Kai is the AI agent that answers your customers. Understanding one rule explains most of its behaviour: Kai answers only from content it retrieves from your knowledge sources at the moment of asking. It does not answer from what the underlying model learned during training, and it does not treat its own earlier replies as facts.
For each new factual question, Kai searches your knowledge sources again — even when the same topic came up earlier in the same conversation. This is deliberate. An answer Kai gave three messages ago might have been incomplete, so reusing it would compound the error instead of correcting it.
A single turn is limited to three searches. For a question with several parts, Kai searches each part once and may rephrase a search once if it comes back empty. After that it answers the parts it can support and says what it couldn't find, rather than searching indefinitely. If you have ever wondered why an AI agent takes a minute to say nothing useful, this limit is what prevents it.
It never fills the gap with a guess. Which of the following happens depends on the question and on how you configured the answer step in your workflow.
The question describes a bug: Kai offers a button to file a bug report, pre-filled with a summary of the problem. This is on unless you switch it off.
The question is a feature request, or asks whether something exists: Kai offers a button to submit it as a feature request. Off unless you turn it on.
The matter is genuine support, billing, account, order, refund or cancellation work that Kai can't resolve — or the person simply asks for a human: Kai hands the conversation to your team. Off unless you turn it on.
In every other case, Kai says it couldn't find an answer and invites another question.
Four narrow cases: greetings and thanks, questions about who it is, closing messages when someone says they're done, and asking for a detail it needs before it can act. Everything else goes through a search first.
When the Show sources setting is turned on, Kai links the articles it used beneath its reply. Customers can open a source and read it in full, which also gives you a way to check where an answer came from when it looks wrong. Show sources is off by default — turn it on in your AI agent's settings if you want citations to appear in replies.
By default Kai runs on an automatic setting that picks a model and falls back to another when one is unavailable. On plans that include model selection you can pin a specific model instead. The model affects phrasing, reasoning on complicated questions, and cost per answer — not what Kai is allowed to say, which is governed by the grounding rule above.
Kai does not reveal its instructions, its configuration or the names of the tools it can call. If someone tries to talk it into acting as a developer or administrator, or asks it to print its own instructions, it declines and offers to help with a normal question instead.
