Which AI tools are actually being used, and where that shows

Most organisations try to find out which AI tools are in use by asking. There is a better source, and you are already paying for it.

The data is in the licences

Zylo's 2026 SaaS Management Index, built on analysis of more than 40 million software licences and 75 billion dollars in spend under management, states that organisations leave an average of 36 per cent of their licences unused.

That figure is not primarily a savings target. It is a measurement of something otherwise hard to reach: the difference between what an organisation decided and what it actually does. A survey answers what people feel they ought to be using. The licence data answers what is actually opened.

The Government Digital Service measured the same thing in telemetry

The UK's Government Digital Service trialled Microsoft 365 Copilot across twelve government organisations and read usage out of Microsoft's built-in dashboard rather than asking anyone. The threshold it used is the same logic that sits behind an unused licence: an active user is one with at least one interaction in the previous thirty days, and the adoption rate is active users divided by licence holders. The experiment covered 20,000 assigned licences, at least a thousand per organisation, and ran from 30 September to 31 December 2024. The report was published on 2 June 2025.

During October the rate climbed to 83 per cent, and adoption of around 80 per cent held for the rest of the experiment. Roughly a fifth of licences therefore saw no interaction at all in thirty days. At application level the distance is wider: Teams reached a maximum of 71 per cent, while Excel and PowerPoint peaked at 23 and 24 per cent.

Two things make that a floor rather than a ceiling. Telemetry existed for only 14,500 of the 20,000 licences, and the report states that data was not available and hence not supplied for the remainder, so the share without interaction describes the measured users and not the whole population. And it was a time-limited experiment with active change management throughout, which is likely to produce more favourable numbers than ordinary operations.

This is a different kind of measurement from Zylo's. Telemetry in a public sector population against platform data in a private one, with the same threshold underneath. That eighty per cent was reachable, and that the rise arrived as change management activities rolled out, says something about what the distance between licence and use is made of.

Meanwhile AI spend grows faster than anything else

The same index reports AI-native applications as the fastest-growing spend category. Spend rose 108 per cent in a year overall, and 393 per cent in organisations with more than 10,000 employees. Expense-based software purchasing rose 267 per cent, and ChatGPT is now the most expensed application.

The two observations sit together uncomfortably. Spend on AI is rising faster than on anything else, in an environment where more than a third of what is bought never comes into use.

Buying faster changes nothing about what decides the outcome: whether the tool earns a place in the working day.

It may be that the tool did not solve the problem the buyer thought it would. It may be that it had no place in the workflow, or that nobody was given time to learn it. Which of those it was can be established, but only if somebody asks before the next tool is bought.

What the data usually shows

The list below ChatGPT is more instructive than the top of it. Zylo's ranking of the most expensed applications continues with Apple iCloud and Canva, and OpenAI's API sits fifth. Eight of the fifty most expensed applications are AI-native, which is 16 per cent.

Torii measures in a different way, by discovering applications in corporate environments rather than reading expense claims. Its review of 2025 found that twenty-six of the fifty commonest shadow IT applications were pure AI tools. Over the year, 694 new AI-native applications appeared in corporate environments, and the average company runs 830 applications in total.

Neither Zylo nor Torii has a Nordic population, and neither of them reports geography. The ordering is probably the same here; the levels should not be assumed to be.

So the question is rarely whether there is AI in the building. It is which, and who knows about it.

And most of it is bought outside IT

The index also reports where the decisions are made. Business units control 81 per cent of software spend while IT directly manages 15 per cent.

That explains something otherwise puzzling, namely how organisations regularly discover tools they did not know existed. Most of what is bought never passes the function that normally keeps the system inventory. When the purchase is made as an expense claim it appears in the finance system and nowhere else.

A separate figure in the same index is about pricing rather than about who bought the tool. 78 per cent of IT leaders report unexpected charges over the past year tied to consumption-based or AI pricing models, and 61 per cent have been forced to cut projects because of unplanned cost increases. What is measured there is the mechanics of pricing, not the purchasing route: a consumption-based charge is just as capable of surprising a centrally negotiated contract. That so many cut projects means those increases displace work that was already planned.

Buying outside IT does, on the other hand, mean the cost is noticed late, because whoever makes the purchase is rarely the person watching consumption.

Those two figures do not come from the licence data but from the index's survey component, which is based on 218 IT leaders. That is a different population and a different kind of measurement: what leaders report, rather than what the transactions show.

The pricing mechanics behind that are covered in AI costs: from licence to consumption.

Four sources you already have

1. The finance system. Subscriptions and recurring invoices, sorted by supplier. That gives the list of what the organisation pays for, including things nobody remembers the reason for.
2. Expense claims and receipts. This is where shadow AI sits, meaning what employees pay for themselves and claim back. Search for the commonest supplier names.
3. Sign-in data from identity management. Who actually signs in, how often, and when last. It is the only source that separates used from paid for.
4. Web traffic to known endpoints. Catches what is used with no contract at all. Handle it carefully and with the data protection considerations that apply, but even in aggregate it says a great deal.

Compiled together, those four sources give a picture of actual usage that no survey response can produce.

What to do with the answer

The temptation is to start cancelling. That is right for part of the list and wrong for the rest.

Licences that were never activated at all are pure savings and can be ended immediately. Licences used rarely by a few people may on the other hand be essential to exactly those people, and withdrawing them saves a little money and costs trust.

The interesting category is the third: tools that were bought with enthusiasm, used for a couple of weeks, and then went quiet. That is where the answer lies to why the next implementation will go the same way, and that answer is worth more than the licence fee.

This is a different question from whether a programme produced results, which requires that somebody decided a metric in advance. That distinction is covered in Do you know whether your AI programme worked. Usage can be measured after the fact because the logs already exist. Effect rarely can.