You are paying for AI tools nobody uses. Here is how to fix it.

Pull your last three months of software invoices and highlight every line that mentions AI, copilot, assistant, or intelligence. For most companies the number is larger than expected, and the usage behind it is smaller than expected.

This is not a procurement problem. It is an adoption problem with an invoice attached.

Step one: find the dead seats

Nearly every AI product has an admin view with last active dates. Export it and sort by activity. You are looking for three buckets:

• Never activated. Seats assigned to people who never logged in. Cancel these this week.
• Tried and abandoned. Active in month one, silent since. This is the group worth saving, because they already had a reason to try.
• Genuinely used. Usually a small number of people. Find out what they are doing and make it the template for everyone else.

Step two: separate tool problems from skill problems

Ask five people in the abandoned bucket a single question: what were you trying to do the last time you used it? The answers sort themselves quickly. If people were asking for something the tool genuinely cannot do, that is a tool problem and you should cancel. If people were asking badly for something the tool does well, that is a skill problem and cancelling it just moves the waste somewhere else.

Cancelling a tool your team never learned to use does not save money. It just hides the gap for another year.

Step three: consolidate

Most stacks we look at contain three or four products doing overlapping work, each used at ten percent. One assistant that everybody genuinely knows how to use beats four that nobody does. Consolidation also makes training possible: you cannot teach depth across six products, and depth is where the return is.

Step four: attach every remaining licence to a named task

For each tool you keep, write one sentence: this tool exists so that [role] can do [task] in [time] instead of [time]. If you cannot write the sentence, you do not have a use case, you have a subscription.

Step five: teach, then measure the task, not the licence

Run real training on live work, then measure the task you named. Did the proposal draft get faster? Did the research step disappear? Licence utilisation is a vanity metric. Time returned to the business is the real one.

What good looks like after ninety days

• Fewer tools, higher spend per tool, dramatically higher usage.
• Every remaining product has an owner and a named task.
• At least two workflows that were manual last quarter now run with AI in the loop.
• People ask for access instead of being assigned it.

The money is rarely the point. The point is that the capability you are already paying for is sitting unused while your competitors work out how to use theirs.