How Folgor uses artificial intelligence, which systems are involved, where people stay in control and what you can do. Written in view of the EU Artificial Intelligence Act (Regulation (EU) 2024/1689).
1. Our approach
Folgor uses artificial intelligence to read documents, check purchases and draft answers faster and more consistently. AI supports the service; it does not replace the people responsible for decisions.
Folgor does not train its own AI models: it integrates general-purpose models from external providers into its platform.
2. Models and providers
| Provider | Models | Used for |
|---|---|---|
| Anthropic | Claude Sonnet | Reading purchase documents and photos, describing the photos customers send in the chat, reading scanned or illustrated manual pages and their figures, the customer assistant, the portal assistant, consolidating the knowledge bases, rewriting the answers of the company's specialists for the customer with the safety warnings, rendering safety texts (warnings, steps, stop protocols) in the reader's language, the Repair agent for Folgor technicians, describing repair photos, drafting technician knowledge in English |
| Anthropic | Claude Opus | Second opinion when an automatic check is uncertain |
| Anthropic | Claude Haiku | Classifying documents, reading manual pages, checking content read from images against the page, translating rule labels, writing the reason when the assistant closes a request, reading the shipping address of a replacement from the purchase document, checking each render of a safety text against its English source, checking that a reply gives no step the safety runner has not released, checking technician knowledge back against the technician's words, translating interface texts generated in English when they are shown, writing the English names of spare parts and checking them back |
| Gemini Embedding | Search in the knowledge bases | |
| Gemini Flash with Google Search | Searching online for solutions, without personal data, where the company's procedure allows it, and in the repair lab once the question to the company's technical contact (or to a Folgor specialist) has ended | |
| OpenAI | GPT-Live | The voice conversation on the support page and in the repair lab, where offered: it listens and speaks, while the answers, the safety warnings and the steps come from the assistant; safety texts are read word for word and checked against the written text, and the voice stops if they differ. Our page asks OpenAI not to store the voice conversation (store: false); OpenAI's own data-retention policy still applies to what it processes |
The models are reached through Vercel AI Gateway. Providers process data under their commercial terms and the safeguards described in the privacy notice.
3. Where AI is used
- Product documents: type, language, version and title of each document uploaded to the portal.
- Support requests: reading invoices and receipts (seller, date, products, sales channel, buyer type, warranty and withdrawal clauses), comparing product photos and labels with the purchase, recognising the product from the photos of the product and its label when the customer has no purchase document, recognising recall codes.
- Knowledge bases: extracting procedures, warnings, facts and figures (illustrations, exploded views, safety pictograms, described in English) from manuals, page by page, with checks against the original text and, for content read only from images or scanned pages, against the page image by a second model.
- Assistants: answering customers on support pages and company users in the portal, from approved sources only; on support pages, proposing a closing activity allowed by the company's procedure and closing a request as resolved or no longer active, with its reason, showing figures of the documentation and describing once each photo the customer sends (personal data filtered out; the company sees the description); in the portal, preparing a report to Folgor that is sent only after the user confirms it (screenshots attached to a report are never sent to the model).
- Repair lab: the Repair agent assists Folgor technicians on each repair (documentation search, guided procedures with their safety warnings, parts, findings, status changes it proposes and the technician confirms); each repair photo is described once without personal data; the agent drafts technician knowledge in English, which a second model checks against the technician's words before the technician confirms it.
- Safety texts: warnings, steps and stop protocols are kept in English and rendered in the reader's language by one model; a second model checks each render against the English, and a render that does not pass is never shown: the release is held and a person takes over (for a stop protocol, a fixed emergency text is shown instead). A second model also checks that no reply gives a step the safety runner has not released.
- Translations: interface texts generated in English (for example reasons, summaries and diagnoses) translated into the reader's language when shown, never safety texts; labels of custom procedure rules, translated from Italian into English and Chinese when a Folgor admin asks; the English name of each spare part, checked back by a separate call before it is stored.
4. Transparency towards people
People who write to the assistant are told that it is an automated system and that its answers are general information and may contain errors. When it proposes a closing activity, the page says who decides it; every closed request comes with its outcome and reason by email.
The support form states that documents, photos and messages are checked with automated systems, including AI.
5. Human oversight
- When a check is uncertain, even after a second opinion, the request is not decided automatically: it goes to the company's team with the reasons.
- If an automatic check fails, the request is marked not eligible with its reason; the company's staff or Folgor can review and reverse the outcome.
- Corrections made by people, for example on the seller or the sales channel, are never overwritten by AI.
- Doubtful knowledge-base items wait for a decision by the company; custom procedure rules are approved by Folgor; commercial policies count only after the company approves them.
- A closing activity proposed by the assistant is decided by the company; Folgor accepts it at once only where the company's procedure says so and documented purchase terms cover it. A rejection is always decided by a person.
- In the repair lab the Repair agent only proposes status changes, which the technician confirms; technician knowledge is saved only when the technician confirms it, is reviewed by Folgor staff and reaches customer support only after the company approves it.
- You can always ask for a person to review your request by writing to the company named on the support page or to info@folgor.com.
6. What AI does not do
- It does not decide refunds, returns, replacements, shipments, repairs or rejections by itself: it can only propose them, as described above.
- It does not record withdrawal declarations: only the customer's own action does.
- It contacts no one on its own: it asks the company's specialist or searches online only where the company's procedure provides it (in the repair lab, it searches online only after its question to the company's technical contact or to a Folgor specialist has ended), and asks a Folgor specialist only when a person must take over and no one at the company can answer; it never forwards a request by email, and promises no other follow-up.
- It does not show repair content to customers.
- It is not used for profiling, emotion recognition, biometric identification or decisions on people's access to essential services or employment.
7. Risk classification
Based on our assessment, these uses are not high-risk AI systems under Annex III of the AI Act and are not prohibited practices under Article 5. The assistants interact with people: the transparency obligations of Article 50 apply and are met as described above.
We review this assessment whenever we add or change AI functions.
8. Data
What AI reads and how long data are kept are described in the privacy notice. Prompts, answers and the contents of invoices are not written to our error logs.
9. Questions and complaints
For questions about AI at Folgor write to info@folgor.com. You can also contact the data protection authority of your country.