The market is halfway across, and this is the useful moment
In 2025, 16.4% of Italian businesses with at least ten employees use at least one artificial intelligence technology. It was 8.2% in 2024 and 5.0% in 2023: the share has more than tripled in two years. But among small and medium-sized enterprises it stops at 15.7%, against 53.1% of large companies.
That gap is an exact photograph of the situation: AI is no longer a laboratory bet, but in smaller companies it has only just started. And the main reason is not cost. Istat finds that a lack of adequate skills holds back adoption in almost 60% of cases — meaning the problem is not being convinced it is useful, it is knowing where to start without causing damage.
(Source: Istat, “Imprese e ICT”, 2025.)
Where to start: from processes, not tools
This is the most widespread and most expensive mistake: choosing the tool first, then looking for something to use it on. It almost always ends with a licence paid for and nobody using it.
The order that works is the reverse. You start from a concrete question: which repetitive activities consume the most hours every week? In most smaller companies the list is surprisingly similar.
- Sorting and filing incoming documents — invoices, orders, delivery notes, correspondence.
- Answering the same questions, from clients or from colleagues.
- Rekeying by hand data that one system already has and another never receives.
- Hunting for information scattered across folders, email and the business system.
- Preparing recurring documents that always start from the same template.
Then you check which of these have data already available in usable form, because that is where AI produces a result in weeks rather than months. Out of those two lists comes a priority based on impact and feasibility.
One case, measured, before anything else
The most useful thing a smaller company can do is pick one process and take it to a verifiable result. Not out of caution: because it is the only way to know whether it works in your company with your data.
A well-set-up pilot has three characteristics: a defined process with a clear boundary, a starting measurement — how long it takes today — and a date by which the result gets looked at. If the number does not move, you have learned something cheaply. If it does move, you extend with experience already gained.
The alternative — buying a platform for the whole company before verifying a single result — is the fastest way of turning AI into a cost with nothing on the other side.
What has to be put in order first
This is the part nobody talks about and it determines success more than the technology does.
The documents. An assistant that searches company archives works if those archives have a logic. If they sit in folders accumulated over ten years under different conventions, that is the first job — and it is worth doing anyway.
Access rights. If AI searches everything there is, it hands anybody things they should not see. The permissions have to be the person’s, not the system’s.
The rules. What may be entered and what may not, who checks the output. One page that gets read is worth more than a long document that gets filed.
Somebody in the company is already using it
This is the situation we find practically everywhere: no company decision, but somebody pasting a document into a free service to have it summarised. It is not disobedience — it is a real need the company has not yet answered.
Banning without replacing moves the behaviour instead of removing it. The method that works is census, replace, then close: know which tools are in use, offer an official one that does the same job without data leaving, and only then take the alternative away.
This work is needed for another reason too: the European regulation on artificial intelligence is fully applicable and requires, among other things, that organisations using these systems know which they are and that the people operating them have adequate training. We cover it under AI Act, GDPR and AI governance.
The rest
See also AI consulting for companies, process automation, custom software development where automation needs a tool that does not yet exist, and smaller companies for the infrastructure side.
The first step
An analysis of your processes: where the hours go, which activities genuinely lend themselves to this and which do not, and what is worth putting in order first. Free and without obligation.