In the largest randomized field experiment on an AI assistant built into everyday office software, workers given Microsoft 365 Copilot spent about 1.4 fewer hours per week in Outlook, a 12 percent drop, and workers who actually used it saved roughly two hours. The study covered 7,137 knowledge workers at 66 large firms. It found no change in how many emails they answered, how many meetings they attended, or how many Word documents they finished.
Key takeaways
- Time spent in Outlook fell 1.4 hours per week for workers offered the assistant (12 percent) and about two hours (17 percent) for those who used it.
- Weekly use settled just under 40 percent of workers offered the tool, and 20 percent never touched it.
- Workers did not answer more email threads, attend more meetings, or finish more documents, so the saved time did not turn into measurable extra output.
- Three of the four authors worked at Microsoft Research, the study measured activity logs rather than work quality, and the sample was large early-adopter firms.
How the experiment was built
Between September 2023 and October 2024, Microsoft offered a limited number of Copilot licenses to selected large companies. Each participating firm recruited at least 100 workers, and the licenses went to a random subset: 3,684 treated workers against 3,453 controls. Controls kept working as before. Treated workers could use the assistant or ignore it, and firms agreed to hold the assignment for six months.
The researchers did not read anyone's messages or documents. Their data came from usage logs inside Microsoft's own applications: minutes in Outlook, number of emails read, meetings attended, documents written. That choice protects privacy, and it also means the study has no measure of whether the work product got better or worse. The NBER working paper says so directly and notes that it has not been through peer review.
Who used the assistant, and how often
More than 90 percent of treated workers tried Copilot at least once. Use peaked at 55 percent of treated workers in the first weeks, then settled to just under 40 percent per week after about 12 weeks. Five percent used it every week, and 20 percent never used it in months four to six. Teams meeting summaries drew the most use, followed by Outlook, then Word.
Which company a worker belonged to mattered more than what the worker did all day. Average weekly use among treated workers ranged from 6.3 percent at one firm to 70 percent at another. By industry, telecommunications and technology firms averaged 45 percent weekly use, against 25 percent at construction and manufacturing firms. A worker's own pre-experiment habits explained only 9.2 percent of the variation in use, and firm identity explained nearly twice as much. The authors read this as a sign that training and management practices shape whether a tool gets used.
What changed in the workday
Workers offered Copilot spent 1.4 fewer hours per week on email in Outlook, 12 percent below their earlier average. Scaling the estimate to people whose use was actually caused by receiving a license gives about two hours, or 17 percent. Users also batched their email: the paper reports about 1.5 additional hours per week of stretches without Outlook activity. Time in applications outside regular working hours fell by roughly 0.25 hours per week, 9 percent, which suggests part of the saved time went back to the workers themselves.
No significant average change appeared in total time in video meetings or in time spent writing documents.
What did not change
Treated and control workers replied to the same number of email threads, joined the same number of Teams meetings, and completed the same number of Word documents. The authors found no sign that workers took on new kinds of tasks, and they saw no negative effect on close coworkers' work patterns. They describe the result as the same work done faster in one area, with no reorganization of jobs. Because the study could not see work quality, it cannot say whether faster email handling helped or hurt the quality of replies.
The results at a glance
| Measure | Result reported in the paper |
|---|---|
| Sample | 7,137 workers, 66 firms, six-month assignment |
| Outlook time, all workers offered the tool | 1.4 hours per week less (12 percent) |
| Outlook time, scaled to actual users | About two hours per week less (17 percent) |
| Time outside regular hours | About 0.25 hours per week less (9 percent) |
| Meetings, documents, email threads answered | No significant difference from controls |
| Work quality | Not measured |
A way to test an assistant on personal work (illustrative)
The study design can be borrowed in miniature. This checklist is an illustrative example, not a method from the paper.
- Pick one recurring task, such as clearing an inbox or writing a weekly status note.
- Time it for two weeks without the assistant, using a simple log.
- Time the same task for two weeks with the assistant, including the minutes spent checking and correcting its output.
- Compare totals, and read a sample of finished items from both periods for errors.
- Repeat after a month, since use in the experiment peaked in the first five weeks and then fell, possibly because of novelty.
Limits of the evidence
The firms were very large and agreed to join with only modest incentives, so they may be more eager adopters than typical employers. Licenses were scarce at the time: each firm could buy 300, which means workers were a small slice of their organizations. The tool tested was the 2023 and 2024 version of one product. Three authors were Microsoft Research employees, and the acknowledgments thank Microsoft's customer research program, which is context for readers weighing the findings. This page does not rank products or recommend buying one.
FAQ
Did the study show that AI assistants raise productivity?
It showed time savings on email for users and no measurable change in output counts. The researchers state that they had no productivity or performance measure, so the question of output quality stays open.
Does the two-hour figure apply to any worker?
No. It applies to workers whose use was caused by receiving a license, in large firms, in one application. The smaller 1.4-hour figure describes everyone offered the tool, including the 20 percent who never used it.
Why did usage differ so much between companies?
The paper cannot say for sure. Firm identity explained far more of the variation than industry or personal habits, and the authors point to training and managerial practices that their logs could not capture.
Where did the saved time go?
Part of it went to fewer hours worked outside regular times, about 0.25 hours per week. The logs show no extra emails answered or documents finished, but they cannot see other work done outside Microsoft's applications.
Related reading
- Official US and Canadian productivity statistics and what they show about AI
- Bureau of Labor Statistics productivity data by US industry
- Three cost series for building and running AI systems
- The timeline of Baidu's Ernie Bot, another assistant product
More in AI Tools and Industry.
Last verified: October 2026
Sources
- Shifting Work Patterns with Generative AI, Dillon E, Jaffe S, Immorlica N, Stanton C, NBER Working Paper 33795, May 2025, revised November 2025