At a glance
- Company
- HM Revenue & Customs (HMRC)National tax authority
- Size
- Enterprise (1,000+ employees) The trial covered 3,500 licences in a national government department.
- Where and when
- United Kingdom. September – December 2024, published July 2026.
- Who produced the figures
- Official recordThe figures come from HMRC’s own evaluation. The main source was a survey of 1,364 users, 45% of those given a licence at random, with usage data, focus groups and a task-based exercise. The time saved is what staff estimated, not timed.
What they did
HMRC, the UK tax authority, tried Microsoft 365 Copilot in three phases. Copilot is an AI assistant inside Outlook, Teams, Word, Excel and PowerPoint. The first two phases, in mid-2024, used 300 staff who had volunteered.
The third phase ran from September to December 2024:
- 3,000 licences went to staff picked at random, among staff whose roles mainly use Microsoft Office. They were spread across the authority’s business groups in proportion to their size.
- 500 more licences went to selected staff. These were volunteers who needed Copilot as a reasonable adjustment for a disability or health condition, or who had taken part in the earlier phases.
HMRC then surveyed users, read the usage data, ran focus groups with 35 people and a task-based exercise with 49.
The numbers
- 83% of staff with a licence used it during the trial.
- 7.1 out of 10: average satisfaction.
- 61% said they would be disappointed to lose the licence.
- 2–3% of the working week: the time staff estimated they saved. HMRC equates this to about 60 minutes a week for a full-time employee.
- About 20%: how much HMRC reduced the self-reported savings, because not every licence holder used Copilot and active users were likelier to answer the survey. HMRC calls the result “a baseline estimate rather than a precise measure of impact”.
- 46% of staff who did not use their licence said security and data privacy concerns were the main reason.
HMRC’s analysts also estimate that up to 50,000 licences would bring about £50 million a year in net productivity benefits. That is a projection, not a measured result.
What the evaluation leaves open
- The time saved is an estimate by staff. The report compares no timed results between licence holders and staff without a licence.
- The 500 selected staff are not reported separately. Their results cannot be told apart from those of the staff picked at random.
- Staff asked for more help. The main barriers after security were the need for practical training fitted to their work, and time to learn the tool.
The wider trial
Across government, users in an earlier trial with about 20,000 staff estimated that Copilot saved them 26 minutes a day. HMRC describes its own estimate as more conservative. A pilot at the Department for Business and Trade timed some tasks and found no evidence of higher productivity.
What the case shows
- Picking users at random gives a fairer picture. Volunteers are often the keenest users. HMRC’s random allocation and its 20% reduction both work against that.
- An hour a week is a small share of the week. The same tool, measured in different ways, gave 26 minutes a day across government and about an hour a week at HMRC.
- A data rule shapes how many people use the tool. Nearly half of the staff who did not use Copilot named security or privacy as the main reason.
What this says about the practices
Practices the company applied
- Start with ready-made AI in software you already use, before building your own HMRC tested an assistant built into the Microsoft Office software its staff already used, with 3,000 licences for three months. Do first · Effort: low (days, one person)
- Measure time saved per task against a baseline HMRC picked licence holders at random instead of only taking volunteers, and reduced self-reported time savings by about 20% because survey respondents were likelier to be active users. Do first · Effort: low (days, one person)
- Decide what data may go into which AI tools During the trial, staff could use Copilot only on material below the Official Sensitive security classification. Do first · Effort: low (days, one person)