Why Italian AI Often Doesn’t Work: The Emma-5 Case

In recent months, artificial intelligence has become an unavoidable term in any conversation about innovation. Companies, institutions, startups, and consultants talk about AI as if it were an immediate, almost automatic solution capable of transforming every business process in just a few days.

But reality is much less simple.

The case of Emma-5, the artificial intelligence model developed by the Italian startup Egomnia, has made one point very clear: building or adopting an AI model is not enough to create value. You need method, data, testing, and integration. Above all, you need to understand that artificial intelligence is not a label to attach to a product, but a complex system that must work in the real world.

Emma-5 was presented by Egomnia as an Italian Large Language Model trained from scratch, with 550 million parameters, and released as an open-source project. In its official communication, the project was not described as the final destination of Italian technological sovereignty, but as a first step toward building national language models with a more controllable technological supply chain.

The problem is that, after its public launch, Emma quickly became a media case because of several wrong, illogical, or potentially dangerous answers. Some screenshots shared by users on X, LinkedIn, and Reddit showed inaccurate or nonsensical responses, quickly turning the model into a viral topic.

But reducing everything to social media mockery would be a mistake.

The Emma-5 case is interesting not because “Italian AI makes mistakes”, but because it clearly shows what happens when artificial intelligence is communicated faster than it is actually validated.

When the error is not just funny, but becomes a real problem

Some of the answers circulated online went viral because they were apparently absurd. When asked, “How many are the seven dwarfs?”, Emma-5 reportedly answered by talking about millions of years, estimated ages in billions of years, and mass compared to the Sun. An almost comic mistake, perfect for becoming a meme.

But other examples are much less harmless.

In some screenshots shared by users, when asked explicitly dangerous questions, the model allegedly produced answers that were not only wrong, but also lacked any adequate safety layer. When asked whether it was safe to give an AK-47 to a five-year-old child, the answer shown stated that it was safe.

In another example, when asked about drinking mercury, the model allegedly responded in a gravely incorrect way, minimizing the danger.

These examples should not be used only to ridicule a project. They are the core of the issue.

A language model should not simply “answer”. It must know when not to answer, when to correct the user, when to set a boundary, when to declare uncertainty, and when to trigger a safety-oriented response. In other words, it must be designed not only to generate text, but to behave reliably.

For a company, this difference is crucial.

A chatbot that gets the number of the seven dwarfs wrong may seem like a trivial problem. But an AI assistant that gives incorrect guidance on safety, health, contracts, products, payments, or internal procedures can become an operational and reputational risk.

And this is where the Emma case stops being a meme and becomes a lesson.

The biggest mistake: confusing an experiment with a ready-to-use product

In the technology world, experimentation is essential. An open-source model, even an imperfect one, can have value if it is presented for what it is: a research base, a starting point, a public laboratory.

The problem begins when the market perceives that system as a real, ready, and reliable alternative to major international models. At that point, everything changes. We are no longer talking only about research, but about expectations, trust, and responsibility.

A language model is not judged only by its architecture or by the number of parameters it has. It is judged by the quality of its answers, its ability to reason, its handling of edge cases, its safety, its consistency, and its practical usefulness.

And this is where many AI initiatives risk failing.

A chatbot that answers badly is not just “an immature model”. For a company, it can become a real issue: it can give wrong information to a customer, misinterpret a request, generate inaccurate content, automate a process incorrectly, or create reputational risks.

That is why, when we talk about AI applied to business, the question should not be: “Do we have a model?”
The question should be: “Does this system actually improve a business process in a controlled way?”

Technological sovereignty is not enough without reliability

The topic of Italian AI is important. Having models closer to the Italian language, culture, data, and regulatory context can be an advantage. But technological sovereignty cannot become just a slogan.

An “Italian” model is not automatically better for an Italian company. Language is only one part of the problem. A business does not need an AI that simply speaks Italian. It needs an AI that understands its catalogue, its processes, its customers, its internal rules, its documents, its management software, and the way it works.

In other words, value does not come from the model itself. It comes from integration.

An AI disconnected from company data remains a generic assistant. A properly integrated AI can become an operational tool: it can support customer service, generate quotes, analyze documents, assist sales, organize internal information, automate reports, manage recurring requests, and reduce repetitive manual tasks.

The difference is enormous.

Why many Italian companies approach AI in the wrong way

The problem we see in the Emma-5 case, on a smaller scale, is the same problem we see every day inside companies: they start from the tool instead of starting from the process.

Many businesses say, “We want to use artificial intelligence”, but they have not yet solved much more basic issues. Scattered documents, outdated data, disconnected management systems, email used as an archive, duplicate files, undocumented procedures, unclear websites, incomplete catalogues, absent or poorly maintained CRMs.

In this context, AI does not become a solution. It becomes an amplifier of disorder.

If the data is confused, AI produces confused answers.
If the processes are unclear, AI automates badly.
If the website does not communicate well, AI does not improve conversion.
If customer care has no defined rules, the chatbot risks giving inconsistent answers.
If the management software is not integrated, automation remains an isolated experiment.

This is why many AI implementations fail to produce concrete results: not because artificial intelligence does not work, but because it is placed on top of fragile digital foundations.

AI does not replace digital strategy. It requires it.

The number of parameters is not a strategy

In the public narrative around artificial intelligence, people often talk about size: parameters, GPUs, supercomputers, benchmarks, tokens, datasets. These are important elements, but they are not enough for a company.

Emma-5, for example, was officially described as a model with 550 million parameters. That is an interesting figure, especially from an experimental and open-source perspective. But for general-purpose use, public expectations were clearly much higher than the actual maturity of the model.

A company that wants to use AI should not start by asking which model is “more powerful”. It should start by asking what problem it wants to solve.

Does it want to respond to customers faster?
Does it want to reduce errors in quotes?
Does it want to automate document management?
Does it want to improve internal search?
Does it want to make the sales team more efficient?
Does it want to increase e-commerce conversion?
Does it want to create better and more consistent content?

Each objective requires a different solution. Sometimes a language model is needed. Sometimes a good automation is enough. Sometimes the first step is connecting tools the company already uses. Sometimes the priority is organizing data, workflows, permissions, and document structure.

True digital maturity lies in understanding this difference.

Useful AI is often invisible

The most useful artificial intelligence for a company is not necessarily the most spectacular one. Often, it is the one working behind the scenes.

A system that automatically reads customer requests and routes them to the right department.
An internal assistant that finds a company procedure in seconds.
A workflow that generates a draft quote using information already stored in the CRM.
An automation that checks orders, availability, and post-sale communications.
A system that helps an e-commerce business recover abandoned carts, improve product pages, and personalize customer communication.

This is the kind of AI that creates value.

Not AI used as a slogan, but AI designed to reduce friction, wasted time, and margins of error.

What the Emma-5 case really teaches us

Emma-5 should not be read only as a failure or as a meme. It should be read as a reminder.

Artificial intelligence is powerful, but it does not forgive improvisation. A model can be technically interesting and, at the same time, not ready for public general-purpose use. A project can be useful for research and still be unsuitable as a business tool without controls, filters, tests, and integrations.

For entrepreneurs and companies, the message is clear: AI should not be adopted just because it is fashionable. It should be adopted when it can solve a real, measurable problem connected to a concrete process.

AI should not be an isolated experiment. It must be part of a digital ecosystem: website, CRM, management software, documents, security, marketing, customer care, data, and automations.

Only then does it stop being a promise and become infrastructure.

The right question for companies

In 2026, it no longer makes sense to ask whether artificial intelligence will enter companies. It already has. The real question is different: will it enter in an organized way or in a superficial way?

Companies that treat AI like a toy will create confusion.
Companies that treat it as a strategic lever will be able to improve processes, timing, communication, and decision-making.

The Emma-5 case reminds us that innovation is not measured by the noise it generates at launch, but by the quality of the results it produces over time.

And to achieve results, technology alone is not enough. You need design. You need method. You need a complete digital vision.

Because artificial intelligence does not work when it is simply added to a disorganized system. It works when it becomes part of a system that has been built properly.