Menu di accessibilità (premi Invio per aprire)

October 26, 2026

AI in business: what comes before the technology

A conversation with lawyer Arianna Ciracò on governance, the AI Act and data sovereignty

Having a talk with Arianna Ciracò

AI in Business: Legal & Ethical Dimensions with Arianna Ciracò, Esq.

A company meeting room, a privacy training session arranged so that an audit would have something to document. Halfway through, the business owner raises his hand and says: “counsel, my company is extremely innovative”. He has just spent a considerable sum on a biometric time and attendance system. A manufacturing company, with no particular trade secrets to protect. Not a pharmaceutical firm.

Asked why, he replies that it is a cutting-edge product. New.

The episode was recounted by lawyer Arianna Ciracò, Data Protection Officer, in her conversation with Claudio Sorbi for the “Behind the prompt” series. A textbook example, as she calls it, and the important part comes straight after: that business owner had made his choice in complete good faith.

Epistemia: when form replaces verification

The conversation starts from a term coined by Walter Quattrociocchi’s research group at Sapienza University in Rome: epistemia. It describes “that seductive illusion of plausibility” produced by the output of a generative system, an illusion people come to prefer over checking.

A text that is fluent, coherent and assertive resembles knowledge closely enough to pass for it. Fluency becomes reliability, and verification starts to look unnecessary.

Inside a company this mechanism has a direct consequence. When someone in a meeting asks for certainty about a result, they are asking a statistical technology for something it cannot give by design. A generative system does not provide answers: it produces results. Meaning is assigned by whoever reads them, against context and against their own domain experience. Without that step the result is a loose cannon, as Ciracò puts it in the video: it may be correct, because it happens to be statistically plausible, or entirely out of context. From the outside, the two look identical.

Senior, junior and semantic capital

From here follows a question about how work is organised, which comes before the technology, and it is Claudio Sorbi who brings it into the conversation, starting from what Aidia observes internally. AI works very well in the hands of a senior professional, someone who can check the result and use the tool to speed up development. Handing artificial intelligence to a junior profile, without that filter, creates problems.

The immediate conclusion is that senior staff alone will do. Sorbi follows that line through to the end precisely to show where it breaks: and then? Senior professionals will eventually retire or move on, and in the meantime the next generation will not have been able to gain the necessary experience. It gets crushed by that dynamic.

There is a second reason, he adds, which is that a tool has to be used as a tool. It accelerates, it takes in the data it was trained on, it returns results. It does not replace critical judgement or the ability to innovate. What Ciracò calls semantic capital, the body of experience, sensitivity and intuition that remains human, is not passed on by reading outputs.

Why AI makes disorganised environments worse

There is a widespread expectation: bringing artificial intelligence into a process that works badly ought to put it back in order. The opposite happens.

The more disorganised the context and the more complex the data, the easier it is to make things worse. The tool gives you a boost, and Ciracò uses a telling image: it is extra speed in the car. If the direction is wrong, arriving sooner does not help. The risk, she adds, becomes exponential.

This is not an isolated view. Both the AI Act and the GDPR are built on a risk-based approach, that is, on the assumption that these systems are risky by default and that whoever introduces them is responsible for containing the effects.

It is Sorbi who brings into the conversation the term that frames the question: guardrails. Ciracò picks it up from her own side and breaks it down. The guardrails are made of law and ethics: the law sets out what may and may not be done, with the AI Act as the reference text, while ethics indicates, within that compliance framework, which choices serve a social good. Not a good in any romantic sense, she specifies, but one directed at human advancement.

Three things therefore come before AI: attention to the data coming in, attention to processes, and defined governance for the expected result. Without that preliminary work the system returns the disorder it finds, amplified. On this point the two arrive together at the same formula: it is not a magic wand.

The gap for Italian SMEs is not about budget

The report “Italy in the AI era” by Fondazione Leonardo, presented to the Italian Chamber of Deputies on 19 March 2026, describes a national market worth 1.2 billion euros, up 58%, with 5 billion by 2030 as a reasonable target. Within that growth, however, a clear gap remains between medium and large enterprises on one side and SMEs on the other.

The usual reading blames the delay on resources. Experience on the ground tells a different story. The push to act is there, driven by the feeling of falling behind competitors. What is missing is knowing what it is actually for, in that department, with that data.

What is missing, in other words, are guiding partners who work alongside a company starting from year zero, rather than arriving with a ready-made solution. And what is often missing is the awareness that this initial work is necessary at all, because it costs time and produces nothing visible.

There is also a sentence you rarely hear from anyone selling technology: AI is not always the answer. Plenty of other technologies will, depending on the case, solve the problem better.

When the company itself provides the tool

Much is said about employees using AI tools without authorisation. There is, however, a quieter variant, and probably a more widespread one.

An employee opens their work inbox and finds a generative assistance feature already switched on, included in the licence the company provided. They click, ask for a draft reply, a summary of a thread. From their point of view they are using a feature of their work software. In reality they have just begun processing data whose perimeter nobody has assessed, and they did so because the company itself gave them the opportunity without explaining the implications.

Article 4 of the AI Act addresses exactly this. It places on providers and deployers an AI literacy obligation covering staff and anyone using the systems on behalf of the organisation, with training that is graded against baseline skills and tailored to the role performed. Not one generic session for everyone.

Alongside the AI Act, the GDPR and the Italian Workers’ Statute remain in force, setting precise limits on the use of certain technological tools and requiring that the worker be informed. When AI enters processes that involve people, those two texts carry as much weight as the European regulation.

It is demanding and expensive work, Ciracò acknowledges, and it ties up resources from the executive level down to day-to-day operations. The board needs to understand the risks the organisation is exposed to, otherwise the training further down has no mandate behind it.

Transparency: what Article 50 requires

The transparency obligations in the AI Act fall on providers and deployers. The provider must design systems that make it clear, at the first interaction, that the user is interacting with an AI system.

The interesting dividing line concerns content. A text generated with AI but subject to human review and editorial responsibility falls into a different category from output released without that step, which instead comes under the obligations placed on the deployer. For anyone who publishes, communicates or produces documentation, the difference is operational before it is legal.

The same applies to enterprise versions of the most widely used tools. Paying for a subscription does not mean you can do anything you like: the provider sets a scope of use, and knowing it is the responsibility of whoever adopts the tool.

Lock-in and jurisdiction: two things you notice late

Dependence on a supplier does not show itself at signature. It shows itself when you try to switch. Interoperability and portability do exist, Ciracò observes, but they only go so far: in the meantime you have built a working environment around that provider, and a habit has formed. Even where migration is technically possible, the cost of unpicking that habit remains. In between, unilateral changes to the contract can occur.

On data location the decisive distinction is a different one, and it is worth spelling out because it is so often treated as settled. Choosing a European region is a further guarantee, and the Data Privacy Framework is a valid instrument for transfers to the United States. The fact remains that a US supplier belongs to a jurisdiction where government power is, in the lawyer’s words, far more penetrating and pervasive. If an executive order requires an American company to hand over certain content, that obligation applies even when the data centres sit in Europe.

Region settles the location requirement. Sovereignty is something else, and it requires control of the perimeter.

Aidia’s perspective

During the conversation Claudio Sorbi described a request that comes up with some regularity: bring AI into the company covering every area, legal responsibility included. The answer was to stop. “We can support you at business and strategic level as consultants, but we are aware of what our competences and our limits are”. Hence the decision to widen the network, bringing in a specialist for each type of question.

This is not caution. It is the only way to avoid going back over something already built: at best you spend more and timelines stretch, at worst you rework the architecture of the solution.

A concrete example from the video illustrates the point. The conversational assistant for a food e-commerce platform has to handle allergens and intolerances. Which data to include in the training dataset, on what criteria, with what weight against other requirements: that is an ethical, legal and sector-specific assessment, which the developer then implements. Leaving it to whoever writes the code means moving a decision about responsibility onto someone without the elements to take it.

The reverse holds too, Ciracò notes: treating the software company as a pure executor of choices already made elsewhere produces the same kind of problem.

The same reasoning returns at the contractual stage, where the lawyer identifies a distinction that makes the difference when a client disputes the service: does the product guarantee a given result, or does it serve to carry out a given process? And in the first case, to what declared level of reliability? Technical annexes, specifications and requirement documents exist to answer that question before a lawyer has to ask it.

This way of thinking also shapes how AVA is built, Aidia’s proprietary AI platform: it runs on-premise or on a private European cloud, connects the company’s software ecosystem to any language model with dynamic switching, keeps data inside the perimeter and neither shares it with third parties nor uses it to train public models. Integration happens on the systems already in use, with no forced migrations.

To close, Ciracò recalls that the AI Act is human-centric legislation: the aim is the advancement of the individual within a context where they make use of a technological solution. It holds as a criterion even when compliance is not the subject.

Frequently asked questions

What is needed before introducing AI into business processes?

You need input data of verified quality, defined processes and governance that establishes who controls the outputs. Without these conditions the system amplifies existing disorder. You also need domain expertise able to validate the results, because a correct output and an out-of-context one look exactly the same.

Does Article 4 of the AI Act really mandate training?

Yes. It places on providers and deployers the obligation to ensure an adequate level of AI literacy among their staff and anyone using the systems on behalf of the organisation. The training must be graded against baseline skills and tailored to the specific role performed within the company.

Do data centres in Europe guarantee data sovereignty?

They guarantee location, not sovereignty. A supplier based in the United States remains subject to US jurisdiction even when the servers are European, and an executive order can require it to hand over data. Sovereignty requires control of the perimeter, not just of the geography.

Does AI-generated text always have to be disclosed?

Article 50 of the AI Act draws a distinction. Content that is generated but then subject to human review with editorial responsibility falls into a different category from output published without that control, which comes under the transparency obligations placed on the deployer.

Practical steps

  • Map the AI features already active: check which tools in use include generative features enabled by default, and what processing perimeter they involve.
  • Establish who validates the outputs: for each process, define which domain expertise checks results before they turn into decisions.
  • Read the contract before the demo: data portability, unilateral changes, applicable jurisdiction, as well as where the servers are located.
  • Separate result from process: clarify contractually whether the supply guarantees a measurable result or supports a process, and to what declared level of reliability.

If you are weighing up where AI is worth introducing and where good automation is enough, talk to our team.


Sources

  • Fondazione Leonardo ETS (2026). “Italy in the AI era”, presented to the Italian Chamber of Deputies on 19 March 2026 - corrierecomunicazioni.it
  • Quattrociocchi W. and others, Sapienza University of Rome. Research on epistemia and the limits of large language models - conf26.garr.it
  • Regulation (EU) 2024/1689 (AI Act), Articles 4 and 50 - eur-lex.europa.eu
  • “Intelligenza Artificiale in azienda: aspetti legali ed etici con l’Avv. Arianna Ciracò”, from the “Behind the prompt” series - Aidia YouTube channel
Marta Magnini

Marta Magnini

Digital Marketing & Communication Assistant at Aidia, graduated in Communication Sciences and passionate about performing arts.

Aidia

At Aidia, we develop AI-based software solutions, NLP solutions, Big Data Analytics, and Data Science. Innovative solutions to optimize processes and streamline workflows. To learn more, contact us or send an email to info@aidia.it.