Microsoft 365 Copilot is seen as user-friendly and reliable and yields clear efficiency gains for administrative work in a research institute, while researchers' productivity views improve only with role-specific training and governance measures.

Main Finding

In a repeated cross-sectional survey of employees at a large non-university research organization piloting Microsoft 365 Copilot, administrative staff report higher perceived usefulness and reliability overall, while scientific staff show significant positive shifts over time (T01 → T02) in perceived usefulness, ease of use, and perceptions of productivity/workload reduction. Copilot is seen as broadly user‑friendly and technically reliable and delivers the greatest incremental value for clearly structured, text‑based knowledge‑work tasks. The authors emphasize that context‑sensitive rollout, role‑specific training, and governance are critical to realize sustainable acceptance and organizational value.

Key Points

• Differential baseline perceptions
  • Administrative employees initially rate Copilot’s usefulness and reliability higher than scientific staff.
• Learning / routinization effects
  • Scientific staff’s mean perceived usefulness rose from 0.42 to 1.09 (scale –3 to +3; p = 0.022).
  • Ease of use for scientific staff increased from 0.74 to 1.31 (p = 0.016).
  • Output quality perceptions for scientific staff trended upward (0.26 → 0.73; p ≈ 0.067).
  • Reliability remained comparatively stable across groups/time.
• Task heterogeneity
  • Highest added value for structured, text‑centric tasks (e.g., drafting, summarizing, template generation, administrative workflows).
  • Less clear added value for open‑ended scientific reasoning, experiments, or tasks requiring domain‑specific validation.
• Practical recommendations from authors
  • Context‑sensitive implementation strategies, role‑tailored training and support, and clear governance (privacy, IP, compliance) are necessary for sustainable adoption.
• Limitations flagged by authors
  • Non-random quota sampling of license holders; modest sample sizes.
  • Data are repeated cross‑sections (not a longitudinal panel) and based on self‑reports; not representative of entire organization (~32k staff).
  • Multiple Copilot updates occurred during the observation window (8 documented updates), complicating attribution.

Data & Methods

• Setting and sample
  • Pilot rollout of Microsoft 365 Copilot in a large German non‑university research organization (Fraunhofer context).
  • 550 employees received licenses and invitations; two survey waves:
    • T01 (Nov–Dec 2024): N = 106 (66 scientific, 40 administrative)
    • T02 (Mar–Apr 2025): N = 90 (51 scientific, 39 administrative)
  • Pseudonymized IDs were available but overlap between waves was small, so analyses treated waves as repeated cross‑sections.
• Measures
  • Core acceptance constructs: perceived usefulness (PU; 5 items), perceived ease of use (2 items), output quality (1 item), reliability (1 item), voluntariness, and task‑specific usefulness (9 items).
  • Items on 7‑point Likert scale, coded –3 to +3.
  • PU scale reliability: Cronbach’s α = 0.97 (T01), 0.95 (T02). Ease‑of‑use inter‑item r = 0.58 (T01), 0.66 (T02).
• Analysis
  • Descriptive statistics by group and wave.
  • Welch t‑tests for group/time differences; Cohen’s d reported for effect sizes.
  • Task‑level usefulness analyzed descriptively (no multiple inferential tests to avoid α inflation).
• Contextual factor
  • Eight product/model updates to M365 Copilot between Nov 2024 and Apr 2025, including UI, integration, and stability changes.
• Key numerical results (selected)
  • Perceived usefulness (scientific): 0.42 → 1.09 (d_time ≈ 0.42, p = 0.022)
  • Perceived usefulness (administrative): 0.94 → 1.35 (d_time ≈ 0.38, p = 0.092)
  • Ease of use (scientific): 0.74 → 1.31 (d_time ≈ 0.45, p = 0.016)
  • Output quality (scientific): 0.26 → 0.73 (d_time ≈ 0.34, p ≈ 0.067)
• Limitations reiterated
  • Self‑report measures, potential selection bias toward tech‑interested users, small samples, inability to infer individual longitudinal change or causal productivity effects.

Implications for AI Economics

• Heterogeneous productivity effects
  • Productivity gains from generative AI are task‑ and role‑dependent: routine, structured, text‑based administrative tasks show clearer, earlier gains than exploratory scientific work. Economic models forecasting AI impacts on labor should incorporate task heterogeneity within occupations and organizations.
• Learning and adoption dynamics matter
  • Observable positive shifts for scientific staff across a short pilot suggest adoption and learning effects can materially change perceived usefulness over months. Economic assessments (cost‑benefit, ROI) should account for time‑dependent uptake and skill acquisition rather than one‑off productivity multipliers.
• Importance of complementary investments
  • Realizing AI’s value requires investments in role‑specific training, process redesign, and governance. From an economics perspective, total gains = technology capability + complementary human capital + organizational change; omitting complements will overestimate realized benefits.
• Governance and externalities
  • Concerns around reliability, IP, data protection, and compliance (especially in research organizations) can constrain deployment or impose mitigation costs. Policy and firm‑level governance choices will influence net economic returns and distributional outcomes.
• Procurement and update cycles
  • Product/model updates (8 during the study window) affect perceived performance and stability; procurement decisions and cost forecasting should consider ongoing vendor updates, maintenance, and monitoring costs.
• Research agenda suggestions (for AI economics)
  • Move from perception surveys to causal, within‑subject longitudinal designs and objective productivity metrics (time‑on‑task, output quality validated by peers).
  • Estimate heterogeneous treatment effects by task type, skill level, and occupation to inform labor demand forecasts.
  • Model the optimal allocation of training and governance spending across organization units to maximize net productivity.
  • Incorporate dynamic adoption curves and software update externalities into macro/micro estimates of AI’s economic impact.

Reference: Schmidt et al., "Generative AI in knowledge work: Perception, usefulness, and acceptance of Microsoft 365 Copilot" (Zeitschrift für Arbeitswissenschaft, accepted 25 Feb 2026; DOI: 10.1007/s41449-026-00517-5).