AI transparency

• Search indexing: email text is converted into embeddings (numeric representations), so you can search by meaning.
• Answers: a language model writes an answer to your question from the emails the search retrieved.
• Drafts: a language model drafts a reply for you to review.

Embeddings and, by default, answers and drafts are generated with OpenAI models. Anthropic models can be selected instead for answers and drafts. Enterprise customers can use their own provider accounts (BYOK). The providers act as our subprocessors; see subprocessors.

Answers and drafts are marked as AI-generated in the app. Each answer cites the emails it is based on, and you can open every cited email.

Quri first searches your mail with hybrid search, which combines meaning with exact keywords such as invoice numbers. The model then writes an answer from the retrieved emails and cites them.

• Answers can be incomplete or wrong, for example if a relevant email hasn't been indexed yet or is outside your plan's history window.
• The contents of attachments are not searched.
• The model can misread tone, dates or context. Check important answers against the cited emails.

Drafted replies are never sent automatically. You review and edit each draft and decide whether to send it through your own mailbox.

Quri does not train AI models on your email. OpenAI does not use data sent through its API for training. Where configured, OpenAI processes requests with EU data residency.

Quri does not make decisions about you or other people that have legal or similarly significant effects. It is a tool that helps you find and understand your own email.

Write to us at support@askquri.com.