Artificial intelligence transparency

In the event of any discrepancy, the Italian version prevails. Read the Italian version

What AI does in Graphorum

Graphorum uses artificial intelligence models to read and sort email, prepare replies on email, WhatsApp and Telegram, summarise documents, remember facts with their source, coordinate agents, speak by voice and prepare social media content. Under art. 3 of Regulation (EU) 2024/1689 (“AI Act”), Graphorum is an AI system:

• AM Solutions is the provider: it develops it and places it on the market under its own name;
• the Customer is the deployer: it uses it under its own authority, for its own business.

Which models it uses

• Local models: they run on the Customer's mini PC or in its cloud instance. They are the standard engine.
• Your model: a model trained only on the Customer's data and only for the Customer, which grows up to 2.5 billion parameters, reasons, sees images, uses tools and speaks several languages. It may derive from a base model published under an open licence, named on its card, which also explains how it was trained.
• External AI services: optional and off by default. They are used only if switched on, job by job, with providers that do not use the data to train their own models. The list is in the data processing agreement.

Graphorum is not a general-purpose AI model within the meaning of Chapter V of the AI Act: customers' models are specific adaptations, for a single customer, and are not made available to others.

Models and licences

The local models are published by their authors under open licences, which Graphorum respects. These ask that their creators be credited:

• The PC's voice: Pocket TTS by Kyutai, code under the MIT licence, weights under the CC BY 4.0 licence. Graphorum uses the public weights without voice cloning, exactly as Kyutai distributes them.
• The Pocket TTS voices: Kyutai's ready-made voices, derived from openly licensed recordings: the VCTK corpus of the Centre for Speech Technology Research at the University of Edinburgh (CC BY 4.0), Alba Mackenna (CC BY 4.0), CML-TTS (CC BY 4.0), Mozilla Common Voice, Kyutai's Unmute project and LibriVox (CC0). Jürgen, Rafael and Estelle are distributed by Kyutai without stating their origin. Voices under a licence that forbids commercial use are not included.
• The natural voices of Supertonic 3 (optional, downloaded only if you choose them): Supertonic 3 by Supertone, code under the MIT licence, weights under the OpenRAIL-M licence, used exactly as Supertone distributes them.
• Usage limits of Supertonic 3: the OpenRAIL-M licence forbids using the model and the voices it creates to break the law; to exploit or harm minors; to create or spread false information in order to harm others; to spread personal data that can harm a person; to defame or harass; to impersonate another person; to spread generated content without clearly stating that it was generated by a machine; for fully automated decisions that affect people's rights; to discriminate against people or groups or exploit their vulnerabilities; to give medical advice; for justice, law enforcement or immigration purposes. Whoever uses Graphorum accepts these limits and makes those who receive the audio respect them. The licence text, which comes with the model, is authoritative.
• The other natural voices (optional, downloaded only if you choose them): Chatterbox by Resemble AI, code and weights under the MIT licence, which adds an invisible mark (Perth) to every audio so it can be recognised as generated; Qwen3-TTS by Alibaba Qwen and VoxCPM2 by OpenBMB, code and weights under the Apache 2.0 licence. Graphorum uses them exactly as their authors distribute them, without cloning the voice of real people.

Risk level and uses to avoid

Graphorum is designed as an assistant for office work by companies, professionals and individuals and is not intended for the high-risk uses of Annex III of the AI Act, for example:

• recruitment, assessment of or decisions on workers' promotions, dismissals and task allocation;
• assessment of people's creditworthiness or setting of life and health insurance premiums;
• access to essential public services, education, justice, migration;
• biometric identification or emotion recognition.

If you want to use it for one of these purposes you must tell us beforehand: specific assessments and obligations and a dedicated agreement are required. Graphorum may not be used for the practices prohibited by art. 5 (manipulation, exploitation of vulnerabilities, social scoring and the others listed in the regulation).

Transparency towards people (art. 50)

• People who write to the Customer know that an AI is replying. When an agent replies on its own by email, WhatsApp or Telegram, Graphorum adds the company's “Graphorum, AI assistant” signature by default. The Customer can turn it off in Settings; if it does, it must inform people in another way, unless it is obvious from the context.
• Generated content. Text, images, audio and video created with AI are recognisable as such: AM Solutions applies the marking required by art. 50(2), using the technical solutions available. When the Customer publishes images, audio or video that look real (deepfakes), or texts on matters of public interest that have not been reviewed by a person, it discloses this (art. 50(4)).
• Voice. Graphorum's voice is synthetic: it must not be used to make someone believe they are talking to a person.

Human oversight

• The Regulator: every structured AI answer is checked by code (dates, amounts, emails, phone numbers, VAT number, tax code, names found in the documents read), corrected or discarded. If no model is sure, Graphorum doesn't make things up: it asks.
• Autonomy: the Customer decides, channel by channel, whether the AI stays off, prepares drafts for approval or replies on its own.
• Approvals: actions that cannot be undone go through a person, unless the Customer leaves them unrestricted.
• Log: every action by users and agents is recorded, with who, what, when and from where.

What the Customer must do

• Use Graphorum in accordance with these indications and the instructions for use, and monitor its operation.
• Ensure the AI literacy of the people who use it (art. 4). On activation, AM Solutions provides basic training and materials to keep.
• Keep the notice on automatic replies active or inform people in another way (art. 50).
• If it is an intellectual professional, inform its clients of the use of AI systems, as required by the Italian AI law (Law No. 132 of 23 September 2025).
• Report seriously wrong results or unexpected behaviour to us immediately.

Data and training

Each Customer's model is trained only on that Customer's data and only that Customer uses it. Customers' data does not train the models of other customers, of AM Solutions or of third parties. The Customer can download its model at any time and delete it together with its data. The protection of personal data is described in the privacy policy and in the data processing agreement.

The AI Act timeline

In November 2025 the European Commission proposed (the “Digital Omnibus” package) postponing some deadlines for high-risk systems. This page follows the text in force on the date of the last update and will be updated if the deadlines change.

Questions and reports

For questions about Graphorum's AI or to report a problem, open a request from the support form or write to [email protected].

Version 1.2 of 16 October 2026 · AM Solutions di Vera Alexandru Nicolae. Earlier versions can be requested by email.