ROBERT KARASZEWSKI

Human-Led AI Strategy

Where Does the Time AI Saves Actually Go?

Written by

Robert Karaszewski

For most people, the time AI saves goes into other work, not into rest. In a study of 25,000 Danish workers, 85% of AI chatbot users said they moved saved time to other job tasks, and fewer than 10% took more breaks or leisure. Yet their earnings and recorded hours did not change, which raises a harder question: who decides what that time becomes?

Ask a team how much time AI saved last quarter and you will get a number. Ask where it went and you will usually get a pause. Part of the answer is known: checking and fixing AI output takes its share first. How much of the saved time survives rework is a separate question. This article follows what is left.

Where does the time AI saves actually go?

Mostly into other work. Anders Humlum and Emilie Vestergaard surveyed 25,000 workers at 7,000 Danish workplaces in 11 AI-exposed occupations and linked the answers to administrative records through December 2024. Of those who used chatbots, 85% reported moving saved time to other job tasks, about 30% spent more time on the same tasks, and fewer than 10% took more breaks or leisure. The answers overlap.

These are shares of people, not hours, so they do not show how a saved hour was divided. They are also self-reports, and the authors note that workers may overstate what the tools do.

How much time is there to trace?

Less than the headlines suggest. Humlum and Vestergaard put reported savings, averaged across users and all work hours, at a few percent of work time.

The strongest behavioral evidence is narrower. A randomized field experiment gave Microsoft 365 Copilot to workers across 66 large firms, with 7,137 participants tracked through application data rather than surveys. Those who used it spent about two fewer hours a week on email, from a base of about 12. Three of the four authors worked at Microsoft, and the data come from the tool’s early rollout.

The study does not measure productivity. Its authors also concede that much of the saved time appears to have shifted to activities they cannot observe.

Why doesn’t the saved time show up in results?

Probably because savings land where one person can act alone, while results depend on work several people must change together. That is the Copilot authors’ reading. Compared with the control group, treated workers replied to the same number of email threads, attended the same number of meetings, and completed the same number of documents. An inbox is one person’s to reorganize. Who owns a document, or what a meeting is for, takes agreement.

The rollout was individual by design, so a team-wide rollout could behave differently.

Results further up look just as flat. The Danish study rules out effects on earnings or hours above 2%, for workers and for workplaces. In a survey of nearly 6,000 executives in the U.S., U.K., Germany, and Australia, 89% reported no AI effect on labor productivity over three years, measured as sales per employee. Both are working papers, and the executive figure is a self-assessment.

Connecting these findings is my inference, not a result of any one study: time redirected one person at a time gets absorbed into the job as it already exists.

When does saved time turn into more work?

When nobody decides otherwise and the tool makes starting easy. In the Danish data, about 8% of users took on entirely new tasks even where employers did nothing, and that share roughly doubles where employers support adoption. About a third of those new AI tasks involve overseeing AI and about a quarter involve integrating it. Workers reporting the largest savings were also the most likely to have taken on new tasks.

An eight-month ethnography by Aruna Ranganathan and Xingqi Maggie Ye, at a U.S. technology company of about 200 people, found employees widening their scope, letting work seep into pauses and evenings, and running several threads at once, often unasked. It is ongoing research at one firm.

The evidence conflicts on one point. In the Copilot trial, treated workers spent about a quarter of an hour less per week working outside regular hours. Scarce licenses in large firms versus open access at one technology company may explain the gap, but no study here tests that.

How do you trace your team’s saved hour?

Pick one workflow and account for each saved hour across five destinations before reporting it:

  • Lost to checks and fixes. Who reviewed or corrected the output, and was their time counted? Subtract this first.
  • Moved to other existing work. Which task got more time, and did its output change? If not, the hour was absorbed.
  • More of the same task. Did volume or quality rise, and did anyone need it?
  • New work created by AI. Who now reviews, integrates, or sets rules? Give this work an owner.
  • Returned to the person. Less overtime or real focus time is a legitimate choice. Record it as that, not as output.

Then separate who can act. An employee can move their own hour. The meeting, handoff, or approval it feeds belongs to whoever owns the workflow. Treat this as a working hypothesis: redesign at that level is where a saved hour can become a result. That missing record feeds what I call the Missing Down-Slope, which explains why reported AI gains erode once the hours are booked.

The pressure point shifts by market. U.S. firms are often tempted to convert hours into headcount before tracing them. In the Gulf, where AI programs often launch quickly at senior level, the ledger works best sponsored at that level. In Central and Eastern Europe, where teams often run lean, backlog is likely to absorb the time first.

Frequently asked questions

Should employees keep the time AI saves?

They can, if that is a decision. The Copilot data suggest some time already returns to workers. A company that wants output instead must say what the time is for.

Can saved time justify cutting headcount?

Not on hours alone. Danish workplaces that encouraged chatbot use showed no differential change in employment or wage bills. Decide workflow by workflow, after output has moved.

Does giving the whole team access solve it?

Not by itself, on current evidence. Danish workplaces combining encouragement, tools, and training reported the highest use and benefits, but no detectable change in earnings or employment. In the Copilot trial, having a treated close colleague made no detectable difference, though few teams had high coverage.

Sources and further reading