Copilot Adoption: The First 90 Days After the Licences Arrive

Part of our topic guide on AI Skills for Business.

By James Cotton, Founder, iO-Sphere

What you bought, and what came with it

Microsoft 365 Copilot, renamed Microsoft Copilot in 2026, is the paid licence; its free tier, Microsoft 365 Copilot Chat, is now Microsoft Copilot Chat. The UK price pages still say Microsoft 365 Copilot for the paid licence.

Microsoft's UK enterprise pricing page shows the add-on at a list price of £23.10 per user per month, paid yearly, excluding VAT, on an annual subscription that auto renews and on top of a qualifying Microsoft 365 licence; the enterprise solutions page also shows £24.26 paid monthly, still on an annual commitment.

Copilot Chat comes at no additional cost to Microsoft Entra account users on an eligible Microsoft 365 subscription, and Microsoft's comparison table gives it the paid licence's web-grounded chat, enterprise data protection, file uploads and image generation, with some organisational grounding in Outlook. The licence adds broad reasoning across the organisation's own Microsoft 365 data, the Graph API, Work IQ and semantic indexing, full agent experiences and priority access.

The licence also brings SharePoint Advanced Management. Microsoft's requirements page says it comes with the paid licence, and Microsoft's UK price tables show it included with every paid option on them and not with Copilot Chat.

Days one to thirty: who can see what

Copilot shows each person only the organisational data they can already open. Microsoft's privacy page leaves it to the customer's own permissions to give the right people access to the right content. A folder once shared too widely is one prompt away for everyone it reached.

The Government Digital Service (GDS) saw the risk in its cross-government experiment. Its findings report, published on 2 June 2025, noted that searching with a user's own access "may highlight issues when users have access to files they shouldn't."

Microsoft's data-foundation blueprint for Copilot has three pillars: remediate oversharing, set up guardrails and meet regulations. Microsoft's requirements page calls configuring this foundation not required but strongly recommended.

Oversharing comes first, which is why, with the licences bought, this plan assigns no seat until the data work is done. Seats already out stay where they are; no new seat goes out before day thirty, and the usage report that day becomes the baseline for the seats already out.

The configuration guide starts with the SharePoint Advanced Management Content Management Assessment, which flags sites with oversized audiences, broken inheritance or inappropriate sharing, and sites that are inactive or ownerless. Restricted Content Discovery and Purview Data Loss Prevention for Copilot then keep sensitive sites and content out of Copilot's reach while site owners review access, down to the file.

Microsoft says to remove those interim protections once permissions are fixed. The minimum licence for the work is the foundational Purview capability in Microsoft 365 E3.

The older control, Restricted SharePoint Search, cannot be newly enabled since 31 July 2026; Microsoft calls it a short-term measure limited to 100 sites, says it is not a security boundary, and points to Restricted Content Discovery instead. Which data staff may put into any AI tool is a policy decision, not a permissions setting, and sits in a staff AI use policy rather than in SharePoint.

Days one to thirty: where the seats go

Microsoft's Copilot Success Kit gathers its rollout guides, among them a User Enablement Guide, a scenario library and a licence allocation guide. Its implementation framework runs in four stages (get ready, onboard and engage, deliver impact, extend and optimise) across human change and technical readiness.

Microsoft names leadership, human change and technical readiness as the essentials, and for leadership asks for an executive sponsor who takes an active, visible part, builds a coalition with executive peers, speaks directly to staff and uses Copilot themselves.

On seats, the "Get ready" stage says to select initial departments with the Scenario Library, target a critical mass of users, and "be intentional with assignment and concentrate seats".

Microsoft's licence allocation guidance, a deck last saved in May 2025 and linked from the kit page updated 6 August 2026, calls the method of assigning licences one of the primary reasons customers struggle to show business value. It names four inefficient approaches: leadership first, a request and approval workflow, people tied to the pilot, and a use-it-or-lose-it rule, which it calls "optimized for usage, not value".

Its alternative is to anchor on one high-value function, the roles its Copilot uses need, and the people in those roles who already work in Microsoft 365 apps every day, using Copilot Chat usage data to find who is active already. Licensing many people in one group, Microsoft says, raises the chance that skills pass between them.

The User Enablement Guide's 12-week milestone view has three bands: the first 30 days, days 30 to 60, and recurring tasks. Its first band opens with the purchase decision, which is already behind you. It puts assembling the team around weeks one to two, and selecting the first cohort and launching training content and office hours around week four.

"Install apps, assign licenses" sits around weeks four to five, in the same box as "Ensure data security controls are in place". This plan follows the blueprint's order, so the seats chosen here go out once the data work is done.

Days thirty to sixty: what counts as use

The admin centre's Copilot usage report counts enabled users, who hold a licence, and active users, who tried a Copilot feature they started themselves, over 7, 28, 90 or 180 days. A day's activity typically appears within 48 hours. Opening the Copilot pane in Word does not count; submitting a prompt does.

Its user table lists everyone licensed in the past 180 days, including people who never used Copilot, with prompts submitted, active days and last activity date; names are hidden by default and can be shown. Microsoft says the audit log is not meant for usage reporting.

The Copilot Dashboard in Viva Insights needs no paid Viva Insights or Copilot licence to view and covers the previous 28 days, with up to six days' delay. Processing takes up to seven days after a Copilot licence with the Viva Insights service plan is assigned. Below 50 licences it loses benchmarks, trendlines, the managers' group view and survey-based sentiment.

It counts a user as active after one Copilot activity in 28 days, and hides metrics for any group smaller than the minimum size the organisation sets.

Its 12-week view is the one to watch. Power users average 15 or more Copilot actions a week and used it in at least 9 of the 12 weeks; habitual users average 1 to 14 a week, also in at least 9; novices used it in fewer than 9.

"Copilot assisted hours" is Microsoft's estimate, not a measurement: it credits 6 minutes per search, summary or creation action, from two Microsoft studies of 163 and 147 knowledge workers published in November 2023.

"Copilot assisted value" multiplies those hours by a default $72 an hour, a US Bureau of Labor Statistics rate in dollars. Microsoft says Copilot is unlikely to be solely responsible for differences the dashboard shows, and a known issue under-reported assisted hours from 1 June 2025 to 18 February 2026.

Days thirty to ninety: what the government trials found about people

Four UK government evaluations covered Copilot deployments that began in late 2024, and were published between June 2025 and July 2026. Each defined use its own way, so each figure keeps its own definition and base.

The GDS experiment, which gave 20,000 staff in twelve government organisations a licence from 30 September to 31 December 2024, concluded that training was essential. Driving adoption took substantial effort, support fell to each department so experience varied, and some users lacked the capacity for the upskilling. GDS found a correlation between confidence with AI and time saved, and saw lower benefits where familiarity with AI was lowest.

Counting a user as active after one interaction in 30 days, on usage data for 14,500 users, GDS saw adoption stay around 80%. Its 7,115 survey respondents reported an average saving of 26 minutes a day, from the midpoints of answer ranges; 17% noticed no clear saving.

A press release from the Department for Science, Innovation and Technology on 2 June 2025, the day GDS published its report, said civil servants could save nearly two weeks a year, a figure averaged across all 20,000 people, a different base from the report's 7,115.

The Department for Business and Trade (DBT) found self-led training more effective than formal sessions: in its diary study of 300 diaries, at least 2 hours of independent training predicted satisfaction and recommendation at a 99% confidence level. Sentiment about formal sessions was generally positive, but departmental and Microsoft sessions were often seen as introductory or high-level and less relevant to people's roles.

DBT piloted 1,000 licences from October to December 2024, about 70% to volunteers, and published its evaluation on 28 August 2025. Counting weekly use, a mean of 64% of licence holders used Copilot at least once a week.

At the Department for Work and Pensions (DWP), 89% of the users who answered its survey explored Copilot on their own and 61% learned from peers; staff asked for targeted, role-specific training and quick-reference materials, and peer demonstrations played a key role in uptake. DWP gave Copilot to 3,549 staff from October 2024 to March 2025 and published on 29 January 2026.

Of 1,716 DWP users who answered its survey, 65% reported daily use, and a regression against 2,535 non-users estimated a saving of 19 minutes a day. DWP named its own limits: non-random allocation, self-reported measures, no pre-trial baseline.

HMRC found its users needed practical, tailored training and time to learn. In an evaluation published on 9 July 2026, it reported that 83% of the 3,000 staff randomly given a licence from September to December 2024 used it, and its 1,364 survey respondents reported saving 2% to 3% of the working week, about 60 minutes for a full-time worker, on a method HMRC called "deliberately cautious".

Microsoft's skills progression runs from the top ten prompts, such as summarising a meeting, to departmental scenarios, then to line-of-business systems and automating business processes, with peer-to-peer learning throughout. Its champions, launched in the "Onboard & engage" stage, are peers who supplement the help desk, motivated by helping others, and represent the business more than IT support.

What the tasks showed

DWP's largest self-reported reductions were in searching for information, 26 minutes, and email writing, 25. GDS respondents reported 24 minutes saved on drafting documents, 19 on presentations and 9 on scheduling meetings, and GDS observed limitations with complex, nuanced or data-heavy work and an ongoing need for human oversight.

DBT's diary study, 1,411 tasks adjusted for unused outputs and novel tasks, put drafting a report or briefing top at 1.3 hours saved per task, then summarising research at 0.8, and meeting summaries, searching and brainstorming at 0.7. Data analysis saved 0.6, emails 0.2 and slide decks nothing, while generating images and scheduling took longer, by 0.5 and 0.6 hours.

DBT also observed six Copilot users against five non-users. Copilot users did Excel data analysis more slowly (25:01 against 20:33) and less accurately (1.5 against 2.7 out of 5), but summarised reports faster (12:37 against 41:34) and more accurately (4 against 2.5).

They wrote emails slightly faster (07:30 against 07:43) and better (accuracy 4.3 against 4), and produced PowerPoint slides over 7 minutes faster (10:47 against 18:30) but less accurately (1.5 against 5). DBT called the samples small and the results supplementary. It found no robust evidence that time saved became productivity, which it had not made a key aim.

Your own first function can keep a short task diary through the quarter, recording the task, whether Copilot was used, the time taken and whether the output was used; DBT's diary study is the same shape of evidence.

Microsoft's own numbers

Microsoft's figures are vendor research, worldwide. It told investors on 29 July 2026 it had over 30 million paid Microsoft 365 Copilot seats, which counts seats bought, not people using them. Its 2026 Work Trend Index, a survey of AI users in ten markets, found 86% treat AI output as a starting point, and an analysis of over 100,000 Copilot chats classed 49% of conversations as cognitive work.

Week twelve: the habitual test

By week twelve the first function has a full 12-week record, and the habitual test applies: Copilot used in at least 9 of the 12 weeks, at an average of at least one action a week. Microsoft's milestone view analyses usage reports and Dashboard data around week ten.

Its recurring band, around weeks nine to twelve, holds a service health review, scenario and user recommendations and AI Council insights, and repeats each cycle.

The review decides which function goes next. If the first group is mostly habitual and the diary shows hours saved on named tasks, the case is made; if it is mostly novice, fix the training and the tasks before buying seats, and a skills gap analysis shows where the next function starts. Seating one function at a time is a different job from a whole-workforce AI programme.

Most people asking for AI skills want to get more of their own work done, so our view is to pick three recurring tasks from the actual week and measure each in hours saved, the start of a practice loop that repeats each quarter with three new tasks.

At iO-Sphere, team training on the team's own tasks runs beside a rollout, in the tools the team already has, with practitioners coaching, in closed cohorts where we look for 10 to 12 people, priced per engagement.

By the end of the quarter, the dashboard shows how many of the first function's people are habitual and how many are still novices, and the user table shows how many never submitted a prompt at all. The diary says which tasks saved hours, which saved nothing, and which went slower with Copilot than without it.

That points to the function that goes next, and to what its training covers: the recurring tasks its own people do every week, starting with the kinds the first function's diary showed saving time, practised on their own work in the weeks after their seats arrive.

Frequently asked questions

Does Copilot use our prompts to train its models?

Microsoft says prompts, responses and data reached through Microsoft Graph are not used to train foundation models, including those Copilot uses. Its privacy page adds that Copilot services have opted out of Azure OpenAI's abuse monitoring, which includes human review of content, and that generated responses are not guaranteed to be factual, so outputs still need checking.

Where is our Copilot data processed?

Organisational use of Copilot falls under Microsoft's Data Protection Addendum, with Microsoft as processor, as its enterprise data protection page sets out. Web search queries go to Bing without user and tenant identifiers, but there Microsoft acts as an independent controller and the EU Data Boundary does not apply. Microsoft's privacy page says customers outside the EU may have queries processed in the US, the EU or elsewhere.

Is there a short introduction Microsoft itself offers?

A 30-minute course "to get started with Copilot" is linked from Microsoft's UK enterprise page. It fits the week licences are assigned, as a first look before people move on to their own tasks.