Where we use artificial intelligence, and where we do not

Fymera One

At the bottom right of this page there is Fymera One: this site's assistant. It is automated, and it tells you so itself, first thing in every conversation — not because it is elegant, but because Article 50 requires it and because knowing what you are talking to is the least you deserve.

• It answers with what is on this site, not with what the model knows about the world. Services, projects, opening hours, contact details: if something isn't written somewhere here, for it that thing does not exist.
• When it doesn't know, it says so. It doesn't try to wing it: it passes the question to us and asks you for a contact. An invented answer costs more than an "I don't know".
• It writes nothing on its own. It can suggest a free slot for a call, but booking is your click on a button, and so is sending us a message. It doesn't decide, doesn't negotiate and doesn't commit us to anything.
• It only learns with a person in between. When it can't answer, the question goes onto a list. One of us writes the answer, one of us approves it, and only then does it join what the assistant knows. It does not learn from conversations by itself.

What you write passes through OpenAI, the model provider, which is based in the United States, and is not used to train anything. We keep conversations for ninety days; the IP address stays complete for thirty days, after which we zero out its last part. The full detail — legal bases, recipients, rights — is in the privacy policy.

Fymera One is also the name of the project through which we build assistants like this one for our clients: same approach, same limits, same principle that a person is the one who decides.

In our own work

We use artificial intelligence tools in parts of our work. Here is the list, with equal attention to what we do not do: that is the half usually left undeclared, and it is the half that actually matters to anyone trusting us with a project.

• sì Scrittura di codice — we use it. Coding assistants during development. Every line that reaches production is read and understood by a person: if none of us can explain what it does, it does not ship.
• sì First drafts of copy — we use it. Drafts of descriptions, emails and documentation. What we publish is rewritten and checked by us, and the editorial responsibility is ours.
• sì Internal analysis and summaries — we use it. Summaries of technical documents and meeting notes, for internal use.
• no Decisions about people — we do not use it. No automated system screens applications, assesses colleagues or decides anything about an employment relationship. People are chosen by a person.
• sì Automated replies to visitors — we use it. Fymera One is this site's assistant: it answers about services and projects, and can suggest a time for a call. It says so first thing in every conversation. When it doesn't know an answer it says so and passes the question on to us, rather than inventing one.
• no Negotiating or deciding in a person's place — we do not use it. Fymera One informs, it doesn't negotiate: it doesn't quote, doesn't agree prices and doesn't commit us to anything. And it writes nothing on its own — booking a call or sending us your details is always your click.
• no Content that could pass as genuine — we do not use it. We do not publish generated images, audio or video that could be mistaken for real footage of people, places or events. Project renders are labelled as renders.

In client projects

When we build an automation or an assistant for a client, five rules apply. They are not a moral commitment: they go into the project document, and anything not written there does not get built.

• Client data trains nothing — The data you entrust to us is not used to train models, ours or anyone else's. It is in the contract, not a verbal promise, and we choose providers based on what theirs say.
• L'assistente dice cosa è — Every system that talks to a person declares that it is automated at the first exchange. It has been a legal requirement since 2 August 2026, and it is the first thing we write anyway.
• Generated content is marked — Content produced by a generative system ships with the machine-readable marking the regulation requires.
• There is always a human checkpoint — In processes touching money, contracts or people, the automation prepares and proposes; a person confirms. This is not caution: it is the point where a mistake becomes visible before it becomes a problem.
• It can be switched off — Every automation we deliver can be turned off without stopping the rest of the system. An automation you cannot switch off is a constraint, not a tool.

What changed on 2 August 2026

From that date the transparency obligations of Article 50 of the European AI regulation apply. They cover those who build AI and — this is the part that catches people out — those who merely use it. In short:

• a system that converses with a person must declare that it is automated, clearly, at the first exchange at the latest, unless it is already obvious from context;
• content generated by a generative AI system must be marked in a machine-readable format;
• anyone publishing a deepfake must declare that the content is generated or manipulated, perceptibly;
• anyone using emotion recognition or biometric categorisation systems must inform the people affected.

Systems already on the market before 2 August 2026 have until 2 December 2026 to comply. Penalties for breaching the transparency obligations reach 15 million euro or 3% of worldwide annual turnover — whichever is higher — with reduced caps for small and medium enterprises.

The high-risk obligations in Annex III follow later: 2 December 2027, and 2 August 2028 for systems embedded in already-regulated products. They do not cover the kind of work we normally do, but if a project fell inside them we would say so before starting it, not after.

Legal references

Who is accountable

Fymera Srl is answerable for the choices described here. For questions about this page, or to work out whether one of your systems needs declaring, write to contatti@fymera.it: a person replies.