The problem
The trial sat inside the government’s push to drive AI adoption across the public sector. The Government Digital Service (GDS) set out to learn what value an AI assistant would bring when deployed across a large part of government, defining value as efficiency, task completion and user satisfaction (Government Digital Service).
The pain is routine administrative work: drafting documents, summarizing email chains and meetings, scheduling and searching for information. Before the trial, earlier departmental pilots had been limited to 300 licences at most, too small to show what would happen at the scale of a department.
What they did
GDS ran the cross-government experiment from 30 September to 31 December 2024. Each participating organization committed to at least 1,000 licences, and the 12 included the Department for Work and Pensions, HMRC, the Home Office, the Ministry of Justice, the Department for Business and Trade (DBT), the Welsh Government and the Office for National Statistics (Government Digital Service). Most organizations switched off internet search so the tool drew only on internal documents, within each user’s existing permissions.
GDS collected usage data from Microsoft’s dashboard, a survey that drew 7,115 responses, feedback forms and five focus groups. Support and training were left to each department, so the report says the user experience was not consistent.
DBT ran its own, more rigorous evaluation of its 1,000 licences (Department for Business and Trade). About 70% went to volunteers and 30% to randomly selected staff. Its evaluation team used a one-week diary study, interviews with users and a control group, and observed tasks in which outputs from users and non-users were blind-assessed.
Results
According to the GDS report, trial participants saved an average of 26 minutes a day, which it says would add up to 13 days over a working year. These figures are self-reported: users picked a time band, and GDS averaged the midpoints. Only 17% of users were not noticing any clear time savings, and more than a third reported saving over half an hour a day. Adoption reached 83% in October and stayed around 80%. In the survey, 82% said they would not want to return to working without the tool, and satisfaction averaged 7.7 out of 10.
DBT’s evaluation is more cautious. It found 72% of diary respondents satisfied or very satisfied, and small time savings across most tasks, largest for written work such as drafting reports (Department for Business and Trade). Scheduling and generating images took longer with the tool. In observed tasks, users summarized reports faster and better than non-users, and produced slides faster but with lower accuracy, while Excel data analysis was both slower and less accurate. In the diary study, 22% of respondents said they had identified hallucinations at some point in the pilot.
The government has moved ahead regardless. HMRC began what it called the largest Copilot rollout in UK government in October 2025, with 32,000 licences that year rising to 50,000 in 2026, after a pilot it said projected two million hours of personal productivity savings a year (THINK Digital Partners). Staff must complete a 90-minute training course before their licence is switched on.
What didn’t go to plan
The central finding of DBT’s report cuts against the headline: “The evaluation did not find evidence that time savings have led to improved productivity”, and control-group colleagues had not noticed productivity gains from those using the tool (Department for Business and Trade). The department had no counterfactual, its diary response rate was 32%, and its value-for-money assessment was not published.
Accuracy was a recurring concern. Beyond the hallucinations DBT’s respondents reported, the department found that quality assurance of outputs was inconsistent across users and tasks. The GDS report notes limits with complex, nuanced or data-heavy work; one policy participant said the tool “struggles with nuanced or context-heavy data requiring human judgement” (Government Digital Service).
Uptake was uneven inside the suite: Copilot adoption in Excel and PowerPoint peaked at 23% and 24%. GDS also could not measure how the time saved was spent. Neither report states licence costs.
Lessons
- Test self-reported savings against observed work. DBT’s observed tasks showed slower, less accurate Excel analysis even though diary users reported time savings for data analysis.
- Measure productivity, not just minutes. The GDS figure is a self-reported average; DBT found no evidence that saved time became more output.
- Match the tool to the task. Drafting and summarizing gained most, while scheduling and image generation took longer at DBT.
- Make training a condition of access. GDS linked lower AI familiarity to lower savings, and HMRC now requires a 90-minute course before a licence is enabled.
- Publish the value-for-money case before scaling. DBT’s assessment was not published when HMRC began a rollout of tens of thousands of licences.
Adapt it in your organization
Is this for you?
- A large share of staff time goes to drafting, summarizing, email and meetings.
- Your documents and permissions are well managed, since the assistant searches whatever each user can access.
- You can run a control group or observed tasks, not only a satisfaction survey.
- You can fund training and local support in every team that gets licences.
- You know the licence cost per user and what saved time would need to be worth to cover it.
A 90-day version
Pilot: give the assistant to a few hundred staff, part volunteers and part randomly selected, with a matched control group, and measure four common tasks through observed sessions as well as surveys. Owner: chief operating officer. Success measure: observed time per task down at least 20% on drafting and summarizing, with no fall in blind-assessed accuracy.
Metric to track
Observed time and accuracy on the same tasks with and without the assistant, alongside self-reported minutes saved per day. The GDS trial reported 26 self-reported minutes a day; DBT’s observed tasks ranged from large gains in summarizing to losses in Excel analysis. For a contrast in a different function, see the Klarna case.
Failure modes to pre-empt
- Reporting self-reported minutes to the board as productivity gains.
- Rolling out without training, so low-familiarity teams see little benefit.
- Using the assistant for data analysis without checking its accuracy.
- Exposing badly permissioned files through the assistant’s search.
Questions to take to your leadership team
- What would we have to observe, not just hear from users, before buying licences for everyone?
- Which tasks should we steer people away from using the assistant for?
- What is the licence cost per user, and how many saved minutes a day would cover it?
Executive FAQ
How much time did Copilot save UK civil servants?
An average of 26 minutes a day, self-reported, in the GDS trial of 20,000 users; DBT’s separate evaluation found no evidence that this raised productivity.
Is the UK government still using Copilot?
Yes. HMRC began a rollout of 32,000 licences in October 2025, rising to 50,000 in 2026, with mandatory training first.
About the curator & AI disclosure
Curated by Luciano Ayres
Luciano Ayres is Director of Artificial Intelligence at ArcTouch, with more than 20 years of experience in software and technology leadership and specialized studies in deep learning. Connect with him on LinkedIn.
AI-generated briefing. This case was produced through an automated AI research and writing workflow curated by Luciano Ayres. Sources are linked for independent verification; summaries and analysis may contain errors and do not replace professional due diligence.