At a glance
- Source
- Dillon, Jaffe, Immorlica and Stanton (2025). Shifting Work Patterns with Generative AI. NBER Working Paper 33795; arXiv 2504.11436, version 4 (November 2025). doi:10.3386/w33795 · Not peer-reviewed: other researchers have not yet checked it for a journal.
- How it was done
- Within each firm, workers were assigned by chance to get Microsoft 365 Copilot or to carry on without it, for six months. The outcomes come from the software’s own activity records, not from what workers said. Randomised experiment. People were assigned by chance to work with or without the tool, and the two groups were compared.
- Who took part
- 7,137 knowledge workers at 66 large firms, 3,684 of them given the tool.
- Where
- Large firms in several industries, most operating in more than one country, from September 2023 to October 2024.
“Apart from these individual time savings, we do not detect shifts in the quantity or composition of workers’ tasks resulting from individual-level AI provision.”
What it found
The tool was built into the email, meeting, chat and writing software the workers already used.
- Who used it. 80% of the workers given the tool used it.
- Email. In the second half of the six months, users spent about two hours less on email each week. They read 4% fewer emails and did not take longer over each one. They replied to as many email threads as before, and no more slowly.
- Hours. Users did less work outside regular hours.
- Meetings and documents. Time in meetings and on writing documents did not change.
- The work itself. Apart from the time saved, the study found no change in how much work people did or what kind.
Work one person controls, such as their own inbox, changed. Work that depends on other people, such as meetings, did not.
When the result holds
Each worker got the tool on their own, inside an unchanged way of working. The study tests what handing out a tool does. It does not test what happens when a team redesigns its work around the tool.
Earlier versions of the paper reported a larger email saving. This page uses version 4, of November 2025.
Limits
- Three of the four authors work for Microsoft Research, and the tool is Microsoft’s. Microsoft’s marketing team recruited the firms and ran the experiment.
- The firms were very large and chose to join. They may adopt new technology more readily than most.
- Activity records show time spent in applications, not the quality of the work.
- A working paper, not peer-reviewed.
Practices that use this study
This study supports:
- Measure time saved per task against a baseline Record how long a task takes and how good the result is before you add AI. Without that starting figure you cannot tell a real gain from a hopeful one. Do first · Effort: low (days, one person)
- Start with one costly process, not a company-wide rollout Pick one process that visibly costs time or money. Apply AI there, with a person checking the output, and expand only when the numbers move. Do first · Effort: medium (weeks, a few people)
- Start with ready-made AI in software you already use, before building your own Your office, accounting, support and sales software probably has AI features now. Try those on one task, with a few licences and a measured result, before you pay anyone to build something custom. Do first · Effort: low (days, one person)
- Redesign the workflow instead of adding a tool to it Giving people an AI tool inside an unchanged process changes little. Decide again who does each step, in what order, and where a person checks. Do next · Effort: high (months, several teams)