AI and models

When you chat with nerix.ai, an AI model writes the answer. It is not a person. The website, the product and the demonstrations say so, and every answer in the product shows that it was produced by a model.

A request is answered by one model from one provider. Depending on what your organisation allows, the provider can be:

• Mistral AI, hosted in the European Union. This is the default.
• OpenAI.
• Anthropic.
• Google.
• DeepSeek. Its models process data in the People's Republic of China. The product prints a caution about this where you choose it.
• Soniox, for speech recognition (turning speech into text).

Your organisation decides which providers are allowed, and can limit the choice to providers hosted in its own data region. Only the passages needed to answer are sent, and only for that conversation. The sub-processors page lists each provider and where it processes data.

Each answer names the model that actually produced it. When the answer comes from your documents or from the web, the sentences carry numbered citations that open the passage they came from, so you can check them. If nerix.ai cannot find support for something, it says so instead of guessing.

AI models make mistakes. They can misread a document, leave something out, or state something that is not true with confidence. Check important answers against their sources, and do not rely on nerix.ai as legal, medical, financial or other professional advice. The terms of service say the same.

Your documents, conversations and Sheets are never used to train any model, by us or by a provider. In the product this reads "Never" and cannot be changed.

Answers and drafts that a model produced are shown as AI output, with the model's name. Demonstration answers on this website are labelled as coming from an example workspace and are not real customer content. Where the law requires AI-generated material to be marked, we mark it. If we add machine-readable marking of generated files, this page will describe it.

PII redaction before text reaches a provider (optional, on by default), a record in the audit log of who asked what and what it cost, and a picker that prints each provider's cautions at the same weight as its strengths.

If an answer is harmful, wrong in a way that matters, or looks like it leaked something it should not, write to info@nerix.ai with the conversation link or a description and the time. A person will read it.

See the privacy policy for how personal data is handled and the sub-processors page for the full list of companies that process data. The model catalogue lives at /models.