UK Department for Work and Pensions
UK DWP trial gives 3,549 staff Microsoft 365 Copilot and estimates 19 minutes saved a day against a comparison group
The UK Department for Work and Pensions gave Microsoft 365 Copilot licences to 3,549 central-office staff for six months and compared their survey answers with a sample of colleagues without licences. The department's model estimates an average saving of 19 minutes a day across eight routine tasks. It warns that the results are self-reported and not from a randomised trial.
Key figures
Copied exactly from the sources. Copy any figure with its citation.
19
minutes
Estimated average time saved per day across eight routine tasks, licence holders versus comparison group, from the department's regression model with all controls1
0.56
points
Increase in job satisfaction on a 7-point scale compared with non-users, from the department's regression model1
0.49
points
Increase in perceived quality of work on a 7-point scale compared with the comparison group, from the department's regression model1
90%
Users in the survey who indicated that Copilot helped them save time1
The problem
A large government department spends much of its staff time on routine tasks such as drafting emails, searching for information and summarising documents. DWP wanted evidence on whether a paid AI assistant would save time and improve work, and whether earlier government results held up when tested against a comparison group.
What they built
DWP ran a trial of the licensed version of Microsoft 365 Copilot. This is an AI assistant built into Word, Outlook, Teams and other Microsoft apps, with access to the user's organisation data. Licences went to central-office staff in 12 groups tied to business areas such as digital, policy and finance.
How they did it
Between October 2024 and March 2025, DWP allocated 3,549 licences through volunteers and peer nominations. Researchers sent a survey to all licence holders and to a random stratified sample of 9,300 staff without licences, and received 1,716 and 2,535 responses. They compared the two groups with regression analysis, a statistical method that adjusts for differences such as grade and AI interest, and added 19 interviews with staff.
Results
The model estimated an average saving of 19 minutes a day across eight routine tasks. The report says the largest reductions were in searching for information and writing emails. Job satisfaction was 0.56 points higher and perceived work quality 0.49 points higher on a 7-point scale than in the comparison group. In the survey, 90% of users said Copilot helped them save time, and 65% used it daily. Users said they reinvested the time in project delivery, planning and supporting senior leaders.
What to copy
Fix the outcome measures and the analysis plan before collecting data, and survey a comparison group as DWP did, so that the result can be read next to other trials. Collect a baseline too, which DWP did not.
All reported figures
| Measure | Figure | Citation |
|---|---|---|
| Estimated average time saved per day across eight routine tasks, licence holders versus comparison group, from the department's regression model with all controls | 19 minutes1 | |
| Increase in job satisfaction on a 7-point scale compared with non-users, from the department's regression model | 0.56 points1 | |
| Increase in perceived quality of work on a 7-point scale compared with the comparison group, from the department's regression model | 0.49 points1 | |
| Users in the survey who indicated that Copilot helped them save time | 90%1 | |
| Users in the survey who reported using Copilot daily | 65%1 |
Lessons for your project
Build a comparison group into the trial. DWP surveyed licence holders and a random stratified sample of staff without licences, which lets it estimate a difference and not just report opinions.
Control for how keen people are on AI. The estimated saving grew when the model accounted for AI interest and experience, so skipping this can misstate the effect.
Plan training that fits roles. Most users taught themselves, and many asked for role-specific sessions and quick-reference tips.
Keep human review in place. Users described Copilot as a starting point and said its output needed their editing and judgement.
Caveats
- The department evaluated its own trial. All outcomes are self-reported in surveys, with no objective measure of time.
- Licences went to volunteers and peer nominations, first come first served, so it was not a randomised trial and selection bias is possible. No baseline was collected before the trial.
- The 19 minutes comes from the model that controls for AI interest and experience. Models without those controls give lower estimates, between 10 and 12 minutes in the report's Table 4.3, which the report explains by AI-keen staff spending more time on tasks anyway.
- The trial covered central-office staff only. Jobcentre and other frontline staff had no licences, and the free version of Copilot reached all staff only in April 2025, after the fieldwork.
- The sources name the product, Microsoft 365 Copilot, but not the underlying language model.
- DWP also took part in the separate cross-government Copilot trial run by the Government Digital Service. This case covers DWP's own evaluation, published separately.
Sources
- An Evaluation of DWP's Microsoft 365 Copilot Trial
GOV.UK (Department for Work and Pensions) · 2026-01-29 · Research paper · gov.uk
Summarized in our own words from the sources above. Spotted an error? See corrections.