Why licences are not adoption.

Microsoft passed thirty million paid Copilot seats at the end of July. Independent estimates put the share that is actually used in an average week at twenty to thirty percent. For all the rest, an organisation pays around 360 dollars a year for a door nobody opens.

From the outside we seem to be doing better. More than three quarters of Belgian companies with over 250 employees have at least one AI technology in house, good for a place in the European leading group. Ask those same employees whether AI saved them time over the past year and 48 percent say yes. The world average is 62.

We buy AI with the leading group and use it in the peloton.

The diagnosis that follows, I know by now. It would be down to people's resistance. There will be an ambassador network and a forty-minute e-learning that nobody sits through.

And yet that non-user is often right. They pay the learning curve today, out of their own diary, while the return only shows up next quarter on somebody else's dashboard. If they get it wrong with a model, their name is on it. If they do nothing, there is nothing. In a house where nobody has said out loud what is allowed, sitting still is the best protected choice there is. 39 percent of Belgian employees say they need extra AI knowledge, 14 percent were given training, the rest are guessing.

Resistance is then a tidy report of how the work is put together. Anyone who reads that report as a character flaw in their staff simply buys licences again next year.

Because a licence is a key, and a key opens nothing as long as nobody dares to say which rooms you may enter, at what risk, and who gets the invoice when it goes wrong. That is design work. It belongs before the purchase order, well before the communication campaign that follows.

So do not count logins. Count the decisions that fell differently last month because somebody let a machine contradict their own assumption. That number is on no quotation, not on mine either, and that is exactly why it proves something.

Which decision fell differently for you this past month because of a machine? If the answer does not come within thirty seconds, your rollout is on the cost line only, for now.

🔥 THREE FORGOTTEN QUESTIONS

1. Who in your organisation decides what an AI mistake costs, and does that person already know it?

2. What if your rollout only succeeds the moment your people dare to contradict the machine, while your measurement system keeps logging exactly that as resistance?

3. If three quarters of large Belgian companies buy the same licences from the same handful of suppliers, where is the difference you can still sell to a client?

Figures in this piece

Microsoft 365 Copilot: 30 million paid seats, Microsoft Q4 FY2026 (29 July 2026). Weekly use 20 to 30 percent of the paid seats: independent estimates, indicative, consistent with Gartner findings 2026. Cost of an unused seat: about 360 dollars a year.

Belgian companies with 250+ employees with at least one AI technology: 76.4 percent (Statbel / Eurostat, 2025 figures).

Time saved through AI: 48 percent of Belgian employees against 62 percent worldwide (Ipsos AI Monitor, surveyed March and April 2026).

Need for AI knowledge 39 percent against 14 percent who were given training: High Council for Employment, report on AI in the Belgian labour market (February 2026).