AI data use
An AI-native helpdesk processes customer data with language models. This page discloses which feature sends which data where – and what is guaranteed in the process.
Five principles
Training on customer data is contractually excluded and technically disabled with every AI provider (no-training guarantee).
Before text enters AI search or context examples, we strip IBANs, card numbers, credentials, tokens and names.
Language models run in the EU by default (Paris, zero data retention – inputs are not stored there). US frontier models are a Business-plan option, never the default.
Monthly budget, round limits and a kill switch per workspace – reached means: the AI pauses, the helpdesk keeps running.
Per Art. 50 EU AI Act, the chat widget, web form and bot emails mark AI use automatically – with a clear path to a human.
Which feature processes which data?
| Feature | Data processed | Processing |
|---|---|---|
| Autonomous replies & drafts | Ticket thread, matching knowledge-base articles, similar resolved tickets (redacted) | EU mode (default) or the selected model mode; no storage at the provider |
| Guardrail checks (competence & risk) | Excerpt of the current ticket | AI classification before every autonomous reply |
| Screenshot analysis | Image attachments of the current ticket | Only when the bot requests it for the case at hand |
| Sentiment detection | Text of inbound messages | Prioritises upset customers, hands over to humans earlier |
| Search & similarity (embeddings) | Redacted text excerpts from tickets and knowledge base | Scaleway (Paris, EU); vectors live in our database in Germany |
| Knowledge-base translation | Article texts (no customer data) | On demand, never automatic |
Who are the providers?
All AI providers, their role and the legal basis are listed in our public subprocessor list – including a 30-day notice period for changes.
What never happens
- No training of AI models on your data – with any provider
- No selling, no advertising, no sharing with third parties outside the subprocessor list
- No permanent storage of your content at model providers (zero data retention or privacy mode)
- No AI reply without guardrail checks – money, legal and credentials are decided by a human
Questions? Write to datenschutz@golle-it.de – or read the technical details on the security page. Security & privacy →