Early Microsoft Copilot adopters markedly shift work toward document-focused tasks: users show roughly a 21% rise in productivity app actions but only a 7% rise in communication activity, and send fewer small-group emails — boosting individual output measures while potentially narrowing interpersonal information flow.

Main Finding

Adoption of generative AI (Microsoft Copilot) in large firms is associated with substantial increases in individual productivity-oriented activity and smaller increases (and some decreases in specific tasks) in communication activity. Concretely, users who used Copilot ≥100 times in the 20 weeks after enablement showed a ~21.2% weekly increase in productivity app actions and a ~7.1% weekly increase in communication app actions (DiD estimates, p<0.05). Higher Copilot intensity (100→500 uses) is associated with larger gains in productivity actions (≈21.2% → 27.0%) and also larger—but smaller—gains in communication actions (≈7.1% → 13.0%). The net effect is a measurable shift of work balance toward individual, documentation-focused tasks.

Key Points

• Magnitude and dose-response
  • Productivity app actions (Word/Excel/PowerPoint) increase by ~21.2% for Copilot users with ≥100 uses (20-week post period); increases rise up to ~27% for ≥500 uses.
  • Communication app actions (Outlook/Teams) increase by ~7.1% for ≥100 uses, up to ~13% for heavier users.
  • Productivity-app share (productivity actions / (productivity + communication actions)) shifts upward modestly (DiD: ~1.4% → 2.1% across intensity thresholds); baseline communication actions were ~2.43× productivity actions.
• Heterogeneous action-level effects
  • Productivity actions increase broadly across specific actions (e.g., Word: Paste +4.4%, Open App +7.1%, both p<0.05).
  • Communication shows mixed changes: reductions in information-consumption and management actions (e.g., Outlook “read” decreases ~5.5%; “move” emails decreases ~7.0%, p<0.05).
  • Small-group email behavior declines (for emails <10 recipients): total small-group emails −4.7%, unique recipients −1.6%, conversation rounds −2.1% (p<0.05); these declines are larger with greater Copilot use.
• Functional links
  • Use of Outlook AI features (summarize, suggested replies, etc.) is correlated with reductions in reading/management actions; suggested replies particularly associated with reduced “read” actions.
• Interpretation / mechanisms
  • Three candidate mechanisms: (1) expanded individual bandwidth (people solve more tasks without colleagues), (2) reduced information overload (less time spent consuming low-value info), (3) AI primarily complements documentation/knowledge tasks (so productivity apps benefit more).
• Caveats
  • Observational study with DiD+PSM to mitigate confounding, but residual selection or unobserved time-varying differences may remain.
  • No access to prompt or content-level data (privacy constraints), so quality of output and which communications were avoided cannot be assessed.

Data & Methods

• Data source: Microsoft 365 digital trace data from multiple large international companies that enabled Microsoft Copilot in early 2024.
• Outcome measures: counts of user-initiated (human) actions in productivity apps (Word, Excel, PowerPoint) and communication apps (Outlook, Teams). Also action-level counts (top-10 actions per app) and a productivity-app use ratio (productivity actions / (productivity + communication actions)).
• Treatment definition: users who enabled Copilot and used it at least a threshold number of times (primary threshold: ≥100 uses) during the first 20 weeks after enablement; additional analyses used 200/300/400/500-use thresholds.
• Comparison group: users who enabled Copilot after the 20-week observation window; control set constructed via propensity-score matching on pre-enablement activity with exact matching on managerial position.
• Empirical strategy: Difference-in-Differences (DiD) comparing pre vs. post enablement trends between early adopters (treatment) and later adopters (control). Additional cross-sectional correlations relate specific AI function counts to changes in specific actions.
• Time window: 10 weeks pre-enable, 20 weeks post-enable for main analyses.
• Significance: many effects reported at p<0.05; detailed tables and robustness checks are in the supplementary information (SI).

Implications for AI Economics

• Task reallocation and labor demand
  • AI appears to reallocate effort toward documentation and individual knowledge work. Jobs or tasks that center on information-synthesis, document creation, or producing artifacts may see productivity gains and potentially higher marginal output per worker.
  • Roles whose value derives from coordination, information triage, or routine communication management may face reduced demand or changing task mixes (substitution/complementarity effects).
• Coordination costs and organizational design
  • Evidence of fewer small-group emails and fewer rounds of back-and-forth suggests lower coordination frictions—firms may be able to compress coordination overhead and flatten some communication layers.
  • Reduced communication could lower transaction costs, but may also reduce informal exchanges that seed innovation; organizations may need deliberate practices to preserve cross-cutting communication.
• Innovation, information diffusion, and externalities
  • A shift away from communication toward focused documentation work can increase individual output but risks reducing the frequency/diversity of interpersonal contacts that generate serendipitous information flows and novel ideas—possible negative externality on innovation.
  • Policymakers and managers should monitor not just productivity metrics but team-level innovation outcomes, cross-team knowledge spillovers, and long-run creative output.
• Inequality and distributional effects
  • Gains are concentrated in knowledge-intensive tasks; workers who perform heavily document-centric work or possess complementary skills to AI may benefit more, potentially widening within-firm or within-occupation inequality.
  • Conversely, employees whose tasks are mainly communication/coordination might see their relative importance decline unless roles are redesigned.
• Measurement and evaluation
  • Digital trace action counts are a useful high-frequency proxy for behavioral change but do not capture output quality, wages, promotions, or welfare. Economic evaluations should pair trace measures with outcome variables (error rates, revenue, innovation counts, career outcomes).
• Policy and managerial responses
  • Encourage complementarities: train workers to use AI for higher-value tasks; redesign incentives to reward collaboration that AI cannot easily substitute.
  • Preserve beneficial communication: introduce norms/tools that maintain inter-personal contact and knowledge diffusion (e.g., structured cross-team interactions, incentives for knowledge sharing).
  • Monitor long-run impacts on innovation and labor demand; consider support for role transitions where coordination/triage tasks decline.
• Research agenda
  • Quantify impacts on output quality, team innovation, promotions/salary, and employment composition.
  • Heterogeneity by role, seniority, industry, and firm structure to understand distributional implications.
  • Long-run causal evidence (panel linkage to performance metrics, randomized rollouts, prompt/content-level studies where privacy permits).

Summary takeaway: Generative AI adoption materially raises documentation/productivity activity and moderately affects communication activity, producing a systematic reallocation of workplace behavior toward individual artifact creation. This has clear productivity upside but raises important questions for coordination, innovation, and labor reallocation that economists and managers should study and address.