Who gets the time we save?
If a task takes an hour less, what happens to that hour? It might become another assignment, time to learn, a conversation or a genuine pause. Independent workers, employees and shift workers have different control over that choice. Count the whole task, including instructions, checking and corrections. Several agents running at once may finish sooner while still occupying someone’s attention. Time saved becomes meaningful when a person can decide how to use it.
The evidence points in different directions. A 2026 research paper, not yet peer-reviewed, surveyed 5,512 Korean workers and reported time savings and more leisure during the working day among recent generative-AI users. Berkeley researchers, observing one technology company over eight months, found broader responsibilities and work spilling into pauses. The methods differ, so neither tells us what will happen everywhere. They show why the use of saved time deserves as much attention as the speed of the tool. Suh & Oh · preprint, 2026 Ye & Ranganathan · Berkeley Haas, 2026
Begin with a clear sense of when the work is finished, including checking and revisions. Then ask what any savings could make room for: an unhurried lunch, learning, care or a reliable stopping point. Some boundaries can be set personally; others need agreement with clients, collaborators or employers. The test is whether people gain usable time without hardship or pressure passed to someone else.
equilibri asks: how can time saved become time people are free to use?
Research & its limits / time
Where researchers found time savings
A randomized experiment across 66 firms and 7,137 knowledge workers tested access to AI in familiar workplace applications. In the second half of the experiment, workers given access who used the tool spent about two fewer hours a week on email and worked less outside regular hours. Researchers found no detectable change in the amount or mix of tasks. The study supports savings on particular activities; it does not show a general shift to leisure or test coworking. Dillon et al. · American Economic Review: Insights, forthcoming
More leisure during work? Early findings
In Suh and Oh’s survey, recent users reported 3.8% less active working time on average. The estimated reduction across the workforce, including non-users, was 1.4%. Users’ reported share of leisure during work rose by 1.3 percentage points. These are survey estimates from Korea, not recorded time logs or results from a randomized experiment. Experiences of fulfillment also varied: less active work does not necessarily mean a better job. Suh & Oh · methods and results, 2026
When faster tools lead to more work
Berkeley’s ongoing workplace study describes expanding tasks, fewer clear stopping points and overlapping AI workflows. Separately, an NBER working paper links occupations more exposed to AI with longer workdays and less leisure in U.S. time diaries. A job’s estimated exposure is not a record of someone using an agent, and an association does not prove cause. These findings make workload agreements an important part of testing AI’s benefits. Jiang et al. · NBER working paper, 2025
Why saved time can be hard to measure
In February 2026, METR judged its follow-up developer experiment unreliable as an estimate of current speed gains. Who took part and how to count time across simultaneous agents both complicated the results. The lesson is to compare complete workflows, separate human effort from machine running time and explain which tasks were tested. A reported speedup alone cannot tell us how much leisure becomes available. METR · methods update, 2026
Machine running time is not always free time
Anthropic’s February 2026 analysis found a median Claude Code turn—a stretch of activity before handing control back—of roughly 45 seconds, while the longest-running end of the distribution exceeded 45 minutes. Some tasks therefore offer long unattended intervals, but that is not typical of every task. A machine running independently does not, by itself, establish that a person has saved time or can take a break. Anthropic · observational product research, 2026
Shorter weeks by agreement
A six-month four-day-week study followed 2,896 employees in 141 organizations, with 12 comparison companies. Shorter hours with maintained pay were associated with better reported well-being. Organizations volunteered to participate, so the study was not randomized. It examines a deliberate change to working arrangements, rather than AI adoption. Fan et al. · Nature Human Behaviour, 2025
Speed still needs a quality check
A review combining 106 experiments found that people working with AI performed better than people alone on average, but worse than whichever performed best alone: the person or the AI. Results varied by task. Any claim of saved time therefore needs to include checking, correction and coordination while keeping the required quality of work consistent. Vaccaro et al. · Nature Human Behaviour, 2024
