Does generative AI shift knowledge workers away from communication?

The paper argues that adopting generative AI changes what knowledge workers do, not only how much they produce. Using digital trace data from the Microsoft M365 suite at multiple large international companies, difference-in-differences analyses find that users who used the AI system more than 100 times over a 20-week post-adoption period increased productivity application actions by 21.2 percent (content creation in Word is the named example) and communication application actions by 7.1 percent. The authors summarize the pattern as "a shift from communication and coordination toward individual, documentation-focused work."

The reasoning rests on the size of the gap. Both categories rose, but the productivity increase is "much larger than for communication applications, indicating a change in the type of work performed." That gap is what lets the authors say AI adoption is "not merely a tool that impacts productivity" but "may fundamentally change how people work by reshaping their work habits." So the shift is in the balance between activities, and the reported figures do not show communication falling in absolute terms. The paper's framing sits between two positions its introduction sets up, that AI automates and displaces work or that it augments human capability, because it describes a change in the mix of work rather than a change in its total.

Against the nearest notes, this is a field-scale counterpart to Does AI really save time, or just change how we spend it?, which found reallocation inside a single task in a skill-formation study. Both say that a single volume measure hides a change in activity type, but they measure different things: that note tracks time moving toward prompting and reading AI output, while this paper tracks actions in productivity versus communication applications. The excerpt does not say whether interaction with the AI tool is itself counted in either category. It also leaves When does AI actually boost worker productivity? untouched, because here "productivity" means a count of application actions, not output quality or skill. And Does concentrated AI exposure enable workers to adapt and reallocate? describes reallocation across tasks at the level of firms and employment, whereas this paper observes an activity mix within individual users.

The excerpt is silent on much that would be needed to read the shift as good or bad. It gives no number of companies or users, no account of how comparison users were selected, and no statement on whether heavy users differed from other users before adoption, which matters because the 100-use threshold defines the group. It does not say whether more actions mean more or better output, and it says only that "further analyses suggest potential mechanisms" without naming them. What follows at the strength the evidence allows is narrow: among heavy users, AI adoption coincides with a rebalancing toward individual documentation work, so a productivity measure that counts only output activity would miss a change in how much coordination the same people do.