AI fluency: artificial intelligence in business is a matter of common sense, not technology

There are two ways to get artificial intelligence wrong in a company, and in recent months we have seen both. The first is being swept away: pasting everything into an assistant, taking whatever it says at face value, publishing it. The second is rejecting it outright, out of fear or principle, while competitors use it and staff use it anyway, quietly. In between there is something with a name, AI fluency, and it is hardly ever a technical matter. It is common sense applied to a new tool.

A note for readers outside Italy: the figures in the next section describe Italian businesses, from the Bank of Italy, Istat (the national statistics institute) and Eurostat. If you run a hotel or a company in Italy from abroad, they describe the market your staff, suppliers and competitors work in.

We are not saying this from the sidelines. Since this summer our website has had an editorial process that uses artificial intelligence under written rules, and over the past few months we have learned more from corrected mistakes than from theory. This article is about what working well with the machine means for a small business or a hotel.

AI in Italian businesses, in figures

The freshest figure comes from the Bank of Italy, which between February and May 2026 asked companies with at least 20 employees whether they use artificial intelligence tools in their processes. The answer: 32%, including those only experimenting with it, against 27% a year earlier, when the share had already doubled. Another 10% say they will start within the year. But when the question becomes “do you use it extensively?”, only 4% of companies below 50 employees say yes. For micro-enterprises there is a separate Bank of Italy survey, collected in 2025 and published in September 2026: artificial intelligence is in use, even in a limited way, in 5% of cases. Counting those only experimenting, the figure rises to 14.4%, with accommodation and food service slightly below, at 12.7%.

The wider picture is still Istat’s, collected in 2025 and published in December: among companies with at least 10 employees, 16.4% used at least one artificial intelligence technology, double the year before; 15.7% of small and medium-sized firms, 53.1% of those with 250 employees or more. In the Eurostat data from the same survey, the share drops to 14.2% among firms with 10 to 49 employees. These are different surveys, with different thresholds and definitions, and their figures should not be lined up as if they were a series. But they say the same thing: plenty of experimenting, little real use, and small firms falling behind. The edition with 2026 data has not been published yet.

The most telling figure, though, is a different one. Istat asked the companies that had considered artificial intelligence without then adopting it, 11.5% of those not using it, what had stopped them. The first answer is not cost, which comes fifth, nor privacy. It is the lack of skills, at 58.6%. The same point comes up in a study published by the Bank of Italy in June 2026, in its Occasional Papers series: skills shortage emerges as the single factor that most holds adoption back, and most companies already using it say they have not yet seen any effect on productivity. On the people side, the figures point the same way: in 2025, according to Eurostat, 8.0% of Italians aged 16 to 74 had used generative artificial intelligence for work in the previous three months, against 15.4% across the EU.

If the brake is skills, the right question is not which tool to buy. It is which skill is needed, and the answer is less technical than you might think.

Four skills, and three are about judgement

The framework we use as a reference was developed by two university lecturers, Rick Dakan and Joseph Feller, together with Anthropic, and describes AI fluency as four distinct skills, known as the 4Ds because they all begin with D. We tell them in our own words and with work scenes you will recognise, because that is where you can tell whether a skill is there or not.

1. Delegation: deciding what to hand over, and how much freedom to give

The first skill comes before any prompt, that is, before the instruction you write to the assistant: it is deciding whether a task should go to the machine, which part of it, and how far to let it work on its own. At a hotel front desk, the assistant can prepare the reply to a group quote request: it gathers the dates, checks the tone, drafts a text. It does not decide the discount. Before a supply contract is signed, it can summarise it and flag the clauses on renewal and penalties; it does not decide whether to sign. Someone who delegates well knows where the line is, and moves it over time as they see how the tool works. Someone who delegates badly makes one of the two mistakes described at the start: hands over everything, or nothing.

2. Description: explaining what you want as you would to a new hire

The second skill is the one usually called “writing good prompts”, and it is the only one that resembles a technique. It even has a name, prompt engineering, and it has become a discipline of its own: it has handbooks, courses, a professional role, and the companies that build the models publish guides on how to instruct them. The techniques are real: giving a role, showing examples of what you want, asking it to reason step by step before answering, fixing the output format. But the foundation is simpler than the name suggests: it is the same ability you need with someone who has just joined the company. “Write me an email to the client” produces a generic email. A real brief says who it is for, what happened before, what needs to happen next, what cannot be promised, and how to recognise a good answer. The difference between the two results is not in the tool, it is in the person instructing it. And someone who can brief a person usually learns quickly to brief a machine.

3. Discernment: judging what it gives back

The third skill is the one most often missing, and the figures show it. In a report published by Anthropic in February 2026, across nearly ten thousand real conversations with an assistant, only in 8.7% did the person verify the facts that mattered, and in 15.8% they questioned the assistant’s reasoning. The sample leans towards early adopters, so probably more at ease than average. A University of Melbourne study across 47 countries, co-funded by a consulting firm and so to be taken with caution, looks at the same thing from another angle: among workers who use it, two in three admit they have taken what it produced as correct without checking it.

Here we can tell a story of our own. Since September every article on this site starts from a dossier of primary sources, is written with the help of artificial intelligence, and before going out passes through a reviewer, also automated but with a different job, and a reading by a person. The checks regularly catch things that would have slipped through: a percentage that was not in the original document, a figure taken from the wrong edition of a report, a sentence that claims more than the numbers support. In September the reviewer stopped a bold conclusion, in an article that had already been read, because the data said something else; in this very piece it found a number attached to the wrong sample. That is not a flaw in the system: it is proof that discernment cannot be delegated to whoever produces the text. It has to sit somewhere else, and in the end with a person who answers for it.

4. Diligence: putting your name to it, towards readers and towards data

The fourth skill is responsibility, and it has two sides. The first is towards readers: since 2 August 2026, Article 50 of the EU AI Act requires transparency, with different obligations for those who provide and those who use certain artificial intelligence systems, and we wrote about it in July. At the end of July the European Commission updated its answers to frequently asked questions, which among other things distinguish the case of texts that have gone through human review: they are worth reading, and it is worth not drawing conclusions from them on your own. We disclose generated images and say openly how we use artificial intelligence in our texts too, as in this article, because transparency costs less than suspicion.

The second side is towards data. Someone who pastes the customer list, a contract or a guest’s record into a public assistant “because it’s quicker” is doing the same thing as the parallel spreadsheet on the desktop we described in our piece on technostress: solving today’s problem with the tool at hand, and shifting the risk onto the company. It is called shadow AI, and it is not an exception. A University of Konstanz survey of 1,105 people in Germany, so not in Italy, found that among workers who use artificial intelligence, the employer was the source of the main tool for 55%; in small organisations, training had reached 11% of staff and written rules 10%. We found no reliable Italian figures, but we have no reason to think things are better here. And data leaving the company without control is also a door for attackers, as we described in our piece on AI-driven ransomware.

What AI fluency is not

It is not knowing how to code, nor even knowing the models. It is not a certification. And it is not the list of things left to humans, creativity, empathy, which gets repeated and helps nobody work better on Monday morning. The real human share, across the four skills, is more concrete: deciding what to hand over, judging in context, answering for the choices. These are things a good finance manager or a good front office manager already do, with people. Artificial intelligence asks them to do it with a machine as well.

On how to choose where to start with tools, agents and processes to automate, we wrote in August. This piece is about something else: the skills of the people who then use those tools.

Our view

AI fluency cannot be bought or installed. It is a skill that belongs to people, and it is built by working, getting things wrong and correcting them, as we have done with our own site. But it rests on a technical foundation that companies often lack, and that is the part that concerns us: tools chosen on purpose, tidy access rights, data that does not leave the company. If the only assistant available is the free public one, people’s diligence is working against an infrastructure that lets it down. If nobody has decided what can be pasted and what cannot, using AI in a company means leaving delegation to each person on their own. That is the daily work of whoever runs a company’s IT, and it is the same reasoning we apply to cyber security: rules do little if the tools make them awkward.

We do not sell courses and we do not promise them. What we can offer is a conversation: how artificial intelligence is entering your company, who already uses it and for what, what leaves your systems without anyone having decided it. If you would like to have it, get in touch: we start from there, not from a product.

Frequently asked questions

What is AI fluency?

It is the ability to work well with artificial intelligence: deciding what to hand over and with how much autonomy, explaining what you want, judging what it gives back and answering for what is published or decided. The reference framework, developed by two university lecturers, Rick Dakan and Joseph Feller, together with Anthropic, describes it as four skills. In our view three of the four are about judgement, not technique.

How many Italian companies use artificial intelligence?

According to the Bank of Italy, between February and May 2026, 32% of companies with at least 20 employees used artificial intelligence tools, including on an experimental basis, but below 50 employees extensive use concerns 4%. A separate Bank of Italy survey of micro-enterprises, collected in 2025, finds artificial intelligence in use, even in a limited way, in 5% of cases, 14.4% counting experimentation. According to Istat, in 2025, 16.4% of companies with at least 10 employees used at least one artificial intelligence technology: 15.7% of small and medium-sized firms, 53.1% of those with 250 employees or more. According to Eurostat, the EU average for firms with at least 10 employees was 20.0% in 2025.

Why do many companies still not use artificial intelligence?

Istat asked the companies that had considered artificial intelligence without then adopting it, 11.5% of those not using it: the first reason is the lack of skills, cited by 58.6%, followed by unclear rules, data quality, privacy and cost. In other words, the main brake is not the price of the tool but knowing how to use it.

What is shadow AI and why is it a risk?

It is employees’ use of public artificial intelligence tools without the company knowing or having decided it, often pasting in customer data or internal documents. The risk is that the data leaves the company and nobody answers for what the tool produces. The answer is not a ban, which only moves the problem, but tools chosen by the company, clear rules on what can be used, and training for people.