Does Copilot change what knowledge workers actually do?

Microsoft researchers ran a field experiment between September 2023 and October 2024 with 66 large firms and 7,137 knowledge workers (3,684 treated), randomly assigning Microsoft 365 Copilot licenses within each firm's pool of recruited participants and holding the assignment for six months. The excerpt reports that "in the second half of the 6-month experiment, the 80% of treated workers who used this tool spent two fewer hours on email each week and reduced their time working outside of regular hours." Beyond that, the authors state plainly: "we do not detect shifts in the quantity or composition of workers' tasks resulting from individual-level AI provision." The measures come from Microsoft's own telemetry on time use and activity counts (emails read, meetings attended, documents authored); the authors note they "do not observe the content of any work nor any measures of productivity or performance evaluation."

The paper's own explanation for why email moved and little else did turns on who controls the task. "Email is solitary and each worker is free to develop her own process for inbox management. In contrast, shifting what work gets done in meetings, how long they last, or who takes primary responsibility for writing documents requires coordinating with colleagues and agreeing on new norms." Consistent with this, the authors find no spillover of a worker's Copilot access onto connected coworkers' Teams or Word activity, and only an imprecise, not-significant suggestion that treated workers with treated close colleagues saved somewhat more time on email. The authors read this as a signature of early adoption: "the complementary process innovations and organizational restructuring that drive larger changes in workflow had not yet taken place."

This null result on task composition sits in direct tension with Does generative AI shift knowledge workers away from communication?, which used M365 trace data to find heavy Copilot users shifting toward productivity (documentation) actions relative to communication actions. The two studies may not actually disagree: that paper compares workers who chose to use the tool 100+ times against those who did not, a self-selected-intensity design, while this is a randomized intent-to-treat comparison that includes treated workers regardless of how much they used Copilot. It is also a sharper contrast with Does generative AI actually save workers time or intensify it?: where that ethnography describes AI lowering the cost of starting one more task and so expanding scope and hours, this RCT finds freed time on a solitary task (email) with no detected growth elsewhere — though the RCT cannot say whether that freed time was quietly absorbed into other, unmeasured work.

The excerpt does not establish what happened to the two hours a week once freed — redeployed to other tasks, idle, or folded into unmeasured activity. It studies a pilot period in which each treated worker typically held the only Copilot license on their team, so the design itself rules out the coordinated, team-level tool adoption the authors say is needed for larger effects; the null finding on task composition may describe this early, uncoordinated rollout rather than what individual-level AI access does in general. Participating firms also opted in for modest incentives and "may be more proactive in adopting new technologies than the average firm," so even this muted result may not generalize to a typical employer. The defensible implication is narrow: absent coordinated process redesign, individual AI access produces a measurable but contained time saving in solitary work and no detected change in the shape of a job — a different and more conservative finding than studies of self-selected heavy use, and the gap between them may trace to how "AI use" is defined as much as to what AI does.