Microsoft 365 Copilot Rollout: What 20,000 Users Showed

Most Microsoft 365 Copilot business cases are written before anyone has measured a single licence. The vendor deck promises hours back per person per week, finance multiplies that by headcount, and the board approves a rollout on a number nobody in the building has observed. The UK government did something rarer. It gave Microsoft 365 Copilot to 20,000 civil servants for three months and wrote down what happened, and one participating department then ran its own, more sceptical evaluation of a 1,000-licence pilot over the same quarter.

Read together, the two reports are the most useful public rollout dataset a UK buyer can get. The Government Digital Service cross-government findings report recorded an adoption rate of 83% after the October licence roll-out and an average self-reported saving of 26 minutes a day. The Department for Business and Trade evaluation found small time savings, negative ones on some tasks, and no robust evidence that the savings became productivity. Both are true at once, and the gap between them is where an enterprise rollout succeeds or stalls.

This post reads both reports as rollout evidence, not as a product verdict. It is written for the CIO or COO deciding whether to extend Copilot beyond a pilot, and it stays on Microsoft 365 Copilot, the assistant inside Word, Excel, PowerPoint, Outlook and Teams. It cedes GitHub Copilot as a coding agent to our grading of coding-agent artefacts, and it cedes the headcount question to our UK evidence on AI and headcount. It quotes no licence price, because neither report does.

This guide is shipped by the team behind DATS, the five-stage AI consulting system DILR.AI uses to place AI where it pays and keep it out where it does not. Or see our AI operating model work, which covers the governance, RACI and lifecycle a tool like Copilot needs once it leaves the pilot.

What did the UK government's Microsoft 365 Copilot trial actually measure?

The Government Digital Service trial gave Microsoft 365 Copilot licences to 20,000 civil servants across 12 organisations from 30 September to 31 December 2024. It measured adoption from the Microsoft usage dashboard, and time savings, satisfaction and benefits from a user survey with 7,115 responses plus five focus groups. Its quality evidence came from users' own reports, and it could not establish how the saved time was spent.

That scope matters more than any headline figure. Each organisation committed to at least 1,000 licences, and GDS describes it as the largest deployment of Microsoft 365 Copilot in all the organisations, with previous trials and pilots limited to 300 licences at most. The participants included HMRC, the Department for Work and Pensions, the Home Office, the Ministry of Justice, the Office for National Statistics and the Department for Business and Trade, among others. Anonymised usage data covered 14,500 users; for the rest, no data was supplied. If you are scoping a pilot of your own, ask us how we would measure it.

The definitions are worth copying into your own pilot plan before you read a single result. GDS counts an active user as anyone with at least one Copilot interaction in the previous 30 days, and the adoption rate as active users divided by licensed users. That is a generous bar: one prompt a month keeps a person in the numerator. The time savings were self-reported in ranges, from under five minutes to more than an hour, and GDS calculated the average from the midpoint of each range, with the top band capped at 60 minutes. GDS itself says those savings "should be considered alongside additional literature on the true value of these metrics".

So the GDS report answers three questions well: will people pick the tool up, where in the Microsoft suite will they use it, and how do they feel about it? It answers a fourth question, whether the organisation got more done, only by inference. The same distinction runs through our enterprise AI consulting guide, and through the DATS five-stage methodology, which runs from Discover and Diagnose to Scale and Run rather than stopping at a pilot.

How fast did Copilot adoption rise, and where did it fall away?

Microsoft 365 Copilot adoption in the GDS trial reached 83% after licences were rolled out in October 2024 and held at around 80% to the end of December. The headline hides a sharp split by application: peak adoption was 71% in Teams but only 24% in PowerPoint and 23% in Excel, and use inside Word and Outlook fell back from its peak.

The pattern is a communication tool being adopted far faster than a content or analysis tool. Survey respondents reported the same shape: 34% used Copilot daily in Teams and 33% daily in Outlook, while daily use in Word was 25%, with a further 43% using it weekly. OneNote and Business Chat went largely unused, with more than 55% of respondents not using Copilot in them at all. GDS records that Word adoption dropped by 5% from its peak by the end of the experiment and Outlook by 10% from its peak, while Copilot Chat recovered to within 3% of its peak.

Engagement looked strong on the survey's own terms. 39% of respondents used Copilot multiple times a day, a further 43% at least multiple times a week, and only 2% never used it. But the department-level picture from the Department for Business and Trade is more sober, and its measures are not the same as the GDS ones, so they should not be compared directly. DBT's usage dashboard showed a mean of 64% of licence holders active in a given week and 30% active on a given working day, with an average of 72 Copilot actions per user across the 63 working days of the pilot, or 1.14 actions per user per day.

The lesson for an enterprise is about which number goes in the steering pack. In a single organisation, a high 30-day active rate can sit beside a much lower working-day active rate, because one prompt a month keeps a person counted. If your rollout dashboard reports only the first, it will look healthy long after the habit has stopped forming. Settle the measure before you collect it, and prefer the one that would embarrass you if it fell; firms earlier in the journey can start with our AI readiness assessment for mid-market firms.

How much time does Microsoft 365 Copilot save, and how reliable is that figure?

Microsoft 365 Copilot users in the GDS trial reported saving an average of 26 minutes a day, which GDS extrapolated to 13 days across a working year. The figure is self-reported, calculated from range midpoints, and capped at 60 minutes for the top band. Only 17% noticed no clear saving, and more than a third reported saving over half an hour a day.

The task breakdown is more useful than the average. GDS found the largest savings in content creation: drafting documents saved 24 minutes and creating presentations 19 minutes, while communication tasks such as scheduling meetings saved 9 minutes but happened far more often. Savings were broadly consistent across grades and professions, and 8 of the 15 professions in the trial saved at least 26 minutes a day on average. Over 70% of users agreed that Copilot reduced time spent searching for information and on mundane tasks, 82% said they would not want to return to working without it, and 85% agreed it gave good value to the organisation.

Two caveats in the report decide how far a buyer can lean on these figures. First, GDS is explicit that "it was not possible to identify how time saved was spent". Second, it found a correlation between a user's confidence and familiarity with AI and the time they saved, which means the average is partly a measure of the people, not only of the tool. Professions with the lowest satisfaction were also those saving the least, and policy teams in particular struggled with nuance and with sources that held contrasting views.

None of this makes 26 minutes wrong. It makes it a perception of time saved, measured once, in a quarter partly lost to the festive period, which GDS itself names as a limitation. A board paper that treats it as recovered capacity, then multiplies by headcount, is doing arithmetic the report does not support. Our guide to board-level AI investment decisions sets out what an approval paper should hold instead: a stated assumption, the evidence behind it, and the point at which it will be tested.

Why did one department find no evidence of a productivity gain?

The Department for Business and Trade found no robust evidence that Microsoft 365 Copilot time savings became productivity, but says productivity was not a key aim and little data was collected on it. Its diary study also measured savings more strictly, subtracting unused outputs and tasks done only because Copilot existed, which left small savings, negative for scheduling and image generation.

DBT allocated 1,000 licences between October and December 2024, about 70% to volunteers and about 30% to a random sample stratified by directorate and grade, and evaluated the pilot through to March 2025. Its diary study collected 300 consenting diaries, a 32% response rate, recording each task, how long it took with Copilot, and an estimate of how long it would have taken without. Satisfaction was high, with 72% satisfied or very satisfied and a Net Promoter Score of 31.

The method is the point. A diary entry that saved an hour drafting a document counted as an hour. An entry where the output was not used, or the task existed only because the tool did, counted as negative time. That one adjustment moved several results a long way, as the department's own Table 5 shows, and it is a habit worth copying before reading the DILR approach to placing AI.

DBT then ran observed task sessions with users and a control group of colleagues without licences. Copilot users summarised reports and wrote emails faster and to a higher quality than non-users. But they completed Excel data analysis more slowly and to a worse quality, and they produced PowerPoint slides over seven minutes faster but to a worse quality. Its conclusion on productivity is blunt: "We did not find robust evidence to suggest that time savings are leading to improved productivity." The report adds straight away that productivity was not a key aim and that limited data was collected on it, so the finding is an absence of evidence, not evidence of absence, as the DBT evaluation report states.

The same governance logic underpins our AI execution office, the embedded delivery engagement in which the production placements are owned by the client, with a named owner per ship.

Which tasks justify a Microsoft 365 Copilot licence, and which do not?

Across both UK government reports, Microsoft 365 Copilot earned its keep on written work: drafting, summarising research, summarising meetings and searching existing information. It struggled with complex data, nuanced judgement and context-heavy material, and in the DBT diary study it added time to scheduling and image generation. A licence pays where a role carries a heavy writing and administrative load and little complex analysis.

DBT reached the same conclusion from its interviews: roles with heavy administrative burdens and little complex data anecdotally found Copilot extremely useful, while roles working with complex data and information found it less so. GDS heard the same thing from its focus groups. A finance participant said Copilot "struggles with complex data requiring contextual input", and an HR participant raised concern about accuracy in grievance handling and performance evaluations, where an error carries reputational risk.

Accuracy is the second filter. In the DBT diary study, 22% of respondents identified hallucinations during the pilot, 43% did not, 11% were unsure and 22% chose not to answer. DBT's conclusion is that outputs are mostly accurate but that quality assurance should review each output. That turns the licence question into a workflow question. A task where checking the output takes nearly as long as producing it, or where an unchecked error is costly, is a poor fit regardless of what the user feels it saved.

Both reports also found the benefits were uneven across people in ways a uniform rollout misses. DBT found neurodiverse colleagues statistically significantly more satisfied than other users, and non-native English speakers anecdotally reported significant benefits. GDS heard similar accessibility feedback. That is a reason to allocate licences by role and need, not by grade or by who volunteered first, and the same logic runs through our six named enterprise AI solutions, which include enterprise knowledge retrieval and post-call follow-up automation.

What should an enterprise measure in its first 90 days of a Copilot rollout?

An enterprise rolling out Microsoft 365 Copilot should measure four things in its first 90 days: working-day activity by application, adjusted time saved per task type, output quality against a control group, and where the saved time went. The UK government trials show that monthly adoption and self-reported savings, measured alone, will look positive whether or not the rollout is working.

Working-day activity by application. The GDS definition of an active user is one interaction in 30 days. Report it, but put the working-day rate beside it, split by Word, Excel, PowerPoint, Outlook and Teams, because that is where the GDS trial found adoption diverging. A fall in Word or Outlook use from its peak is an early warning that novelty has worn off.

Adjusted time saved per task type. Copy the DBT method. Ask users to log tasks for one week, record the time with Copilot and their estimate without it, and count unused outputs and tasks done only because the tool exists as negative time. The adjusted column, not the raw one, belongs in a business case.

Output quality against a control group. DBT's observed tasks are the only independently scored quality evidence in either report, and they reversed the self-reported result for Excel analysis, though DBT notes they had a very small sample and covered only four tasks. A handful of scored tasks per role, with colleagues who do not hold licences as a comparison, is cheap and catches the tasks where faster means worse.

Where the saved time went. GDS could not answer this, and DBT's control group had not observed productivity improvements from colleagues in the pilot. If no manager can name what their team now does with the hours, the hours have not been banked. Set the target against your own baseline before launch rather than borrowing either report's figure.

These four measures fit inside the AI lifecycle that an operating model engagement designs. Where your AI programme sits inside a central team, our comparison of an AI operating model and a centre of excellence covers who should own the measures once the pilot ends.

What data and governance work has to happen before Copilot licences go out?

Before Microsoft 365 Copilot licences go out, an enterprise needs to fix file permissions, decide the internet setting, write acceptable use guidance people actually read, and plan training and support by team. GDS warned that Copilot retrieves whatever a user can access, so poor permissions surface as oversharing, and DBT found colleagues outside the pilot knew little about the tool's limits or security.

GDS describes Copilot as adopting the permissions of the end user and retrieving only documents that user could normally reach, and then warns that this "may highlight issues when users have access to files they shouldn’t". Most organisations in the trial switched off internet search so that Copilot relied only on internal data sources. For a private enterprise the same point lands harder: years of broad SharePoint and OneDrive sharing become instantly searchable. Fixing access before launch is a data protection task under UK GDPR as much as a technology one, and it is the kind of perimeter question our guide to self-hosted AI for regulated industries treats in depth for models a firm runs itself.

Training is the other half, and the evidence on it is specific. GDS found that "training was essential" and that driving adoption "required substantial effort", with user support left to individual departments and experience inconsistent as a result. DBT found that self-led training was the stronger and more consistent link to higher satisfaction; among departmental training groups, only those with at least five hours were significantly more satisfied. That argues for short, task-specific material people can use on their own work, which matches the bite-sized training videos the trial introduced after users said their work capacity limited participation.

Public sector buyers have a framework to anchor this. The AI Playbook for the UK Government, published by DSIT and GDS on 10 February 2025, sets out 10 principles civil servants should uphold when using AI. Private firms will look instead to their own regulators: the FCA and the PRA for financial services, and the ICO for any processing of personal data. The governance artefacts are the same either way: who owns the tool, who reviews outputs in high-risk tasks, and who can switch a feature off. If you want a second opinion on that list, talk to our team.

What is the best way to roll out Microsoft 365 Copilot in 2026?

The best way to roll out Microsoft 365 Copilot in 2026 is to place licences by role and task rather than by grade, measure adjusted savings and quality before extending, and fix permissions first. A large integrator such as Accenture, Deloitte or KPMG may suit an estate needing tenant configuration at tens of thousands of seats; a placement-first approach suits a firm unsure where the value sits.

That verdict comes with an honest concession. If your organisation has already decided to buy Copilot at scale and the remaining work is configuration, licensing and change management across a very large Microsoft estate, a Microsoft-focused integrator is likely the better fit. The question DATS is built to answer is earlier: the placement diagnostic produces a ranked roadmap of where AI belongs and where it does not, and the operating model covers governance, RACI and lifecycle once a tool is live.

Here is our own reading of two rollout approaches. The rows are our recommendations, informed by what the two UK government reports measured, not findings either report states.

Neither column is the vendor's fault or credit. GDS found that change management and the October licence roll-out lifted adoption to 83%, and DBT found that careful measurement shrank the savings to something smaller but real. A rollout that does both will know far more after 90 days than one that only buys licences. For the stages that come after a pilot, see the DILR approach to placing AI inside enterprise systems.

Frequently asked questions

Is Microsoft 365 Copilot worth it for a UK enterprise?

Microsoft 365 Copilot is worth it for a UK enterprise where roles spend much of their day drafting, summarising and searching documents, and where outputs can be checked quickly. The UK government trials found high satisfaction and real savings on written tasks but no robust productivity evidence, so the value case should rest on measured, adjusted savings in named roles, not on an average.

Whether the licence cost clears that bar depends on your own baseline, which is why both reports are better used as a measurement template than as a return figure. Our enterprise AI consulting guide covers how to build that case, and the wider strategy category holds the related operating-model posts.

Is Microsoft 365 Copilot the same as GitHub Copilot?

Microsoft 365 Copilot and GitHub Copilot are different products that share a brand. Microsoft 365 Copilot is the assistant inside Word, Excel, PowerPoint, Outlook and Teams that the UK government trials evaluated. GitHub Copilot is a coding assistant and agent for software developers. Evidence from one says little about the other, and the government trials did not evaluate GitHub Copilot at all.

For the coding side, see our coding-agent evidence grading, and the About DILR.AI page explains who builds and runs these engagements.

Where to go next

Want to test this in your own estate? See how DATS places enterprise AI, read about our enterprise AI solutions, review the AI approval paper guide, or check mid-market AI readiness before you buy more licences.

Written by the Dilr.ai engineering team, practitioners who ship enterprise AI in production. Follow us on LinkedIn for shipping notes, or subscribe via the RSS feed.