evidoria

← Back to browse

Good practice

DWP's Controlled Evaluation of Microsoft 365 Copilot — A More Rigorous Estimate of UK Civil-Service Time Savings

United Kingdom · London · See the United Kingdom profile

The UK's DWP ran a six-month Microsoft 365 Copilot trial with a genuine comparison group of non-users. Copilot users saved an average of 19 minutes a day — notably less than an earlier self-reported 26-minute figure from a 20,000-person cross-government trial.

19 minutes/day
Average daily time saved by Copilot users (Oct 2024-Mar 2025)
26 minutes/day
Time saved on information retrieval
25 minutes/day
Time saved on email writing
24 minutes/day
Time saved on summarising
9 minutes/day
Time saved on meeting transcription
95 %
Users reporting satisfaction
89 %
Users redirecting saved time to other tasks
0.56 points (7-point scale)
Job satisfaction increase
0.49 points (7-point scale)
Perceived work-quality increase
3,549
Copilot licences allocated
26 minutes/day
GDS self-reported comparison figure (no comparison group) (published June 2025)

Details

Maturity
Pilot
Promoter
UK Department for Work and Pensions (DWP)
Period
Oct 2024 – Mar 2025 (published Feb 2026)
Keywords
public administration, social security, generative AI, workplace productivity

Context

The UK Department for Work and Pensions, the government's largest public-facing department, piloted Microsoft 365 Copilot from October 2024 to March 2025 (evaluation published February 2026), allocating 3,549 licences across 12 licence groups covering core departmental functions and excluding frontline caseworkers.

Activities

DWP compared survey responses from 1,716 Copilot users (48% response rate) against 2,535 responses from a non-user comparison group (27% response rate), plus 19 in-depth interviews, using a Seemingly Unrelated Regression model to adjust for age, gender, grade, role, directorate, health conditions and prior AI experience.

Results

The adjusted estimate found Copilot users saved an average of 19 minutes per day (26 minutes on information retrieval, 25 on email writing, 24 on summarising, 9 on meeting transcription); 95% of users reported being satisfied, 89% redirected the time saved to other tasks, and job satisfaction and perceived work quality rose modestly (0.56 and 0.49 points on a 7-point scale).

Conclusions

DWP's report explicitly notes its 19-minute estimate sits below the 26-minutes-per-day figure from a June 2025 Government Digital Service trial of 20,000 civil servants that used self-reported measures with no comparison group -- illustrating how methodological rigour changes headline AI productivity estimates. DWP also disclosed its own trial's limitations: no randomisation, no pre-intervention baseline, and reliance on self-reported measures.

Implementation

Implementation detail (cost, timeline, staffing, conditions for success) is not yet available for this practice.

Data sources

Where this practice's information was retrieved from, and when.

Similar practices you may find useful