How Kyrovane uses AI

• Answers questions about your workspace from your workspace: the facts are read at the moment you ask, inside your own permissions, and each answer lists the records it used. A question it has no data for gets “Kyrovane does not have that”, not a guess.
• Drafts messages, briefs, replies and reports for a person to read, change and send.
• Proposes next steps: creators to contact, follow-ups that are due, risks it noticed. A proposal is a suggestion with its reasons until a person approves it.
• Classifies what comes in, such as whether a creator’s reply is interested, declining or asking a question, with how sure it is.

• Send, approve, publish or spend on its own. Anything that leaves Kyrovane or changes money goes through the same approvals as a person’s action, and the workspace’s rules decide which need one.
• Change permissions, roles, settings or its own limits.
• Read data your role cannot read, or another workspace’s data. It uses the same row-level security as every screen.
• Invent figures. GMV, sales, follower counts and approvals come from the records that hold them or are not given.
• Take instructions from the content it reads. A creator’s bio, a message or an imported file is treated as data, so text in it cannot tell the assistant to do something.

When a question needs a language model, Kyrovane sends the provider the question and the facts needed to answer it, not your database. The provider is the one configured for the service (Anthropic or OpenAI); where none is configured, Kyrovane answers in a built-in mode that only assembles sentences from the facts it found and sends nothing out.

Your data is never used to train a model, by us or by the provider: every request is sent with the provider’s training opt-out, and the run record says it was. Kyrovane does not fine-tune models on customer data.

Every AI run is kept in your workspace: the question, the answer or the reason it declined, which provider and model answered, which records it used, the tokens it took and what it cost. Admins can see the workspace’s AI use and its cost in Kyrovane; the record is how “why did the assistant say that?” gets answered.

• Read a draft before it is sent, and own what you send.
• Treat a score, a forecast or a classification as an estimate with its uncertainty, not a promise.
• Use the AI within the acceptable use policy: no deceptive messages and no fake endorsements.

A language model can still word something badly or misread a reply. That is why everything it produces that matters waits for a person, and why its sources are shown. If an answer looks wrong, the run record shows what it was given; tell us at support@kyrovane.com.