Licences are not adoption. What the Copilot numbers actually show.

Chapter 01

It turns up in a board paper, a conference slide and a LinkedIn post in the same week. Ninety-two of every hundred Copilot licences go unused. It is a wonderful number: short, damning and exactly what a sceptical finance director already suspects. It is also unsupported by any research anybody can produce. If it is in your deck, take it out before somebody else checks.

There is a real published figure sitting close to it. Recon Analytics, in its AI Choice 2026 study of more than 150,000 respondents, found that where workers have Microsoft 365 Copilot, ChatGPT and Gemini all available to them, 8 percent name Copilot as their primary tool. That is a preference share among people choosing between three assistants. It says nothing about how many licences are used, how often or by whom. Subtract it from one hundred and you have the number in the board paper. A preference share was inverted, relabelled as an activation rate and then repeated until it sounded like research.

No research supports the claim that most Copilot licences go unused, and no published evidence supports a single-digit activation rate at all. You might reasonably say the direction of travel is right even if the figure is invented. That works once. The difficulty arrives in the second meeting, when somebody with a search engine asks for the source. You do not lose the number at that point. You lose the argument, and the argument was sound.

Chapter 02

There is real data, and it is more useful than the invented figure precisely because it is less dramatic. A public body ran the experiment at scale, counted properly and published the result whether it flattered anyone or not.

The UK Department for Business and Trade ran a Copilot pilot across 1,000 licences from October to December 2024 and published the evaluation in 2025. Mean weekly active use was 64 percent. Mean daily use was 21 percent, which works out at about 30 percent once weekends and holidays are excluded. Users performed a mean of 1.14 Copilot actions per person per working day.

Read those together and you get a picture neither camp enjoys. Roughly two thirds of licence holders touched the tool in a given week, so it is not shelfware. Fewer than a third used it on a given working day, and when they did they used it about once. That is a tool people reach for occasionally, not one that has reorganised anybody’s morning.

The second number comes from the Forrester Total Economic Impact study of Microsoft 365 Copilot, published in March 2025 and commissioned by Microsoft, which is a label that belongs on every citation of it. Over three years its composite organisation spends $7.18m on licences, $4.92m on implementation and management and $8.54m on training and employee discovery. Roughly two thirds of the three-year cost is not the licence. Treat the absolute figures as illustrative, because the model is vendor-commissioned. The shape is the part that matters.

Chapter 03

Take the invented statistic out and the case against blanket licensing is still standing, on evidence you can put in front of anyone. Buying licences is not a strategy, because the licence is the smallest and the easiest part of the thing that produces the result.

For every euro of licence in the Forrester model there are roughly two euros of implementation, management, training and the slow business of people working out what the tool does in their own job. Skip that and you have not bought a third of a result. You have bought the part that does nothing on its own, which is also the part that renews automatically every year.

A mean of 1.14 actions per user per working day is not the signature of a bad product. It is the signature of a product that was handed out rather than embedded. A government department ran a defined pilot with an evaluation attached and still landed there, which is worth holding in mind when you estimate what a rollout with no evaluation behind it is producing in your own estate.

The useful question is not how many licences sit unused. It is which people, doing which recurring task, produce something you would miss if the tool disappeared tomorrow. That question has answers, and they are available inside your own organisation within a fortnight, without a statistic from anywhere. An hour with the financial controller and an hour with the director of sales will produce most of the list.

Chapter 04

Everything above is general. The hotel version is sharper, because the shape of a hotel workforce is nothing like the shape of the workforce these products were designed around, and blanket licensing exposes the difference inside the first quarter.

In a bank or a software company most of the payroll sits at a desk with a mailbox. In a hotel group it does not. Housekeeping, kitchen, stewarding, maintenance, spa, most of food and beverage and much of the front of house have no use for a Copilot seat, not because those people are less capable but because the work is physical, guest-facing and does not run through Word, Excel and Outlook. At 6:30am the housekeeping supervisor is allocating rooms and counting linen, and there is no document in that hour to draft.

So the knowledge-worker population in a 300-room property might be twenty people out of two hundred. Buy a seat for everyone with an email address and you inflate the unused pile faster than any office-based business could manage. The figure quoted back at you in the renewal meeting is then real, and it is your own doing rather than the product’s.

The four functions above share one property. The people in them already spend a large part of the week turning existing data into written work, which is the task these tools do well. Everything else, including the enthusiastic general manager who wants one to try, can wait for a second wave and be paid for by the first.

Chapter 05

None of this works without measurement, and measurement here is smaller than it sounds. You are not building a dashboard. You are answering three questions once a quarter, in a meeting that takes half an hour.

First, how many assigned seats were active in the last 30 days, by department rather than in total. A group figure hides everything that matters. Finance almost fully active while sales barely touches it is a management problem in one department, not a verdict on the product. Second, what the active people are doing with it. The 1.14 actions per working day from the UK pilot is a useful yardstick precisely because it is unglamorous.

Third, which named recurring task has actually moved. Not a sentiment score. A specific piece of work that used to take somebody a morning and now takes an hour, with a name and a date against it. Three of those, written down, carry a renewal conversation further than any survey of how people feel about AI.

Pulling back unused licences is the conversation everyone has first, because it is concrete, quick and produces a number the finance director likes. Do it. Twice a year, reclaim what has not been touched in 90 days and move those seats to a function on the list. Then be honest about the size of the prize. Reclaiming seats recovers part of the smallest line in the Forrester split, while the other two thirds, implementation, management and training, is where the return either appears or does not. The reclaim conversation makes the spend look tidier. The adoption conversation makes it worth having.

If the invented figure is already in a deck, replace it rather than quietly deleting it. The honest version is stronger. No evidence says most Copilot licences go unused. The best public evaluation found about two thirds of holders active in a week and roughly one action a day. The real exposure is not idle seats, it is the two thirds of cost that never got spent on making the seats work. An Sbit AI-Readiness Scan is one way to find which functions repay a seat, though a controller with a list of recurring tasks will get most of the way there on their own. Either beats a statistic somebody reversed on the way to a slide.