The seats are bought. Somebody senior wants to know whether it is working, and there is a report with a percentage on it. The useful question is not what the percentage is. It is what it counts, and when it starts being true.
Copilot adoption reporting is narrower and later than almost everyone expects, and the gap between what the number means and what a board thinks it means is where most of the arguments happen. This article is about closing that gap. If you have not bought yet, the question of whether to is a different one and we have answered it separately.
The data is readable at three weeks and worth judging at six to eight
Start with the clock, because it governs everything else.
- Activity for a given day typically appears in the usage report within 48 hours of the end of that day in UTC.
- The Copilot Dashboard always shows the previous 28 days, with a delay of up to six days from today. It refreshes daily.
- A newly licensed person carries up to seven further days before they appear in the analysis at all.
Stack those and the first genuinely readable cohort view of a new rollout is somewhere around three weeks after the licences land. Anyone presenting an adoption percentage in week one is presenting a number about almost nobody.
Readable is not the same as worth reporting. In our judgement, the first adoption read worth putting in front of a steering committee comes six to eight weeks after licences are assigned, not three. The first month mixes curiosity with habit: people trying Copilot to see what it does, alongside people who have started using it for a real, recurring task. An activity count cannot tell those apart. By weeks six to eight the curiosity has mostly burned off, and what remains is the work people have actually folded it into, such as meeting recaps and pulling several documents into one summary.
Worth knowing, because it will trip someone up in a meeting: Microsoft’s own pages disagree on the detail. The usage report page says data typically arrives within 48 hours and offers 7, 28, 90 or 180 day windows. The Copilot Dashboard FAQ says the admin centre report is produced within 72 hours and offers 7, 30, 90 or 180 days. Both are current. If two people pull “the same” report and get different numbers, that is one of the reasons.
“Active user” is a lower bar than your board thinks
A user counts as active when they perform one intentional action with an AI capability, on that day, and Microsoft documents what qualifies. Microsoft’s own example is precise: selecting the Copilot icon in the Word ribbon to open the pane does not count. Submitting a prompt in that pane does. One prompt, and that person is an active user for the day.
Two details make the bar lower still.
Since December 2025, a user who merely views a Loop document generated by the Facilitator feature in a Teams meeting is counted as active usage, both for Loop and in the all-up Copilot figure. Someone reading a meeting summary they did not ask for registers as Copilot adoption.
And prompts are not counted the same way for every feature. “Edit with Word” counts towards prompts submitted in Copilot Chat. “Edit with Excel” and “Edit with PowerPoint” do not. So the Copilot Chat prompt total includes editing done through Word but not the same work done through Excel or PowerPoint, and any comparison built on it measures the counting rules as much as the behaviour.
None of this is Microsoft hiding something. It is all documented. But an active-user rate means the share of licensed people who did one countable thing in the window, which is a different sentence from the one usually spoken out loud.
Check the denominator before you argue about the percentage
The bottom of the fraction moves too, in ways nobody announces.
In the admin centre usage report, the user-level table lists every user who held a Copilot licence at any point in the previous 180 days, including people whose licence was later removed, and people who never used it once. The Copilot Dashboard works the other way: someone who holds only a Copilot licence drops out of every analysis, for every period, once that licence is removed. And within the dashboard, the tenant-level counts of licences assigned and active users include disabled mailboxes, while the group-level counts do not. So the same organisation can produce two different adoption rates on the same day, honestly, depending on which view someone opened.
Before anyone defends or attacks a percentage, agree three things in writing: which view it came from, whether removed licences are in it, and what window it covers.
There is a third denominator problem, and it is the largest. A licence that was bought but never assigned does not appear in an adoption rate at all, because the rate counts people who hold a licence, not licences bought. A rollout can show a respectable active-user rate while a large share of what was paid for sits on the shelf.
In stalled rollouts, the licences were often never handed out
When we are brought in to audit or rescue a stalled Copilot rollout, we count licences purchased against licences assigned, where assigned means the licence is attached to a named user account, whether or not anyone uses it. Of the 18 stalled rollouts we were brought in to audit or rescue, seven still had more than 40% of the licences they had bought sitting unassigned 90 days or more after the order completed. Across all 18, 58% of the seats bought had been assigned at the 90-day mark. The median estate had assigned 67%, and the seven that had stalled averaged 29%, which is what pulls the overall figure down.
These are not a representative sample of Copilot rollouts. They are the 18 that were stuck badly enough for someone to call us, so read the figures as a description of what a stalled rollout looks like rather than a forecast for yours.
In those rollouts we usually found the same sequence. The licences had been bought in bulk, often to secure a discount, and deployment was then held back by one of three things: data security concerns nobody had resolved, a lack of training, or no decision about which users should go first. Each of those can be settled before the order, and each is cheaper to settle then than after it.
The assisted-value number is a model, not a measurement
The Copilot Dashboard will show a value figure, and it is persuasive because it is in currency. It is also a calculation rather than an observation: assisted hours multiplied by an hourly rate that defaults to $72, taken from US Bureau of Labor Statistics data.
If your people do not cost $72 an hour, the figure is wrong by exactly that ratio before anything else is considered. Use it as a direction of travel if you like. Do not put it in a board pack without replacing the rate with your own, and saying that you did.
Below fifty licences, the dashboard will not give you the comparison
A detail that reshapes pilots. Dashboard features are gated on licence count. With 1 to 49 Copilot licences you still get the readiness page, and adoption and impact metrics at tenant and group level, with HR filters. What you do not get: agent insights and the Agent Dashboard, benchmarks, intelligent summaries, survey sentiment, week- and month-level trendlines, the group view for managers and delegation, the combined view across all licence types, and the user-level data export. The dashboard’s own licence table says fifty or more assigned, paid Viva Insights licences unlock the full set whatever the Copilot count. Microsoft’s Agent Dashboard page disagrees on that one feature: it asks for at least 50 assigned Copilot licences and some agent activity.
Group-level figures also carry a privacy floor. A group smaller than the minimum group size, ten people by default and five at the lowest setting, shows no figures at all. In a twenty-person pilot, most team-level cuts fall below it.
So a twenty-person pilot will not get its comparative evidence from the dashboard. That is not the same as having no evidence, and it is worth knowing before the pilot is designed rather than after it has been declared inconclusive. On pilots of that size we leave the dashboard aside and build the evidence directly. Each week we pull the per-user Copilot usage report through Microsoft Graph, which gives each person’s prompts submitted, their active days, and the last date they used Copilot in each app. That shows how often, not in which apps; where a pilot needs prompts per app, the Purview audit log is the source, as we explain below. We pair it with a five-question pulse survey sent to the pilot group, asking about time saved and the quality of what Copilot produced. The usage data shows who is using it and how often. The survey supplies what the dashboard’s sentiment metrics would have, which are among the features switched off below fifty. Self-reported time saved is a claim rather than a measurement and should be presented as one, but a weekly series of both, from the same twenty people, is evidence a board can actually interrogate.
Two things nobody mentions at purchase. By default, every Microsoft 365 usage report, including the Graph downloads, replaces user names with anonymised identifiers; reading the report person by person needs an administrator to clear “Conceal user, group, and site names in all reports” under Settings, Org settings, Services, Reports, and that change is logged. And assigning a Copilot licence automatically assigns a Viva Insights service plan, so buying Copilot quietly puts those employees into a measured population. Neither is sinister, and both are exactly the kind of thing a works council or a privacy lead would rather hear from you in advance than discover in a report.
What to do before the data means anything
The measurement gap is not dead time, and treating it as dead time is the commonest adoption failure we see. It usually looks like this. Someone checks the dashboard around day ten, finds flat or empty charts, and enablement stops while everyone waits to see what the numbers say. Meanwhile, people who get a poor first answer, a summary that is wrong or a draft that is no use, do not raise a ticket. They quietly go back to searching and drafting the way they did before. By the time the dashboard fills in, around a month after the licences land, it is confirming a drop-off that happened a fortnight or more earlier, and the window in which someone could have caught it has closed.
The work in that window is the work that decides what the data will eventually say.
- Confirm the licences are actually assigned, not merely purchased. The gap between bought and assigned is real and invisible in the invoice.
- Pick the three tasks each role will try first, in the applications they already live in. Breadth at the start produces no habit anywhere.
- Run office hours and prompt reviews from week one, because the first weeks are when people quietly give up, and the dashboard cannot show it yet.
- Fix the things that will make it look broken, which is pre-purchase work arriving late if it was skipped: the readiness checks on tenant configuration, oversharing and network inspection.
- Agree the measure now, with its denominator, so the first readout is a result rather than the opening of a methodology argument.
Our position: pick one number and defend it, rather than reporting all of them
In our view the thing to decide before any of these reports is read is which single number the organisation will be judged on, and then to publish how it is calculated alongside it, every time.
The reason is practical rather than philosophical. These dashboards offer several measures that disagree with each other by design, and in the absence of an agreed one, the number that gets quoted is whichever is most convenient to whoever is presenting. Microsoft’s active-user figure counts anyone who did one qualifying thing in the window, so a licence holder who asked Copilot Chat a single question in a month is active. That is how a programme ends up with a high adoption claim that collapses the moment finance asks what it did for a department. We would rather publish a lower number that survives scrutiny than a higher one that collapses the first time somebody asks what it counts.
The stricter measure we prefer is what we call a weekly habitual user: someone who submits at least three prompts a week across two or more of the core Microsoft 365 apps. It always produces a lower number than the active-user rate, and the lower number is the one worth defending. It is a habit test rather than a break-even test: it shows that people have folded Copilot into their week, not that a seat has paid for itself.
One caution if you adopt it: Microsoft’s usage report cannot calculate it. Per person, the report gives a total prompt count for the period and the last date each app was used, not prompts per app per week. Per-app counts by person exist in two places. The Copilot Dashboard’s user-level export has them, but it is in preview and needs at least 50 Copilot or Viva Insights licences. The Purview audit log records every Copilot interaction with the user, the time and the app that hosted it, and keeps those records for 180 days. Auditing is on by default in enterprise tenants but not on Business Basic, Business Standard or Business Premium, where an administrator has to turn it on first, so check it before the pilot starts. Microsoft says plainly that audit data is not intended as the basis for usage reporting and that counts built from it may not match the official reports, so a measure built that way is your own calculation, and should be labelled as one.
The corollary, and it is the same argument we make about Power Platform adoption: a usage percentage is not an outcome. It is the cheapest available proxy for one, and it is worth exactly as much as the definition attached to it.
Where to start
- Write down the measure, its view, its window and whether removed licences are included.
- Count licences assigned against licences purchased, and the people who have never submitted a prompt.
- Replace the $72 hourly rate with your own before anyone sees an assisted-value figure.
- Start enablement in week one rather than waiting on the reports: choose each role’s first tasks before launch, run office hours and prompt reviews from the first week, then read the data at three weeks and judge it at six to eight.
Veratas runs Copilot adoption programmes and the measurement that makes them defensible. If you have seats live and a number nobody trusts, talk to our team.
Frequently asked questions
How soon can we measure Copilot adoption? The data is readable at about three weeks: daily activity appears within roughly 48 hours, the Copilot Dashboard shows the previous 28 days with up to six days’ delay, and newly licensed people take up to seven further days to appear. In our judgement the first read worth showing a steering committee comes six to eight weeks after licences are assigned, because the first month mixes curiosity with habit.
What counts as an active Copilot user? One intentional action in the day. Opening the Copilot pane does not count; submitting a prompt does. Since December 2025, viewing a Loop document generated by the Facilitator in a Teams meeting also counts.
Why do two Copilot reports show different numbers? Usually the denominator, and the two reports count it differently. The admin centre usage report’s user-level table lists everyone licensed for Copilot at any point in the past 180 days, including people whose licence was later removed and people who never used it. The Copilot Dashboard drops anyone who held only a Copilot licence once that licence is removed, and its tenant-level counts include disabled mailboxes where group-level counts do not. Microsoft’s own pages also differ on refresh timing and available windows.
Why does our adoption rate look healthy when licences are unused? Because a licence that was never assigned is not in the rate at all: the rate counts people who hold a licence, not licences bought. Check assigned against purchased separately. In the 18 stalled rollouts we were brought in to audit or rescue, seven still had more than 40% of purchased licences unassigned 90 days or more after the order.
Is the Copilot assisted value figure real money? It is a model: assisted hours multiplied by an hourly rate that defaults to $72 from US Bureau of Labor Statistics data. Replace the rate with your own before using it, and say that you have.
Can a twenty-person pilot produce evidence a board will accept? Not from the dashboard, unless the tenant also has 50 or more Viva Insights licences. With fewer than 50 Copilot licences, benchmarks, week- and month-level trendlines, survey sentiment and the user-level export are switched off, and groups under the privacy minimum show nothing. Pair the per-user usage report from Microsoft Graph with a short weekly pulse survey of the pilot group instead, and present self-reported time saved as a claim rather than a measurement.
Does buying Copilot change anything else in the tenant? Yes. Assigning a Copilot licence automatically assigns a Viva Insights service plan, which places that employee into the measured population for Copilot reporting.
Operations head with 20 years of experience delivering business intelligence, ERP and AI programmes across healthcare, finance and manufacturing. Focus areas include Power Platform adoption and governance, delivery models, and managed services.