Does AI actually give you time back?

“Does AI save time?” and “does AI give you time back?” sound like the same question. The research says they aren’t.

The case for yes

The task-level evidence is strong, and most of it isn’t people marking their own homework.

• Writing. In a randomised experiment published in Science, Noy and Zhang (2023) gave 453 professionals writing tasks from their own line of work. Those given ChatGPT finished about 40% faster, and independent graders rated the work higher.
• Customer support. Brynjolfsson, Li and Raymond (Quarterly Journal of Economics, 2025) followed 5,179 support agents as an AI assistant was rolled out. Issues resolved per hour rose 14%, and 34% for the newest agents. The most experienced barely gained.
• Consulting. Harvard and BCG’s 758-consultant study found people were more than a quarter faster on tasks the AI was good at.
• Government. In the UK Department for Business and Trade’s Copilot trial, summarising a report took under 13 minutes with Copilot and about 42 without, with better results.

The case for “not really”

Zoom out from the task to the week and the numbers shrink.

Humlum and Vestergaard (2025) surveyed about 25,000 Danish workers, twice, across 11 AI-exposed jobs, and linked their answers to tax and employment records. Users said chatbots saved them 2.8% of their working hours — about an hour a week. In the US, Bick, Blandin and Deming found 5.4% for users, a bit over two hours. Both are self-estimates.

Then the Danish team asked what people did with the time:

About 80% put it into other work. Fewer than one in ten took it as a break or as leisure.

Earnings and hours didn’t move — “precisely estimated zeros”, in the authors’ words. And 17% of users said AI had created new work for them, such as teachers redesigning assignments around it. About 5% of people who didn’t use it said the same. One person’s shortcut can be another person’s checking.

The data runs to mid-2024. Tools have improved since.

Your sense of speed isn’t reliable

In 2025 the research group METR randomly assigned real tasks to 16 experienced open-source developers, with and without AI. With AI the work took 19% longer. Afterwards the developers believed it had been 20% faster. Screen recordings showed less coding and more prompting, waiting, checking and idle time.

The rerun is just as telling. METR’s 2026 update leans towards a speed-up now, but its confidence ranges include zero, and METR calls the data “an unreliable signal”. More and more developers refused to take part because they didn’t want to work without AI. Wanting a tool tells you that you like it. It doesn’t tell you it saved you time.

Who owns the hour

The strongest objection to the Danish result is that salaried people on fixed hours can’t go home early. The saved hour goes to the employer. Bick and colleagues add that some people may take savings as “on-the-job leisure” that no survey will catch.

The reply comes from Berkeley. Ranganathan and Ye spent eight months inside a ~200-person US tech company. People worked faster, took on wider work, and let it spread into lunch and evenings — “often without being asked to do so”. It felt like momentum at the time. Over the months they felt busier. It’s one company and the full study isn’t out yet.

It has happened before

Keynes predicted in 1930 that his grandchildren would work “three-hour shifts or a fifteen-hour week”. The usual sequel is Ruth Schwartz Cowan’s More Work for Mother: household machines raised standards and saved no time.

The newer data disagrees. Gershuny and Harms (2016), using US rural women’s time diaries, found cooking and cleaning fell from about 271 minutes a day in the 1920s to 108 in the 2000s. The machines did give time back. It went somewhere: childcare rose from about 35 minutes a day to 88, and shopping from 23 to 80.

That’s the pattern. Saved time goes to whatever is standing closest.

What to do with it

• Decide where the hour goes before you save it. Put it in the calendar as something — a walk, the kids, leaving early — or the gap fills itself.
• Time it once. Same kind of task, with and without. Use a clock, not your gut.
• Check where the saving went. If it takes you four minutes to send and someone else twenty to fix, you’ve moved the time, not saved it.
• Know the edges. In the UK trial, spreadsheet analysis was slower and less accurate with Copilot. Harvard’s name for it is the “jagged frontier”.

---

Sources and reporting notes: data from Denmark, the US and the UK; most time savings are self-reported. The only New Zealand figure found was OpenAI’s own survey (1,007 people, 45% saying 4+ hours a week) — a vendor asking about its own product, not used here. Noy & Zhang, Brynjolfsson et al. and the Harvard/BCG study were read via abstracts and summaries. The Berkeley study was read via HBR’s summary and the Haas release. Cowan is summarised via Gershuny & Harms.