The 26 Minutes a Day Figure Measures Time Saved, Not Work Done

If you are weighing paid AI assistant seats, the headline number from the UK government trial answers a smaller question than it appears to.

The Snapshot

The Government Digital Service ran a cross-government Microsoft 365 Copilot experiment from 30 September to 31 December 2024, giving licences to 20,000 employees across 12 organisations. ^[1][2] GDS reported that users saved an average of 26 minutes a day, but says experimental constraints made it impossible to identify how that saved time was spent. ^[1] For a firm your size, the useful question is not whether staff save minutes but whether those minutes turn into more or better work, and none of these evaluations shows that. That gap matters most to owners of professional services firms being asked to approve per-seat spend on the strength of one number.

Where the 26 minutes comes from

The 26 minutes is a self-reported estimate, not a measured saving. ^[1] Survey respondents picked a range for the average daily time they thought Copilot saved them, from none or less than 5 minutes up to more than an hour. ^[1] GDS then took the midpoint of each range, counted the top answer as 60 minutes, and extrapolated the result to 13 days a year. ^[1]

The report describes no control group and no pre-trial baseline. ^[1] Results came from Microsoft's usage dashboard, an end-of-trial survey, open-ended feedback forms, five focus groups and departmental success stories. ^[1] GDS also warns that time savings of this nature should be read alongside other literature on how such figures translate into real impact. ^[1]

So the number tells you what a large set of civil servants felt, scored at the end of a three-month trial, and nothing about what changed downstream.

Three more departments, three more designs, same blind spot

The Department for Business and Trade piloted Copilot with 1,000 licences for UK-based staff from October to December 2024 and published its evaluation on 28 August 2025. ^[3][4] DBT's main method was a self-reported diary study, supported by usage data, interviews and a small set of observed, timed tasks. ^[4][5] DBT collected no data from participants before the pilot, so no before-and-after comparison was possible. ^[5][6]

DBT did not find evidence that time savings led to improved productivity. ^[7][5] DBT notes that productivity was not a key aim of the evaluation, so limited data was collected on it. ^[7][5] DBT also says its findings are not comparable with those of other departments, because methods and evaluation aims were prioritised differently. ^[5]

The Department for Work and Pensions evaluated its own trial of 3,549 staff running from October 2024 to March 2025. ^[6] DWP compared survey responses from 1,716 users with 2,535 non-users and estimated an average self-reported saving of 19 minutes per day. ^[6] DWP says the study was not a randomised controlled trial: licences were self-selected, there was no pre-trial baseline, and all outcomes were self-reported. ^[6]

HM Revenue and Customs randomly allocated 3,000 Copilot licences in a trial running from September to December 2024. ^[8] HMRC says participants self-reported saving 2 to 3% of their working week, around 60 minutes per week for a full-time equivalent working pattern. ^[8] HMRC reduced those self-reported estimates by approximately 20% to reflect non-users and response bias. ^[8]

We read these four evaluations as using different designs, all reporting time savings, and none measuring productivity directly.

The measured evidence, and what it still leaves out

Self-reports are not the only evidence available. A randomized field experiment across 56 firms and more than 6,000 workers, run by Microsoft researchers, used product telemetry rather than self-report. ^[9] Workers given Copilot licences spent 12 fewer minutes a week reading email on average, a 7% decrease, and regular users spent about half an hour less, an 18% decrease. ^[9]

That experiment measured time use and output proxies such as Word documents created, but it did not measure the quality of work. ^[9][10] There was also no significant change in time spent in meetings, even though meetings were where workers used Copilot most. ^[9][10] Adoption was uneven: on average 38% of licensed workers used Copilot each week, with firm averages ranging from about 9% up to 75%. ^[9][10]

DBT says its small observed-task sessions showed trends similar to what pilot participants self-reported, which it reads as self-report bias having small impacts on its time-saving findings. ^[5] In DBT's observed tasks, about eleven people, Copilot users were faster at summarising reports and building slides but slower at Excel data analysis. ^[5] Their Excel and PowerPoint outputs scored lower on accuracy and quality than non-users' outputs. ^[5]

The time savings look real enough; the link from saved minutes to better output is the part nobody has yet shown.

What this means when you are buying seats

The evidence partly supports treating the 26-minute figure as a self-reported estimate from a large public-sector rollout, not a forecast for a small firm. Suppose saved minutes are absorbed as on-the-job leisure rather than extra output; survey-reported savings would still look the same. GDS found a correlation between a user's confidence and familiarity with AI and the time savings they reported. ^[1]

DBT's report did not include a value-for-money assessment; it said one was being conducted and that further evaluation of value for money was needed. ^[4][7][5] DBT's usage data showed an average of 1.14 Copilot actions per user per working day across 63 working days. ^[7][5]

That points to running a small trial of your own before buying at scale.

What to do with this

• Measure three things in any trial you run: weekly use per person, task-level time, and the quality of the output.
• Track how many of your paid seats are actually used each week, because adoption in the measured experiment ranged widely between firms.
• Pick the tasks the evidence favours for your first test, starting with drafting and summarising rather than spreadsheets and slides.
• Budget for training and support, since GDS found reported savings correlated with users' confidence and familiarity with AI.
• If you are asked to justify the spend, note that DBT said further evaluation was needed to assess value for money. ^[4][7][5]

Limitations

DBT's finding of no productivity evidence is not evidence of no effect, since the design could not measure productivity. ^[7][5] DBT's observed-task comparison rests on a very small sample covering four tasks. ^[5] The main measured Copilot evidence comes from two papers on one experiment run by Microsoft, using Microsoft product telemetry, with firms recruited by Microsoft's marketing team. ^[10][9] None of these studies looked at firms the size of yours, so applying them to a small business is a judgement, not a finding.

Sources

• [1] Government Digital Service (no individual authors named). "Microsoft 365 Copilot Experiment: Cross-Government Findings Report (HTML)". 2 June 2025. https://www.gov.uk/government/publications/microsoft-365-copilot-experiment-cross-government-findings-report/microsoft-365-copilot-experiment-cross-government-findings-report-html (accessed 2026-10-19).
• [2] Government Digital Service. "Microsoft 365 Copilot Experiment: Cross-Government Findings Report". 2 June 2025. https://www.gov.uk/government/publications/microsoft-365-copilot-experiment-cross-government-findings-report (accessed 2026-10-19).
• [3] Department for Business and Trade. "Microsoft 365 Copilot pilot: DBT evaluation report". 28 August 2025. https://www.gov.uk/government/publications/microsoft-365-copilot-pilot-dbt-evaluation-report (accessed 2026-10-19).
• [4] Daisy Thomas (Digital Evaluator, Principal Statistician) and Vicki Woolgar (Internal communications lead for AI), DBT Digital trade blog. "Discover DBT's M365 Copilot evaluation report". 25 September 2025. https://digitaltrade.blog.gov.uk/2025/09/25/discover-dbts-m365-copilot-evaluation-report/ (accessed 2026-10-19). As reported from Department for Business and Trade (Monitoring and Evaluation team, DDaT), "Evaluation of the M365 Copilot pilot in the Department for Business and Trade".
• [5] Department for Business and Trade (Monitoring and Evaluation team, DDaT). "Evaluation of the M365 Copilot pilot in the Department for Business and Trade". 28 August 2025 (gov.uk publication date). https://assets.publishing.service.gov.uk/media/68adbe409e1cebdd2c96a19d/dbt-microsoft-365-copilot-evaluation.pdf (accessed 2026-10-19).
• [6] Department for Work and Pensions (Francesco Arzilli, Callum Seán Lynch, Lucy Page). "An Evaluation of DWP's Microsoft 365 Copilot Trial". 29 January 2026. https://www.gov.uk/government/publications/an-evaluation-of-dwps-microsoft-copilot-365-trial/an-evaluation-of-dwps-microsoft-365-copilot-trial (accessed 2026-10-19).
• [7] Public Technology staff (Civil Service World). "Copilot: DBT trial finds time savings but no productivity improvement". 10 Sep 2025. https://www.civilserviceworld.com/news/article/dbt-trial-finds-copilot-saves-staff-time-but-doesnt-improve-productivity (accessed 2026-10-19). As reported from Department for Business and Trade (Monitoring and Evaluation team, DDaT), "Evaluation of the M365 Copilot pilot in the Department for Business and Trade".
• [8] HM Revenue and Customs. "Evaluating the Impact of Microsoft Copilot in HMRC". 9 July 2026. https://www.gov.uk/government/publications/evaluation-report-phase-3-trial-of-microsoft-copilot/evaluating-the-impact-of-microsoft-copilot-in-hmrc (accessed 2026-10-19).
• [9] Eleanor Wiske Dillon, Sonia Jaffe, Sida Peng, Alexia Cambon (Microsoft). "Early Impacts of M365 Copilot". January 6, 2025 (date shown on paper; arXiv v1). https://arxiv.org/html/2504.11443v1 (accessed 2026-10-19).
• [10] Eleanor Wiske Dillon, Sonia Jaffe, Nicole Immorlica (Microsoft Research); Christopher T. Stanton (Harvard Business School, NBER, CEPR). "Shifting Work Patterns with Generative AI". April 14, 2025. https://arxiv.org/html/2504.11436v1 (accessed 2026-10-19).