Shifting Work Patterns with Generative AI

We present evidence from a field experiment across 66 firms and 7,137 knowledge workers. Workers were randomly selected to access a generative AI tool integrated into applications they already used at work for email, meetings, and writing. 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. 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.

Introduction. Generative AI has opened new possibilities for technology to assist with or automate a variety of tasks. Early studies have already shown that generative AI increases worker productivity in targeted tasks (e.g. Brynjolfsson, Li and Raymond, 2025; Peng et al., 2023; Noy and Zhang, 2023; Dell’Acqua et al., 2023), and advances in this technology have the potential to shift the nature of jobs (Eloundou et al., 2024; Tomlinson et al., 2025) and perhaps the macroeconomy (Autor, 2024; Acemoglu, 2025).

In lab experiments and early field studies on generative AI, participants often focused on completing a single task with targeted, well-suited tools. In everyday settings, workers must first learn which parts of their job benefit from the use of new tools and then change their work habits. This dynamic helps to explain why the transition from the creation of new technologies to effective, widespread adoption in the workplace often takes time (Brynjolfsson, Rock and Syverson, 2021).

In this paper, we present field evidence on the adoption of generative AI tools and their effects on the way knowledge workers do their jobs. Between September 2023 and October 2024, Microsoft partnered with 66 large firms to run a cross-industry field experiment to measure how access to an integrated generative AI tool, Microsoft 365 Copilot (hereafter, Copilot), changed work patterns. Each firm recruited at least 100 workers for random assignment to treatment and control groups. Treated workers received access to this new generative AI tool while controls continued working with their current technologies. Participants in the study came from varied occupations, but all did work that used the Microsoft Office tools into which Copilot was integrated (e.g., emailing, video meetings, chat, ∗Corresponding author: eldillon@microsoft.com. † denotes equal contribution. We thank the Microsoft Customer Research Program, especially Alexia Cambon, Sida Peng, Neha Shah, Chris Gideon, Modern Work Marketing, the Office of Applied Research, and company partners for help carrying out this experiment, Abigail Atchison, Roman Basko, and Fabio Vera for superb data science support, Jack Cenatempo, Esther Plotnick, and Will Wang for excellent research assistance. We thank 3 anonymous referees, Tatyana Deryugina, Seema Jayachandran, Rem Koning, Anders Humlum, and Danielle Li for helpful comments. document creation). Firms were asked to maintain the random assignment for six months, giving treated workers time to explore the new tool and integrate it into their daily work. The study took place during the pilot rollout of Copilot, when firms could purchase licenses for only a small fraction of their workers. Treated workers were free to use or not use the tool during their regular work. Although treated and control workers may have used other generative AI tools (for example, ChatGPT had been available for nearly a year by the beginning of the study), Copilot integrated AI into applications the workers already used regularly in their jobs. This integrated tool allowed new uses, like summarizing email inboxes or searching for local information, and likely reduced the effort to deploy generative AI for writing documents and emails.

Related work. Our work adds to studies on the effects of generative AI in real workplaces (Brynjolfsson, Li and Raymond, 2025; Cui et al., 2024; Otis et al., 2023), while complementing studies that use controlled tasks in a lab setting.3 We also contribute to the broader literature on technology diffusion (Comin and Hobijn, 2010; Forman, Goldfarb and Greenstein, 2018). Surveys indicate rapid adoption of generative AI at work across a broader set of workers than those considered in lab or early field settings. Humlum and Vestergaard (2025a) find 41% of Danish workers in exposed occupations used ChatGPT at work by the end of 2023 and Blandin, Bick and Deming (2024) find 24% of all U.S. workers used generative AI at work by August 2024. In follow-up work, Humlum and Vestergaard (2025b) find that AI chatbots at work have not had measurable impacts on worker earnings or hours worked, largely consistent with our mixed results on worker behavior. This study bridges some of the gaps between these two lines of research: measuring both take up and impacts of use for a generative AI tool on a broad set of workers, albeit at the cost of clear measures of worker productivity.

Method. We use telemetry data collected by Microsoft products to track workers’ behavior during the experiment and for several months beforehand. Our primary outcome measures are related to time use (e.g. time spent in meetings or reading emails) and quantities of activities (e.g., emails read, meetings attended, documents authored) within the Microsoft Office suite, as we do not observe the content of any work nor any measures of productivity or performance evaluation.

1.1 Experimental Design In fall 2023, Microsoft offered select large firms the option to purchase exactly 300 Copilot licenses.

Copilot was not available to smaller firms or to individual consumers, and firms could not purchase additional licenses beyond 300. Within this group, Microsoft’s marketing team recruited firms to join the study with the understanding that they would allocate at least 50 of these 300 licenses to the experiment and recruit at least twice as many participating workers. The Copilot licenses were assigned at random among experiment-eligible workers, but treated workers were not required to use the tools or deploy them for specific work tasks. In exchange for participating in the study, firms received early access to findings from the experiment, including customized firm-level reports. They did not receive discounts or access to additional licenses.

Appendix A includes more details on the 7,137 participating workers (3,684 treated) across 66 firms. The firms span a range of industries (see Figure A.1a). All are very large and most operate in multiple countries. Firms opted into the study in response to only modest incentives and may be more proactive in adopting new technologies than the average firm.

Firms varied in how they recruited participating workers, with some targeting specific functions or workers with a track record of adopting new technologies early and others creating an open call for participation.4 Appendix Table B.1 tabulates occupations for the subset of workers for whom we have that information. As expected, the sample is dominated by knowledge workers who regularly do the kinds of emailing and word processing tasks for which Copilot was designed. Appendix Table B.2 summarizes pre-period work behaviors for treated and control workers in the sample. Unfortunately,

Discussion. In keeping with the largely null direct effects of Copilot for meetings or writing tasks, there are no spillover effects of Copilot access on connected coworkers’ Teams or Word activity. These consistent differences across applications may reflect how easily individual workers can adjust their work to take advantage of the new tool. 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. This pattern is fitting for the early adoption stage we study, where individual workers were exploring the capacities of a new tool, but the complementary process innovations and organizational restructuring that drive larger changes in workflow had not yet taken place (Bresnahan et al., 1996; Brynjolfsson and Hitt, 2000).

In the spirit of the literature on complementary innovation and process restructuring, we test whether there are different effects when multiple coworkers get access to the tool. The third column of Table 3 shows the triply interacted effects of Zi, ZClose, and 1{τft ≥0}. The total effect of being treated oneself while having at least one treated close colleague is given by the sum across the three columns. The point estimates suggest some modest benefits to gaining access to the new tool along with colleagues, treated workers with treated colleagues save 50% more time on email, but again these effects are imprecisely estimated and, with the sharpened q values, we cannot rule out zero differences in effect.18 These results suggest that the co-creation of work and process restructuring that the literature identifies as a pre-condition for effective technology adoption did not seem to take place under the individual-level rollout of the tool that our experiment assesses.

Conclusion. This paper offers a portrait of the first steps of adoption of a new technology. Historically, the main productivity gains from innovative technologies came only with the complementary process innovations that shift the nature of work around these new capacities. These institutional changes take time. through the experiment) Copilot licenses. Under our data use agreements, we are unable to track this Copilot access for workers outside the study. 18The specification here is a 0-1 indicator for whether any close coworker is treated, while a saturation design would account for the share of workers on a team that are treated. Using that approach, reported in Appendix Table D.5, we also detect little difference in treatment effects by whether our studied workers had an above or below median share (11.9%) of close coworkers with access to Copilot. However, even our “high saturation” teams had, on average, only 39% of coworkers with access to Copilot, so we may simply be unable to detect team-level effects from this experimental design.

Among recent technological innovations, generative AI is unique because it does not require new equipment or hardware to use, allowing individual workers to bring generative AI to work even without their employers’ support. As Blandin, Bick and Deming (2024) point out, the discrepancy between the high rates of generative AI use reported in surveys of workers, such as theirs, and the lower rates reported in surveys of firms, as in Bonney et al. (2024), suggests that workers use these tools even before their employers develop formal adoption strategies.

This study provides indications of how this individual use of generative AI shapes the patterns of work for knowledge workers in large firms. Large firms are often among the earliest adopters of digital technologies, presumably because of their ability to capture benefits (McElheran et al., 2024).

We study workers during the first year in which generative AI tools gained broad use. While their colleagues may have been using other generative AI tools, the workers we studied were often the only members of their team with access to Copilot during the early roll-out period.

Limitations. They document workers’ email and meeting behavior in a sample of 102 large firms around 2019. Our sample of workers seems roughly comparable to theirs. For example, their studied workers attended a mean of 39 meetings per month while ours attended 43. Nonetheless, we caveat that participating workers may not be representative of the overall worker population at participating firms, or of knowledge workers more broadly.