The thesis in one line: training is the line nobody puts in the AI budget, and it’s exactly the one that decides whether the tool you bought gets used or sits dead in an expensive drawer. The license gets approved, the launch gets celebrated, and six months later nobody uses it. It’s not that the tool was bad or the model was dumb: buying the license is half the work, and the other half —getting your people to actually use it, in their real jobs— is the half that got cut first because it didn’t look good in the demo.
The license you bought isn’t the license that gets used
There are two numbers people confuse on purpose because one is comfortable and the other isn’t. The comfortable one is how many people have access to the tool: it’s on the invoice, it gets shown at the steering committee, it feels like progress. The uncomfortable one is how many people actually use it to do their work. The whole project lives between those two numbers, and they’re almost always light-years apart. Paying for access is easy; getting usage is another thing, and it’s the one nobody budgets.
A license nobody uses isn’t a pending saving: it’s a spend already made that returns nothing. Every month between «we bought it» and «they use it» is money burned and, worse, it’s the window in which the team decides «this AI thing wasn’t such a big deal» and goes back to the manual way. That decision, once made, costs far more to reverse than training well from day one would have cost.
The adoption gap: 85% can use it, 25% do
This isn’t a consultant’s hunch: it’s a measured pattern. According to a 2026 IBM study of executives, 85% of employees have access to AI tools, but only 25% use them regularly. That gap —the adoption gap— is where the return promised in the proposal evaporates. It isn’t lost because the model fails; it’s lost because most of the people with access never get it into their daily work. The tool is on, and alone.
And here’s the data point that seals it: in the same kind of studies, most employees say the training they get doesn’t help them use AI in their role. So it isn’t that there’s no training —it’s that the training there is doesn’t work. You run a generic «this is ChatGPT and this is a prompt» course, tick the box, and the person goes back to their desk with no idea how to apply that to what they do every day. The adoption gap doesn’t close with a webinar; it closes with training glued to the real work.
Why generic training doesn’t move the needle
The root mistake is treating AI training as a product course and not as a change in how the work gets done. A generic course explains the tool; what people need is to see their own task done with the tool. Those are different things, and confusing them is what fills companies with dead licenses.
- Generic instead of by role. The salesperson, the admin and the support rep don’t use AI for the same thing. A course that works for everyone truly serves none of them. Training that sticks starts from each role’s concrete task, not from the tool.
- Theory instead of their work. Teaching what a prompt is isn’t teaching how to resolve the email that person writes twenty times a day. How to break your own task into instructions an AI actually follows is half the work, and we cover it in automating tasks with ChatGPT.
- An event instead of a habit. A two-hour session on launch day doesn’t build a habit. Usage sets in with repeated practice, examples from the team itself, and someone who answers questions when they come up, not before.
- No owner of adoption. If the training ends and nobody measures who uses what or helps whoever gets stuck, the tool switches itself off. Adoption needs an owner, just like any other project you want to actually happen.
What to budget when you budget AI
If training is the half that decides the outcome, it has to be in the budget from the start, not as an extra that gets cut when the numbers tighten. And not just any training: the kind that closes the adoption gap has a concrete shape.
- Budget training by role, not a single course. Each function needs to see its own task solved with the tool, with the examples and the data it actually works with.
- Measure real per-person usage, not active licenses. The number that matters is how many people use it each week in their work, and that number is the one that tells you whether the training worked.
- Name an owner of adoption. Someone who answers questions, collects the cases that work and spreads them, and spots who’s fallen behind before they give up.
- Start with one real, visible task, not everything at once. A case people see working —the repetitive email, the same old report— convinces more than any speech, and lays the base of how to use AI day to day in the company, something we land in using ChatGPT in the company.
- If what you’re building is an agent on your information, training includes teaching the team to supervise and correct it, not just use it. That handoff is part of the work, and we break it down in training the agent on your own information.
The dead license costs more than the one you didn’t buy
The invoice for a license nobody uses is the cheap part of the problem. The expensive part is what you don’t see: the team that decided AI «wasn’t such a big deal» and now has to be re-convinced, the project reported as a failure at the committee when it was never actually used, and the month after month of opportunity cost of work still being done by hand next to a tool that’s switched on. Cutting training to save money is the most expensive saving there is: you save the small line and pay the big one.
The takeaway isn’t «train more», it’s «budget adoption as part of the project, not as an extra». The tool is 50%; getting it used is the other 50%, and that 50% is bought with training glued to the real work and an owner who sustains it. If you want that side —your team’s adoption of AI, not just its purchase— run by someone who treats it as the work it is, AI adoption for teams is exactly that: we don’t sell you the license, we leave the team using it.