Two published UK government evaluations of the same AI work assistant, Microsoft 365 Copilot, reached different conclusions about productivity. In a trial of 20,000 civil servants from September 30 to December 31, 2024, users reported saving an average of 26 minutes a day, based on a self-reported survey. A separate evaluation inside the Department for Business and Trade found small time savings but no evidence that they led to improved productivity.
Key takeaways
- The cross-government figure of 26 minutes a day comes from survey answers sorted into time ranges, with the average calculated from the midpoint of each range.
- The department-level evaluation recorded the largest time savings on written tasks and extra time on scheduling and image generation.
- In 11 observed task sessions, results varied by task: summaries were faster and scored higher, while the Excel analysis and slide tasks scored lower with the assistant.
- Neither report measured what staff did with saved time, so time saved is not the same as output gained.
The cross-government trial: scale and method
The Government Digital Service (GDS) published its cross-government findings report on June 2, 2025. Twelve organizations took part, each committing at least 1,000 licenses, and the report says it was the largest deployment of the product in any organization at that point; earlier pilots had been limited to 300 licenses at most. Data came from usage dashboards, a user survey with 7,115 responses, feedback forms, and five focus groups.
Roughly 80 percent of licensed users stayed active over the trial, and Teams was the most used application, with a peak adoption of 71 percent. Survey respondents did not measure their time. They picked a range for daily time saved, from "none or less than 5 minutes" up to "more than an hour." The report calculated its 26-minute average from the midpoint of each range and noted that 17 percent of users saw no clear saving. It also projected 13 days saved per year, an extrapolation from that self-reported average. By task, drafting documents showed the largest saving (24 minutes), followed by creating presentations (19 minutes) and scheduling meetings (9 minutes).
The authors state that these self-reported savings should be read alongside other literature on their real value. They also write that the experiment could not identify how the saved time was spent, and that users with more familiarity with AI tools saved more time.
The department-level evaluation
The Department for Business and Trade (DBT) ran its own pilot with 1,000 licenses from October to December 2024 and published an evaluation report in August 2025. About 70 percent of participants were volunteers and about 30 percent were randomly selected. The main instrument was a diary study with 300 usable diaries, a 32 percent response rate. The team also held 19 interviews, including six with colleagues who had no access, and ran observed tasks.
Satisfaction was high: 72 percent of respondents said they were satisfied or very satisfied, and the Net Promoter Score was 31. Written tasks such as drafting and summarizing earned the largest time savings, while scheduling and image generation took longer. The report says the evaluation did not find evidence that time savings had led to improved productivity, and it adds that this was not a key aim, so little data was collected on the question. Of diary respondents, 22 percent said they had identified false or misleading output, 43 percent said they had not, 11 percent were unsure, and 22 percent did not answer.
Observed tasks: four results from very small groups
To check the diaries, DBT staff watched people do short fictional work tasks with and without the assistant and had a second analyst score the outputs blind on a scale of one to five. Each group had two or three people, and the report says the results are supplementary and may reflect factors other than the tool. The table reproduces the reported means.
| Task | Accuracy (with / without) | Quality (with / without) | Time (with / without) |
|---|---|---|---|
| Excel data analysis | 1.5 / 2.7 | 1.5 / 2.7 | 25:01 / 20:33 |
| PowerPoint slide | 1.5 / 5 | 1 / 2 | 10:47 / 18:30 |
| Summarizing two reports | 4 / 2.5 | 4 / 3 | 12:37 / 41:34 |
| Writing an email | 4.3 / 4 | 4.7 / 3.5 | 7:30 / 7:43 |
The pattern differs from the diary results. Diary respondents reported time savings on data analysis, yet the observed Excel task took longer. Summarizing produced the clearest gain on both speed and score. With so few participants, none of these rows should be treated as an estimate for a whole workforce.
Why the two findings can both hold
A survey asks people how much time they believe they saved. An observation measures the clock. A person can honestly feel faster while checking, correcting, and reformatting add time that is not counted. DBT's diaries asked about checking time: for emails, 51 percent of reported cases said checking took less time with the assistant, while 30 percent of slide-deck cases said it took longer. The GDS focus groups raised the same concern, reporting that the tool struggled with nuanced or data-heavy work and that human oversight was needed.
The second gap is between time and output. Saved minutes may go to training, a walk, or more strategic work, as interviewees described, or they may be absorbed by new tasks. DBT found no evidence either way on productivity, and GDS said it could not trace the use of the time.
Questions to ask about any work-tool trial
- Were time savings measured by observation or reported by users?
- Was there a comparison group that did not use the tool?
- Did participants volunteer, and what was the response rate?
- Is checking and correction time counted?
- Which tasks showed gains, and which showed losses?
- Does the study track output or quality, or only minutes?
The product is named here only because it is the subject of the reports. This article does not recommend or rank any tool, and the reports cover UK public-sector settings and one product, so they do not test other tools or other workplaces.
FAQ
Does 26 minutes a day mean 13 working days a year?
That is the report's own extrapolation. It rests on self-reported ranges averaged at their midpoints, and the report says the figure should be read alongside other literature on the real value of such estimates.
Did the trials use a control group?
The cross-government methodology describes usage data, a survey, and focus groups, not a control group. DBT described its design as evidence without a counterfactual, but it did interview colleagues without access and ran observed tasks with a no-assistant group.
Which tasks gained most from the assistant?
Both reports point to written work. GDS found drafting documents saved the most time, and DBT found written tasks produced the largest savings and the clearest gain in the summarizing task.
Did the reports find that users trusted the output?
Not without checking. DBT recorded inconsistent quality assurance across participants and task types, and GDS recorded concerns about incorrect information and the need for human oversight.
Related reading
- Top AI tools by industry
- AI for complex problems and efficiency
- AI and human-computer interaction
- Implementing AI training
More in AI Basics.
Last verified: October 2026
Sources
- Microsoft 365 Copilot Experiment: Cross-Government Findings Report, Government Digital Service (UK), June 2025
- The Evaluation of the M365 Copilot Pilot in the Department for Business and Trade, Department for Business and Trade (UK), August 2025