The finance director of a provincial government stops by the administrator's desk. Six hundred Copilot licences were bought last year; how many are being used? The administrator opens the admin center and has a number within a minute: in the past thirty days, just over a third of licensed users have used Copilot in at least one app. The second question follows straight away: and what is everyone else doing?
The report cannot answer that second question. So you measure Copilot adoption in two layers: what Microsoft shows about the licences, and what AI use is taking place in the organisation, with Copilot as one of the tools. You already have the first layer. The second you have to set up, and you can do that without tracking a single person.
What does the Microsoft 365 admin center show?
The admin center has usage reports for Microsoft 365 Copilot. They show, per period, how many users hold a licence, how many of them used Copilot and in which apps: Teams, Word, Excel, PowerPoint, Outlook and Copilot Chat. There is a per-user detail view with the last activity per app; whether it shows names or concealed values depends on a setting in the admin center. Microsoft rearranges these reports regularly, so the exact columns you see can shift.
That is enough for three questions. How many of the licences were used in a period, and what does the unused part cost? Which apps account for the use, and has Copilot caught on in Teams but not in Excel? And does use rise after a training session, only to slide back afterwards? Those are licence questions, and the report is good at them.
What does it not show?
Everything that is not Copilot. The report covers one service in your own tenant, and by definition can say nothing about ChatGPT, Perplexity, Gamma or the transcription tool a department found on its own. The consumer version of Copilot, opened without a work account, also falls outside it, because it does not belong to the tenant.
On top of that, "active" only tells you that someone used Copilot during the period. Not what for, not whether it helped and not whether that same person opened another tool for the same work an hour later. So the two thirds in the opening scene who did not use Copilot are not necessarily people who did not use AI. The report simply does not know.
| Question | Microsoft usage report | Measurement of all AI use |
|---|---|---|
| How many licences were used? | Yes | No |
| In which Microsoft app? | Yes | No, only the domain |
| Which AI tools are in use alongside Copilot? | No | Yes |
| Is the consumer version of Copilot being used? | No | Yes, as a separate row |
| What share of AI use runs through allowed tools? | No | Yes |
| Who was it? | Per user, unless concealed | Never |
Why are an unused licence and an unknown tool the same problem?
Back to the province. Set the measurement alongside the report, and it shows that ChatGPT in the browser is opened on more devices than Copilot, that Perplexity is opened every day in the legal team and that a meeting transcription tool nobody has ever decided on is in use. A good part of the sixty-odd per cent who did not use Copilot turn out to be using AI after all, just somewhere else.
That is one finding, seen from two sides. Look at the licences and you see under-use. Look at the tools and you see use without a decision. Both tell you that the preferred tool is not the one people actually use. Why they pick other tools alongside Copilot is the next question. Another round of Copilot training will not fix that if Perplexity does something for the lawyers that Copilot does not.
How do you put the two side by side without tracking people?
By settling three things in advance. First: the measurement of AI use counts per tool and per day, without a user ID. "This domain was opened on this device today" is enough to know how many tools there are and how use shifts, and too little to point at anyone. Which figures that leaves you with, and which you can leave alone, is covered in measuring AI use.
Second: you use the Microsoft report at organisation level, with names concealed. The licence question needs no names, and a per-person overview of who used Copilot when quickly turns into employee monitoring, which brings requirements on purpose and necessity, and a role for the works council. The difference between those two kinds of measurement is explained in AI monitoring versus employee monitoring.
Third: put the purpose on paper before you measure. "Knowing whether the Copilot licences are used and whether Copilot has become the preferred tool for AI use" is a purpose the figures fit. "Seeing who is not following the policy" is not, and the figures could not do that anyway.
BeeSensible provides the second layer. The extension notices when someone opens an AI tool and reports only the domain, at most once a day per device, without a user ID or any content. Copilot appears as an ordinary row, and the consumer version and Microsoft 365 Copilot are separate catalogue entries, each with its own risk score. The Compliance screen shows the share of AI use that goes through allowed tools, how that share is trending and which tools still have no decision. Nowhere in the dashboard is there a breakdown per employee. The product page covers what else the module does.
