“Should I add AI too?” is the question going around many SMEs today. The honest answer is neither yes nor no: it is “it depends on the process”. Artificial intelligence is a tool, and like any tool it pays off where it saves time or reduces errors in real work, not because it is in fashion. The point is not whether AI works, but whether it works in your case.
According to ISTAT (the Enterprises and ICT survey, 2025), 16.4% of Italian enterprises with at least 10 employees use at least one AI technology, double the 8.2% of 2024. But the average hides a clear divide: 53.1% of large enterprises use AI, against 15.7% of SMEs (with at least 10 employees, the survey perimeter). The gap is not only about budget: it is also about processes and data being ready. That is exactly where it gets decided whether AI pays off.
Where AI really helps
AI delivers when it works on something repetitive, frequent and already in order. It is good at reading piles of text and pulling out a summary, at classifying, at answering questions about data that already exists in clean form. Three situations where it usually pays off:
- High repetitive volume: the same task dozens or hundreds of times a week (sorting requests, extracting data from similar documents, drafting replies).
- A question instead of a hunt: querying data you already have in plain language, instead of opening five screens or building yet another report by hand.
- Tolerable error: a draft a person reviews, a first classification someone confirms. AI proposes, you decide.
The common thread is that AI works alongside a person, not in their place. It removes the mechanical work and leaves the judgment to whoever should make it. That is how, for example, in Verso Flow Flow AI, built on Claude by Anthropic, answers in plain language about teams' operational data: “what got left behind today?” instead of opening five screens. It does not decide for you, it gets you to the answer sooner.
Where AI is not enough (or makes things worse)
There are cases where AI only adds cost and complexity to maintain. It is not a limit of the technology: it is the wrong case. Keep it out, or hold off, when one of these signals shows up:
- The process is still chaotic: if the work lives between phone calls, chats and scattered spreadsheets, AI does not bring order, it amplifies the mess it finds.
- The data is dirty or incomplete: duplicate records, empty fields, different versions of the same information. AI answers based on what it reads, and if the data lies, it lies with confidence.
- You need a fixed rule, not a judgment: if the decision is “if A then B”, a written rule is more reliable, cheaper and easier to verify than a model that is sometimes wrong.
- The error is not tolerable: if a result goes straight into accounting, into an invoice or into a signed report to the client, every error costs dearly and has to be re-checked by hand anyway.
You can see it in the ISTAT figures too: among enterprises that evaluated AI without adopting it, the most cited barriers are the lack of adequate skills (58.6%, the most cited), poor data quality or availability (45.2%) and costs (43.0%), alongside regulatory uncertainty and privacy. Put plainly: the problem is almost never AI itself, it is the foundations missing underneath.
The decision grid: three questions before you spend
Before signing up for any tool “with AI”, hold the process you have in mind up to three questions. They are the same ones we ask ourselves before proposing AI in a project.
- Is the volume high and repetitive? If the task comes up twice a month, automating it does not repay the effort. If it comes up a hundred times a week, it is worth a look.
- Is it a fixed rule or does it need judgment? A clear rule is solved with a rule, not a model. AI shines where there is text to interpret, summarizing to do, blurry cases to classify.
- How much does an error cost? If the output is reviewed by a person, the error is tolerable and AI saves time. If it ends up in accounting or in a report to the client, the bar rises, and a deterministic rule plus a check is often the better choice.
An honest middle way: if the answers are uncertain, start small. A single, well-bounded process, with a person who reviews the output. You measure whether it really saves time, and only then do you expand. No cathedrals, the same principle we use when we start from a first core in custom software.
Before AI comes the process (and clean data)
AI pays off where there is an orderly process and clean data. Without those, it amplifies the chaos: it gives quick answers from wrong foundations, and gives them in a confident tone. That is why, often, the first useful step is not “add AI”, but put the upstream process in order.
Putting it in order means having the data born where and when the work happens, and making the systems you already use talk to each other instead of re-entering everything by hand. That is what we do by digitizing field operations and with data integration between systems. When the foundations are there, AI becomes a natural step; when they are missing, it is an expensive patch.
How we use it: where it helps, not where it shows off
We use AI for real, where it brings value. We bring it in when it makes a process better and we keep it out when it would only add noise: the same criterion on the products we build and on the custom software for our clients. It is a choice about the result, not the label: software that works, first of all.
If you want to understand where AI makes sense in your company and where it does not, tell us how you work today: we start from the process, not from AI, and we tell you honestly whether it really pays off.