Key findings
- Reported savings ranged from about an hour a week at HMRC to 43 minutes a day in the NHS pilot of more than 30,000 staff, whose method is unpublished. DWP estimated 19 minutes a day once it controlled for staff interest in AI, and 10 to 12 minutes without those controls. The cross-government experiment reported 26 minutes a day across 20,000 licenses.
- DBT's staff diary showed small savings on written tasks and extra time spent on scheduling and images. DBT measured no output, and six control-group interviewees had not observed productivity gains in colleagues using Copilot.
- About 80% of cross-government license holders remained active after change management. HMRC reported use by 83% of licensed staff. Time saved tracked users' confidence with AI.
- Accuracy and trust limited use. DBT users reported hallucinations, and 46% of HMRC non-users cited security and data privacy as the main reason.
- The University of Washington chose licenses paid for by the teams that want them, after finding that a broad rollout would amplify oversharing risk.
Pilot paid Copilot with the roles that draft and summarize most, because that is where all six core evaluations found time saved. Five UK public bodies and the University of Washington published them between November 2024 and July 2026. Two stopped short of licensing everyone. DBT found no evidence that the savings raised productivity, and the University of Washington rejected central funding. Seven more trials, from Australia's Treasury to NCDOT, point the same way.
What was measured
The UK Government Digital Service ran 20,000 licenses across 12 organizations from September to December 2024. DBT gave 1,000 licenses to volunteers and a random group. It collected a staff diary and timed 11 participants against non-users on four tasks. DWP's regression compared 1,716 survey responses from licensed staff with 2,535 from non-users; 3,549 staff held licenses. HMRC randomly allocated 3,000 licenses, directly allocated 500 more, and surveyed 1,364 staff from the random group. The NHS reported a pilot across 90 organizations. The University of Washington piloted 200 users for five months and tested 425 prompts.
Where the time came from
Drafting and summaries led every study. Cross-government users saved 24 minutes drafting documents. DWP's largest savings were information search at 26 minutes and email writing at 25 minutes. HMRC users drafted most in Word, at 53%, and rated meeting summaries and document search most useful. The NHS counted meeting notes and email summaries.
Why two stopped short
DBT's evaluators could not link reported time savings to productivity, and PowerPoint and scheduling scored lowest on satisfaction. The University of Washington found most pilot users learned by trial and error, rated results unevenly from useless to tremendous. They judged that existing oversharing made a broad rollout too risky without remediation.
What other Copilot users report
Public accounts outside government repeat the pattern. Of 108 public accounts from Copilot users and admins, 9 of the 10 that discuss meeting recaps praise them, while answer quality draws 12 critical accounts to 4 positive. Accounts that describe how licenses were assigned favor selected roles first, 8 to 3.
Seven more trials
Australia's whole-of-government trial, published October 23, 2024, reported about an hour a day saved on summaries, searches and first drafts by self-assessment, a figure its evaluators call a likely upper bound; it counts licences as over 5,765 in one place and over 7,769 in another. Australia's Treasury found the rollout smooth and use lower than expected among 218 staff. Ohio University's yearlong pilot found steady savings of 1 to 3 hours a week for only part of 130 users, and Virginia Tech's 177 volunteers reported 38 minutes a day.
Fraunhofer's 550 licensed staff rated Copilot easy to use, with administrators gaining first. NCDOT's 175 employees rated it less useful after eight weeks than before, 3.62 against 3.85 on a five-point scale. New Zealand's MBIE reported better quality and speed on ministerial correspondence. Microsoft's own randomized study across 66 firms measured 1.4 hours a week less email time and found no change in what work people did.
Methodology
GILD analyzed six core published evaluations opened on October 12, 2026, seven more trials added on October 28, 2026, and 108 public accounts from Copilot users and admins. GOV.UK reports came from the Government Digital Service (June 2, 2025) and the Department for Business and Trade (August 28, 2025). Other GOV.UK reports came from the Department for Work and Pensions (January 29, 2026) and HM Revenue & Customs (July 9, 2026). The others were the NHS release on GOV.UK (October 21, 2025) and the University of Washington UW-IT report (November 15, 2024). Every time-saving figure is the publishing body's own. DBT's observed sessions timed 11 participants directly, the other headline figures are self-reported by staff, and the NHS release publishes no method. HMRC cut its survey estimates by about 20% because not every license holder used Copilot and respondents were more likely to be active users. Australia's Digital Transformation Agency published a separate whole-of-government evaluation on October 23, 2024, read among the seven added trials: 69% of its survey respondents said Copilot sped up their tasks, and its pages give two license counts, over 5,765 and over 7,769. Microsoft's 66-firm experiment is Microsoft Research's own. Each study's sample and method are on its linked page below.
Products in this piece
Sources
- Official documentation
- 4
- Public source
- 9
- Microsoft Learn: What is Microsoft Copilot?
- Microsoft Learn: License options for Microsoft Copilot
- Microsoft Learn: SharePoint Advanced Management overview
- Microsoft Learn: Copilot controls overview
- GOV.UK: Microsoft 365 Copilot Experiment, Cross-Government Findings Report (Jun 2, 2025)
- GOV.UK: Microsoft 365 Copilot pilot, DBT evaluation report (Aug 28, 2025)
- GOV.UK: An Evaluation of DWP's Microsoft Copilot 365 Trial (Jan 29, 2026)
- GOV.UK: HMRC evaluation report, phase 3 trial of Microsoft Copilot (Jul 9, 2026)
- GOV.UK: Major NHS AI trial delivers unprecedented time and cost savings (Oct 21, 2025)
- University of Washington UW-IT: Microsoft 365 Copilot executive recommendations and risks (Nov 15, 2024)
- Digital Transformation Agency: Evaluation of whole-of-government trial into generative AI, now available (Oct 23, 2024)
- Hacker News comment on a slow Microsoft 365 Copilot rollout at a non-tech company (Oct 5, 2025)
- Hacker News comment on a company buying ChatGPT Enterprise beside Copilot (Oct 10, 2025)