At a consulting firm of about forty people, 85 % of the licence holders had used their AI assistant during the month: the activation rate a vendor report would have shown. The usage logs still left nearly one seat in three to decide on, and the three heaviest users consumed, at API rates, between 12 and 25 times the price of their seat. The Copilot usage report counts the same way: an employee who had a single email thread summarised during the month is as active there as someone who uses it all day.
What can you check today in the Copilot usage report?
Three readings can be done today, with what Microsoft already provides: the gap between active users over 7 and over 90 days, the share of licence holders still at the trial stage in Viva Insights, and a cross-check of licence holders against your other AI tools.
- Compare the active user rate over 7 days and over 90 days. Over 90 days, someone who tried Copilot in the first week and then stopped still counts. The gap between the two rates gives a first measure of drop-off.
- If you have access to the Copilot Dashboard in Viva Insights, look at the share of "Novice" users, who took at least one action in twelve weeks without regularity, and "Other" users, which also covers licences never opened: two groups whose usage remains to be shown.
- Cross the list of licence holders with those of your other AI tools. We often see the same need, summarising meeting minutes, handled in Copilot, in ChatGPT Enterprise and in the in-house tool, and paid for three times.
These readings point to the licences to examine. What the others return can only be read in the usage itself.
Assigned, active, used: what is the difference?
An assigned licence is paid for, an active licence was used at least once during the period, a used licence produces more than it costs. Microsoft's report counts the first two stages; the third can only be read in the usage itself.
| Stage | What it means | Where you read it |
|---|---|---|
| Assigned | the licence is paid for | the contract, the admin console |
| Active | it was used at least once during the period | the vendor's usage report |
| Used | what it produces exceeds what it costs | the usage logs, read with the teams |
And when the tool is billed per token?
With consumption-billed tools, coding assistants first among them, the seat count says nothing about cost, and a used licence becomes consumption that was worth what it cost. In the estates we analyse, 5 % of developers consume half the credits, and the real cost of the same tool varies from one person to another by a factor of one to several dozen. GitHub Copilot's mechanics, seat and credits, are detailed in our article on GitHub Copilot pricing.
The same calculation settles seat versus pay-as-you-go. For a heavy user, the seat is often much cheaper than the same volume billed per token; for an occasional user, the reverse holds. Our page on the Copilot licence details the cost mechanics specific to Microsoft.
How do you know whether your Copilot licences are really used?
By crossing three readings, from the simplest to the finest. The first two come from Microsoft and say who uses Copilot; the third says what the usage produces, and it is the only one that separates an active licence from a used one.
- The usage report in the Microsoft 365 admin center: who triggered a Copilot feature over 7, 28, 90 or 180 days, and in which app. That is the "active" stage.
- The Copilot Dashboard in Viva Insights, if it is enabled in your tenant: who has made a habit of it, and who stopped at trying it.
- The usage logs, read team by team and checked against interviews: which tasks the tool serves, what they cost and what they return.
How does Microsoft count an active Copilot user?
The usage report in the Microsoft 365 admin center counts as active any licence holder who triggered a Copilot feature themselves, in one or more apps, during the selected period: 7, 28, 90 or 180 days (Microsoft documentation). Opening the Copilot pane does not count; sending it a prompt does. The active user rate divides that number by the licences assigned.
The report also gives the number of prompts submitted, adoption by app and, for each holder, their active days in Copilot Chat and their last activity date in each app. It counts features used, a meeting recap or an email draft, without saying what work they served or what became of the output.
What does the Copilot Dashboard in Viva Insights show?
The Copilot Dashboard in Viva Insights sorts licence holders into four levels, based on their last twelve weeks.
| Level | Microsoft's definition |
|---|---|
| Power | 15 or more actions per week on average, and usage in at least 9 of 12 weeks |
| Habitual | 1 to 14 actions per week on average, and usage in at least 9 of 12 weeks |
| Novice | at least one action in 12 weeks, but usage in fewer than 9 of 12 weeks |
| Other | every other licence holder, including those who never used it |
Microsoft itself separates trying from habit.
And in ChatGPT Enterprise, Claude or Cursor?
Each vendor has its own definition of an active user, and they all come down to having done something with the tool during the period.
| Tool | What makes a user active |
|---|---|
| Microsoft 365 Copilot | at least one user-initiated action over 7, 28, 90 or 180 days |
| ChatGPT Enterprise | having started to use it (an "activated" seat), then being active during the week |
| Claude (Team, Enterprise) | in the Analytics API, having consumed tokens during the day |
| Cursor | having used at least one AI feature during the period |
Vendors add other indicators: messages per user, estimated time saved, task categories, lines of code accepted. Each computes them within its own perimeter, with its own factors, and none says whether a licence is worth its price for you, or whether the same task also runs on another tool. For ChatGPT, see also how to measure ChatGPT Enterprise adoption.
Why does a high active rate not prove usage?
Because a single action during the period is enough to count as active: the rate measures trial, not habit, and not what the usage produces. The narrower the window, the lower it falls, as public evaluations show.
What the time window changes
Two public UK evaluations measure usage over different windows. The Government Digital Service followed 20,000 civil servants with the vendor's definition, one interaction in 30 days, and adoption reached 83 %. At the Department for Business and Trade, in a 1,000-licence pilot, around 64 % of licence holders used Copilot at least once a week, and around 30 % at least once per working day. In Word, Teams or Outlook, daily usage fell to 10 to 15 % of licence holders.
How many licence holders account for most of the usage?
In the estates we analyse, 10 to 15 % of licence holders produce more than half of the usage that saves time: drafting long documents, analysing files, summarising meetings. The rest splits between occasional usage and licences that sit idle.
At the consulting firm from the opening, the seats to decide on, 31 % of the estate, were seats with no trace at all and aborted starts, at fewer than 1.4 prompts per working day: enough for the vendor to count them as active. One workstation had hit the usage cap more than three times per working day and needed a higher tier.
What the vendor report cannot tell you
The report says who used Copilot, and in which app. It says neither what work the tool served, nor what became of the output, nor what happens in the other AI tools.
"We have Copilot, and when we roll it out there's a platform that shows usage. We know who uses it and who doesn't, on the licences." A CIO told us this in a meeting at the end of September. Aurélien Sabaly, co-founder of Idun Group, answered: "Someone sends one email, once, and the licence counts as activated for that month." The report shows who used the tool; leadership then asks other questions:
- what the tool is used for, and which use cases nobody had spotted;
- whether the tasks handed to the AI get done or stop at the first answer;
- who to train, and on what, based on how each team uses it;
- whether what was deployed returns what it costs, and if not, what is blocking it.
Sometimes the most used tool in the company has no licence at all: that is shadow AI, which the vendor report does not see either.
How do you know whether a licence is worth its price?
Read what teams do with the tool in the usage logs the platform already produces, then check it against what they say in interviews. The logs show the tasks, their frequency, their cost and the abandoned ones; the interviews explain the process the usage serves and what holds it back. A month of logs is not always enough: at the consulting firm, some accounts with no conversation during the month had been used before.
The reading covers usage grouped by team, and employees know before the measurement what is looked at and what is not. Our audit of real AI usage does this work on Copilot and on the other tools in the estate, use case by use case.
Before you renew: how many licences should you keep?
As many as actual usage justifies, licence holder by licence holder, and the decision is made before the contract renews.
"We renew by default: nobody dares to take a licence away." CIOs say it because the vendor report gives them nothing to decide with. With real usage, each licence holder falls into one case: they hit the caps and need a higher tier; they use it enough to keep theirs; they could move to a lighter plan, such as Copilot Chat, included in the Microsoft 365 licence, or to pay-as-you-go; or their seat can be reassigned.
The first decision costs nothing: give the idle premium seat to the person who keeps hitting the cap. It has to be made before the contract renews, otherwise it waits for the next renewal.
That leaves how licences are assigned. In many organisations they are handed out from a list, by population or first come, first served, and seats sit idle while IT is swamped with requests. A large French transport group does it differently: its executive committee opens new licences on condition that actual usage is read in the logs, with a written rule that takes a licence back after a month without use.
To place these licences within the whole estate, see our guide to AI in the enterprise.
