AI adoption is a fairy tale. What counts is how much work gets done on its own

In every company that has bought AI licences for everyone, the same thing happens. A few use them every day and have got the hang of it. Some open them a couple of times a day and use them badly. The majority do not open them. It holds for an office of ten and for a group of a thousand: the scale changes, the shape does not.

The owner looks at the numbers and sees that “adoption” is high. Then they look at the work and see that nothing has got faster. Both things are true at once, and the second is the only one that counts.

Using it and using it well are two different trades

Even among those who use it, the distance is enormous. Take any task in administration: replying to a supplier who asks when they will be paid. The first person pastes the email into the chat, takes the answer as it is and sends it. If there is a wrong date inside, the supplier finds out. The second person gives context, the invoice, the agreed due date, the payment method, reads the answer, corrects one line and sends a shorter, correct message.

Same tool, same desk, opposite result. Using AI well is a trade learnt through many repetitions, and at least half of a company will never get to the second version. It is not a fault: those people have a job, and it is not becoming good at a chat.

A perfect rollout does not cure it

The objection is always the same: “those companies did the rollout badly, we will do it well”. No. Even with training, licences for everyone and an enthusiastic manager, the result has the shape of a barbell: a few who consume almost everything, many who consume nothing. And if the many started using it like the few, the bill would go up tenfold. The best case becomes the worst.

Why “adoption” is the wrong word

Adoption is measured with a yes-or-no question. Did they open the program this month? Did they send at least a few requests? But the ability to use AI is not a yes or a no: it is a spectrum that goes from “never opened it” to “pastes emails in to reformat them” to “has three processes running on their own on the accounts”. An adoption percentage squashes all of that into a single number, and the number is true. It just says nothing.

It is not an Italian problem. A 2025 MIT report counted that the great majority of AI pilot projects in companies leave no measurable trace in the accounts. Many projects, very few results, and in between an indicator that reassures everyone.

Incentives keep the gap open

Those who sell AI add capability with every release, and nobody adds the capability to use it. Every new feature raises the bar for anyone who wants to use it well, and widens the distance between the few and the many. The few, inside the company, have no interest in closing it: their advantage is precisely that gap. They do the work of three and nobody asks them to.

Even classic training, “let’s teach everyone to write prompts”, is a small part of the game. The big part is understanding which jobs should never touch a model and which should be automated entirely. And the answer is different in every company.

What to do, then

Train anyway, but to understand who is who. Training is a diagnosis, not a cure: it brings out the interested people who never had the time to go deep.

Give the good ones a place to publish what they build, so one person’s discovery becomes something the others can use. Recognition is the only incentive that works: the good ones trade their edge for reputation.

For everyone else, the work has to get done without them changing how they work. That is: the AI goes into the background. You take the most repetitive processes, invoices to record, transactions to match, reminders to send, and automate them inside the tools that are already there. People approve, correct, reject. They do not have to learn anything.

The number to bring to the meeting

Stop reporting adoption. Report how much of the administrative work today is manual, how much is mixed, how much is automatic. It is the only number that describes a real change, and the only one that can improve month after month without asking anyone to become good at a trade they did not choose.

What we build

Ditta is designed for the rest of the company, not for the few. The virtual employee works on the sources that are already there and does the repetitive part of administration on its own; whoever keeps the books confirms the doubtful cases. Those who want to ask questions ask them in chat or on WhatsApp, as they would a colleague. Those who do not, find the work done anyway. Adoption, in this scheme, is not a problem: there is nothing to adopt.

The starting point was an article by Varick Agents, written for large American companies. The proportions change, the mechanism does not.