this research#

Rahim Hirji is the author of SuperSkills (Kogan Page, 2026), keynote speaker on AI and human capability, and founder of The SuperSkills Intelligence Company. The Census, Humlum-Vestergaard and Ye-Ranganathan figures were all read at source. The reading that connects them to capability debt and the missed reps is this research's own interpretation, kept separate from what the three sources themselves measure.

How this research works  ·  Reviewed quarterly  ·  Found an error? Tell me and it is corrected on the page.