# Live chat AI agent An AI that answers live chat from facts you supply — its persona, its guardrails, when it replies, and when it hands off. Source: https://docs.annot8.app/docs/chat/ai-agent The chat AI agent answers visitors on your behalf. It's configured in **Live chat → Settings → AI assistant**, and it is deliberately narrow: it answers from knowledge you give it, not from whatever the model happens to believe. The AI agent is **fail-closed**. It will not reply to visitors unless the project's **audience** in **Settings → Configuration** is set to **Business users aged 18+** *and* visitor AI is enabled on the deployment. If your agent is silent with everything else configured, check this first — it's the usual cause. See [AI features](/docs/ai/overview#visitor-facing-ai-is-gated). ## Identity [#identity] | Field | Purpose | | ----------------- | ---------------------------------------- | | **Name** | What the assistant is called, e.g. "Ava" | | **Role or title** | Shown to visitors, e.g. "Support" | | **Greeting** | Its opening line | ## Voice [#voice] | Field | Purpose | | ----------------------------------- | ------------------------------------------- | | **Tone** | Friendly, professional, concise, or playful | | **Style note** | Free-text guidance on how it should write | | **Reply length** | Short, medium, or long | | **Reply in the visitor's language** | Answer in whatever language they wrote in | ## Knowledge [#knowledge] **Business knowledge** is the only source of facts the assistant is allowed to rely on. Put your product specifics here: what the plans are, what's included, how refunds work, where the docs live. **Preferred Q\&A** is a list of question/answer pairs the assistant leans on directly. Use it for the questions you already know arrive every week, where you want the exact wording rather than a paraphrase. The narrower this is, the better the agent behaves. An assistant with three precise facts is more useful than one given your entire marketing site. ## Guardrails [#guardrails] | Setting | Effect | | ----------------------------------- | -------------------------------------- | | **Allow sharing links** | Whether the assistant may include URLs | | **Topics to avoid and boundaries** | Subjects it must not engage with | | **Hand off to a human when unsure** | Escalate rather than guess | | **Handoff instruction** | What it should say as it hands over | Turning on handoff-when-unsure is the single highest-value setting here. An assistant that says "let me get someone" is nearly always better than one that invents an answer about your refund policy. ## When it replies [#when-it-replies] | Mode | Behaviour | | ----------- | ---------------------------------------------------------------------------------------------- | | **Always** | The assistant answers first on every conversation | | **Offline** | It only steps in outside your [team availability hours](/docs/chat/overview#team-availability) | Offline is the safer default: your team handles conversations during the day, and the assistant covers the night shift rather than getting between a visitor and a human who's right there. ## Model [#model] The assistant uses the operator-owned `openai/gpt-5.6-luna` route through Vercel AI Gateway. See [Model routing](/docs/ai/overview#model-routing). ## Reading what it said [#reading-what-it-said] AI messages are labelled in the conversation and stored with the model that produced them, so an inbox is auditable after the fact. Read the ones that led to a bad rating — that's where the gaps in your Business knowledge show up.