The rollout went smoothly: two hundred licences, an account for everyone, a short introduction. Eighteen months later a renewal is on the agenda and nobody can say how many of those two hundred people opened the tool last month.
AI licences versus actual usage is the gap between what you pay per period for user rights and the number of people who actually use them. With AI software that gap is structurally wider than with other business software, because licences tend to be bought organisation-wide while usage concentrates in a limited group.
This guide is about the cost side: how to measure real usage, which pricing models exist, what to do at renewal, and when owning infrastructure becomes cheaper. We covered adoption and training as an organisational question separately in training employees on AI tools.
What does an unused licence really cost?
The direct cost is simple: unused licences times the price per user times the term. Per-user pricing for AI assistants runs higher than for most other business software, so an unused share adds up quickly on an annual basis.
Three further costs run alongside that never appear on the invoice.
Administration. Every issued account needs provisioning, permissions and removal when someone leaves. Accounts nobody uses are also the worst maintained.
Risk. An active account that is unused and unmonitored is standing access to company data. This weighs more heavily the more internal sources the tool connects to.
Crowding out. A broadly deployed, lightly used general package makes it harder to get budget for the targeted application that would pay off, because "we already have AI" counts as an answer.
Against those three sits one real counterforce. Cutting licences can push users back to personal accounts, putting company data out of sight. That is the pattern we describe in our guide on setting a shadow AI policy. Reducing without an alternative is therefore not a saving but a relocation of the problem.
How do you measure real usage?
Licences purchased tell you nothing. Accounts created tell you nothing either, and a single sign-in even less. Measure at four levels, because the gaps between those levels are exactly where the money sits.
| Measurement | What it tells you | What you act on |
|---|---|---|
| Licences issued | what you pay | the ceiling on the bill |
| Active in the last 30 days | who still opens the tool | candidates to withdraw |
| Weekly returning usage | where it sits inside the work | the core you want to keep |
| Usage inside a business process | where it touches production | where expanding or building pays |
The jump between the second and third rows matters most. Someone who asks something once a month is not using the tool in their work but trying it occasionally. That is not a criticism and usually not a training problem: it means the tool has no fixed place in a task for that role.
Two practical notes. Take the figures from the supplier's admin console rather than from a survey, because self-reporting overstates usage consistently. And look at the distribution per department rather than the average: an average of forty percent can mean two departments use it intensively and five not at all, which leads to a very different decision.
[ TIME SAVED ]
Save 4 hours per week on manually tracking licences, accounts and renewals
Which pricing models exist and when does the bill tip?
There are three in practice, and the choice determines how your bill behaves as usage grows.
Per user per month. Predictable and easy to buy. The drawback is that the bill grows with the organisation rather than with value: you pay the same for someone who uses the tool daily as for someone who never opens it.
Per consumption. You pay for what you use, which tracks value but makes budgeting harder. This model does not penalise broad shallow usage and rewards concentration where it pays.
Fixed capacity or your own environment. A set amount for compute, regardless of volume. The bill is predictable and falls per unit as usage rises.
The tipping point between the first and third sits where the number of intensive users and the volume together are high enough to carry the fixed cost. With broad, light usage, licences almost always win. With a limited group working on the same task all day, it tips sooner than most organisations expect. The wider cost picture is in our guide on AI costs explained.
What do you do at renewal?
A renewal is the only moment you have negotiating position. Prepare it three months ahead, in four steps.
Take a usage measurement over three months. One month is too short because of holidays and peaks. Record per department how many people return weekly.
Withdraw dormant accounts before the measurement, not after. That gives a cleaner picture and immediately lowers the baseline for the conversation.
Split your population. Intensive users keep their licence, occasional users get a cheaper alternative or a shared facility, and non-users come off. Buying uniformly for the whole organisation is almost always the most expensive option.
Put what you need into the contract. The ability to scale down mid-term, monthly usage reporting, and clarity on what happens to your data and configuration if you leave. Our guide on buying AI software and contracts covers the clauses that belong there.
The biggest saving is rarely a lower price per user. It is the number of users you stop paying for, and the ability to scale down again next year.
When do you move from licences to your own environment?
Three signals suggest you should change pricing model rather than negotiate.
Usage is concentrated and heavy: a limited group runs the same kind of task intensively, every day. That is precisely the profile where fixed capacity becomes cheaper per unit.
Usage sits inside a process rather than with people. Once AI is part of a processing flow you are paying for volume, not for people, and a per-user licence is the wrong instrument.
The data cannot leave the building, for contractual or legal reasons. Then it is not a cost trade-off but a constraint.
When comparing, always count the full cost on both sides: owning an environment brings administration, updates and monitoring with it. That breakdown is in our guide on AI without ongoing API costs. For general office use by many people, licences usually remain the sensible choice, which is also why we do not argue against general assistants: they are simply rarely the answer to one specific, heavy task. We work that distinction out in using AI copilots in business.
Conclusion: measure before you renew
AI licences are typically bought broadly on an expectation and never tested afterwards. The result is a fixed annual cost of which part produces no result at all.
The correction is modest work: measure three months of usage from the admin console, split the population into intensive, occasional and non-using, withdraw dormant accounts, and set the contract up for scaling down. Do that once a year before renewal and you pay for usage rather than for access.
We help with that assessment through AI consulting, including the question of when your own environment becomes cheaper than continuing to pay per user.