UK Government Digital Service
UK government coding assistant trial: users estimate 56 minutes saved daily, 15.8% of suggested code accepted
The UK Government Digital Service made 2,500 AI coding assistant licences available to engineers in more than 50 public sector organisations for three months. Survey respondents estimated they saved 56 minutes a working day, yet GitHub Copilot telemetry showed only 15.8% of suggested code lines were accepted.
Key figures
Copied exactly from the sources. Copy any figure with its citation.
56
minutes
Average time saved per working day, as estimated by trial participants in a survey (424 responses from 31 departments)1
58%
Survey respondents who said they would not want to return to their working conditions before the assistants1
15.8%
Suggested code lines accepted, GitHub Copilot telemetry1
What it took
The problem
Government engineering teams build and maintain the software behind public services. The Government Digital Service (GDS) wanted evidence on whether commercial AI coding assistants would make public sector engineers more productive, since little research covered the public sector.
What they built
GDS ran a cross-government trial of two off-the-shelf AI coding assistants: GitHub Copilot and Gemini Code Assist. These tools suggest code as the engineer types, answer questions in a chat window and help review existing code. An internal commercial review in July 2024 had judged them the most mature options at the time.
How they did it
The trial ran for three months from November 2024 to February 2025. GDS made 2,500 licences available across more than 50 organisations and supported deployment, user engagement and adoption. It collected survey answers at the start, middle and end, tool usage data from the GitHub Copilot dashboard, eight interviews and success stories. The main analysis used 424 survey responses from 31 departments.
Results
Participants estimated an average saving of 56 minutes per working day, with the largest share coming from creating code and analysis. Of respondents, 58% would prefer not to go back to working without the tool, and satisfaction averaged 6.6 out of 10.
What went wrong
The usage data was less flattering than the survey. GitHub Copilot users accepted only 15.8% of suggested code lines, which the report calls slightly lower than industry reports, and only 39% of users reported committing suggested code. The press release says 15% of generated code was used without edits. Telemetry was missing for a month, licences were distributed late in some organisations, and 8% of users saw no significant time saving.
What to copy
Run a time-limited trial with a central team that supports onboarding, and publish the limitations as GDS did. Pair what users say with what the tools record, and avoid sizing a business case on self-reported minutes alone.
All reported figures
| Measure | Figure | Citation |
|---|---|---|
| Average time saved per working day, as estimated by trial participants in a survey (424 responses from 31 departments) | 56 minutes1 | |
| Survey respondents who said they would not want to return to their working conditions before the assistants | 58%1 | |
| Average overall satisfaction score | 6.6 out of 101 | |
| Suggested code lines accepted, GitHub Copilot telemetry | 15.8%1 | |
| Users who reported committing code suggested by the assistant | 39%1 | |
| Users reporting no significant time saving | 8%1 | |
| AI-generated code used without any edits (government press release) | 15%2 |
Lessons for your project
Measure usage with system data as well as surveys. Self-reported savings of 56 minutes a day sat beside a measured 15.8% acceptance rate for suggested code.
Tell staff how long access will last. The report says users who knew licences would be removed after the trial may have held back from using the tools fully.
Secure sponsorship inside each organisation. Engagement depended on local sponsors, and some licences reached users only after the first month.
Avoid a trial window that includes holidays. The central month fell over the festive period, which reduced user and team availability.
Keep humans responsible for the output. Government guidance told developers to understand and own any code the assistant produced.
Caveats
- The government's own Digital Service published the report. The headline time saving comes from participants' estimates in 10-minute steps, not from measured work.
- The report says savings across tasks may overlap and that respondents may have been over-optimistic, so the 56 minutes may be overstated.
- GitHub Copilot telemetry is missing for the second month of the trial. Gemini Code Assist had no comparable telemetry reported.
- Of 2,500 licences, 1,900 were assigned; 1,100 of 1,600 GitHub Copilot licences and 173 of 323 Gemini Code Assist licences were redeemed. The press release says more than 1,000 developers took part.
- The report itself says people were not tracked between the start, middle and end surveys, so changes over time are hard to judge.
- The report uses the figure 39% twice: key findings say 39% of users reported committing suggested code; the body says 39% believed productivity would fall if they lost access. The key-findings wording is used here.
- The report states the trial will not influence future procurement. The press release names DSIT as the trial's organiser; the report is published by GDS. The deployed date is the trial's start month.
Sources
- AI coding assistant trial: UK public sector findings report
Government Digital Service, GOV.UK · 2025-09-12 · Filing · gov.uk
- Government coders using AI to each save 28 days a year and build more tech
GOV.UK (DSIT) · 2025-09-12 · Press release · gov.uk
Summarized in our own words from the sources above. Spotted an error? See corrections.