How the AI works, and what it will not do
A plain account of what the model does with your data, where it can be wrong, and what we refuse to claim. Last updated 15 September 2026.
These are drafted documents, not legal advice. They have not been reviewed by a lawyer. If you are signing the data-processing agreement on behalf of a company, have your own counsel read it first.
What actually happens to a message
A customer's message is embedded so the service can find the passages in your knowledge base most likely to answer it. Those passages and the message go to a language model, which writes a reply. The reply, the documents it cited and the revision of each are recorded.
Training
We do not train models on your data, and we do not build a shared model across customers. We use our providers' APIs; their own terms govern what they do with what we send, and they are named in the data-processing agreement. If you need a written commitment from a specific provider, tell us which and we will point you at their terms rather than paraphrase them.
Where it can be wrong
A language model can produce a fluent answer that is not supported by any document. This service is built to make that visible rather than to promise it cannot happen: every answer records the document and revision behind it, and how often the agent answered without a source is a number on your dashboard and on your trust page.
It can also be confidently wrong about something a document does say, and retrieval can miss the passage that would have answered. Neither is eliminated by anything here.
A person is always reachable
When the knowledge base cannot support an answer, the agent hands over rather than guessing. You configure when that happens. No outcome that costs a customer money is taken by the agent alone - refund requests are filed for a person to decide.
What we refuse to claim
We do not publish a resolution rate we define and are paid on. Resolved, abandoned and escalated are counted separately, ambiguity resolves to abandoned, and a rate is withheld below 30 conversations. We report an overlap as an overlap: the service will not tell you a silence cost you three customers, because it cannot know that.
Telling your customers
The widget identifies itself as an automated agent. You should not configure it to claim to be a person, and several jurisdictions now require that disclosure.