GabrySolution at Aidia: is there an Italian way to AI?

GabrySolution’s interview with Aidia

Gabriele Venturini in Florence to interview the Aidia team and founders: is there an Italian way to AI?

Is artificial intelligence a bubble? Luca Angioloni, CTO and co-founder of Aidia, answers like this: “I believe that, for obvious reasons, it started out as a bubble.” Getting this far, he adds, has required investments that, “if you put them down in numbers on a sheet of paper, it’s madness.” But for Luca it is a bubble that is already producing real value, because the technology is real. The question is where that value actually lies.

It is one of the questions put to us by Gabriele Venturini, a computer engineer and entrepreneur known online as GabrySolution. On his channels he talks about data-driven marketing, social media and artificial intelligence, and he came to Florence to interview our team and founders. His starting point, in his own words: “In Italy there is someone who is really doing AI. I went to see who.”

For us, this conversation is an important milestone. It confirms the interest in artificial intelligence built on real problems, already at work in factories and in our clients’ systems. It also gave us the chance to explain that approach to his audience, answering precise and sometimes uncomfortable questions. In the background was a broader question: is there an Italian way to AI? We believe there is, and it is vertical artificial intelligence, built around a single process and integrated into the systems a company already uses.

The full interview is on our YouTube channel: watch the video.

Where does AI create value in a factory?

On a food packaging line, a defective piece can let bacteria in. “These are contexts where errors are not acceptable,” says Luca. For years the only safeguard was sample-based inspection: when a defect was found, an entire production batch was thrown away. Today, with the technologies we bring to companies, every single piece can be inspected in real time, while the machine is running, and defective ones can be rejected immediately. Luca lists the effects: better production quality, no machine downtime, lower costs and less waste.

Doing this takes the right technology. In this field, Luca explains, “large language models make no sense at all.” The system has to perform around sixty checks per second, processing images and sensor data. A language model is much slower, and at that pace “you get nowhere.” What you need are specialized neural networks. This is what Aidia Control does: our deep learning computer vision system connects to the machine and rejects defective pieces without stopping the line.

That the right technology matters has not always been obvious. At first the problem was trust. AI sounded like science fiction, and Riccardo Celli, CEO and co-founder of Aidia, recalls that clients always asked the same thing: “How can I be sure whether it works or not?” With the arrival of ChatGPT, the situation flipped. Luca remembers a phase when everyone wanted AI everywhere, in every sector, and often with the wrong technology: companies asking for a language model to solve problems that needed a vision model. That is why part of our job is to say when a technology is not needed, or when a different one is.

Choosing the right model, however, requires skills that many companies do not have in-house. According to the Italian National Institute of Statistics (ISTAT), in 2025 15.7% of Italian SMEs with at least 10 employees used at least one AI technology, compared with 53.1% of large enterprises. Among companies that considered AI without adopting it, a lack of skills is one of the main obstacles.

Why can a generative model be so confidently wrong?

It comes down to how it is trained. “The training target of these models is to be plausible in their answers, not for the answer to be correct”, Luca explains. Pre-training, the phase that takes up around 90% of the computing power needed for training, teaches the model to reproduce text in a statistically accurate way. But plausible text can still be wrong in substance.

Luca also points to two structural limits. Models are static and have no real memory: what looks like memory is the context the user provides again with every request. And they only know the world through a representation of it, made of text or images, without the physical understanding that a seven-year-old child already has.

They remain useful tools, as long as they are used where answers can be checked and context can be provided. AVA, our proprietary AI platform, connects the language model to a company’s documents and systems, on-premise or on a private European cloud, and makes it possible to switch models without depending on a single provider. The final check stays with the people who know the process.

What changes for developers, and for those just starting out?

Alessandro Minervini, machine learning engineer and project manager at Aidia, uses AI to do in four hours the programming work that used to take him eight. Even with vibe coding, software development handed to AI starting from a description, it only works, he says, “if you know what you want to do.” Otherwise the AI proposes a solution different from the one you needed, and a lot of time goes into correcting it before you get the result. Bernardo Menicagli, software architect, sees AI as a tool that will be used more and more and that will become ever more embedded in the software development workflow. With one clear limit: “there will always be a need for human oversight behind it.”

Both positions lead to the same point. AI speeds up those who already know how to do the job and leaves those still learning more exposed. Carolina Magrini, head of marketing, puts it plainly: “Junior roles will pay a bit more of the price.” She is also convinced that the change will open up other opportunities. For companies, one question remains open: if AI now does the tasks through which juniors used to learn the trade, how will tomorrow’s seniors be trained?

What is Italian about AI built this way?

For Carolina, the answer is cultural. Food, fashion, automotive: the sectors in which Italy is recognized worldwide share quality, customization and craftsmanship. Applied to technology, they become what we call “Italian artisanal tech.”

Luca chooses a tailoring image. The sewing machine made the work faster, and often better, but the value of a garment still lies in its design, whether by Gucci or Brunello Cucinelli: “the machine alone doesn’t invent it, doesn’t come up with it, doesn’t know what it will want to make.” Looking ahead, he believes AI will become the operating system of work, an orchestrator that calls on tools written by developers for specific functions. If that happens, we believe the advantage will go to those who have built the best tools around real processes.

The limit of this path is financial. The research is there, and Carolina cites the University of Pisa as an example. What is missing is capital: in Italy, Luca observes, companies are less inclined to take on debt and prefer to grow with a safety net, while in the United States you start with investors and figure out the rest later.

What remains if the bubble bursts?

Luca does not deny that AI is still a bubble: “it was a bubble and it still is today.” It was the bubble itself, he explains, that created the market and the demand and produced value. It remains to be seen whether it will burst or hold until it is no longer a bubble. On one point, however, he has no doubts: the technology “is real, it is useful, we will all use it,” and it will remain, “it doesn’t disappear, whether the bubble bursts or not.”

There is already a precedent. The dot-com bubble of the early 2000s burst, but the internet stayed and became part of everyday work. As we see it, this time too what will remain above all is AI built on real processes: the inspection of every piece on a production line, vertical tools tailored to a company’s systems.

Frequently asked questions about the Italian way to artificial intelligence

Is there an Italian way to artificial intelligence?

For Aidia, there is. It is vertical AI, built around a specific process and integrated into the systems a company already uses, such as ERP, MES or production lines. It draws on the customization and craftsmanship of the sectors in which Italy is recognized worldwide, such as fashion and automotive. Its main limit is the difficulty of raising capital to grow the companies that develop it.

Do you need a language model for quality control in a factory?

Usually not. Quality control on high-speed lines requires specialized computer vision systems, able to analyze dozens of images per second and reject defective pieces in real time. Language models are better suited to documents, data and management software, where inspection speed is not the main requirement.

Why does generative AI sometimes give wrong answers?

Because it is trained to produce statistically plausible text, not to verify its content. A confident, well-written answer can therefore be wrong. Using it in a business requires context, meaning access to the right data and documents, and a check by people who know the process.

Where to start when evaluating AI for your company

• Choose the model based on the task: computer vision for high-speed inspections, language models for documents, data and management software.
• Give AI the right context: without clear goals and relevant data, the model’s answers become random.
• Check the answers: a plausible generative output must be reviewed by someone who knows the process.

If you want to find out whether and where AI can help your processes, talk to the Aidia team.

Sources

• ISTAT (2025). “Imprese e ICT, anno 2025” (Enterprises and ICT, 2025) - istat.it
• Assolombarda, Research Centre (2025). “Utilizzo dell’Intelligenza Artificiale nelle imprese” (Use of Artificial Intelligence in businesses), based on ISTAT data - assolombarda.it