Skip to content
← Insights
Nov 14, 20264 min read

Who gets the time an AI tool saves?

A faster task leaves a policy decision about the time recovered. Workplace studies separate individual tool effects from the organisational changes needed for shorter hours.

A desk clock beside a closed laptop and blank schedule, surrounded by an hourglass, document tray, and small pause button.

Finishing a task sooner creates a decision about the time left over. A shorter working week happens only when the organisation changes what it expects people to do with that time.

That is the unresolved part of the promise that AI will end busywork. A faster draft may reduce one person's effort. The same tool can also make it cheap to ask for more drafts, leaving colleagues with more material to read. The result depends on how work is assigned after the tool arrives.

Separate task time from the working week

The field experiment described in Shifting Work Patterns with Generative AI studied access to an integrated assistant in workplaces. It found changes in work that individuals could alter, including less time spent on email, while meeting time did not change significantly. That distinction limits the idea that giving someone a tool automatically reorganises shared work.

A meeting exists in several calendars. One person's faster preparation does not cancel everyone else's appointment. The decision about whether the meeting is still needed belongs to the people who set the coordination process.

I would therefore begin a pilot with a named activity and the result it must deliver. If the activity is preparing a weekly status report, the question is what decision the report supports. A report nobody uses can be removed. A report that supplies information for a real decision still needs an accepted standard, whether its first draft is written manually or generated.

The measurement should include the next person in the process. A writer who produces a longer report more quickly may create extra reading work for a manager. The local saving and the total saving can differ.

Describe what happens to recovered time

A team can use recovered time for more output, a backlog, or a shorter schedule. Those are different policies. Calling each one productivity improvement conceals the question employees actually face: what will be expected after the tool is introduced?

A 2026 study on generative AI and time reallocation examines work and leisure using a survey of Korean workers. The authors found little relationship between reported time savings and changes in output. Its design studies how adoption relates to the allocation of time. A survey relationship needs to remain distinct from an experimental estimate of how a specific workplace policy changes hours.

For a team-level experiment, I would record the intended use of any saving before starting. If the goal is an earlier finish, the evaluation needs actual working hours. If the goal is more accepted output, the evaluation needs output and its correction cost. Counting tool use cannot substitute for either outcome.

The schedule must also include the work that remains after a generated result. Reviewing a draft can require attention that the generation time does not capture. A fast response from the model may be followed by a slow factual check.

A shorter-week policy has its own evidence

The 2025 study Work time reduction via a 4-day workweek examined employees in organisations adopting shorter hours without a pay reduction. It reported improvements in well-being compared with the pattern in control companies. The study followed an organisational intervention; it was not a test that AI alone creates a shorter week.

That distinction gives the discussion a more useful starting point. An employer considering reduced hours needs to design a schedule and examine the work required to support it. AI may be one tool used during that redesign. Its presence does not remove the coordination decisions.

A customer-facing service also needs a coverage plan. Shorter individual hours can coexist with a longer service window if shifts are arranged accordingly. Whether that arrangement works depends on the team and the service requirements. A technology claim cannot choose the staffing arrangement for them.

Make the acceptance rule visible

The end of an activity needs a definition. A completed report has the required information and has reached the person using it. A completed software change has met its agreed checks and review requirements. A generated file is an intermediate result when further decisions remain.

I would use that accepted endpoint when comparing the old process with the new one. Record exceptions instead of measuring only the smooth cases. If a generated result occasionally creates extensive rework, that cost belongs in the evaluation period.

The same policy should prevent the recovered time from disappearing into unplanned demands. A team trying to reduce hours needs a way to decline additional work that would undo the schedule change. A team trying to increase output needs a way to keep the volume within the review capacity available.

The original appeal of removing busywork is reasonable. Time spent doing a task that no longer contributes to a decision can be returned to something else. Turning that possibility into a shorter week requires an explicit agreement about which work ends and what people are allowed to stop doing when it is finished.

AIWork