What the chatbot knows

It reads your knowledge base, and only the public part of it. The chatbot only reads from public knowledge base articles. Mark articles as public in the knowledge base to include them.

That is the whole training step. There is no separate model to teach and no list of question-and-answer pairs to maintain — fix the article and the bot's answer changes with it.

It cannot see your CRM. No customers, no deals, no invoices. A visitor cannot get it to reveal anything about anybody, because it has no access in the first place.

Shaping how it answers. Under AI behavior:

• Custom instructions (optional) — Appended to the chatbot's system prompt. Use this for brand voice or domain-specific rules. For example: e.g. Always answer in formal English. Mention our 30-day money-back guarantee where relevant.
• Default language — The language the chatbot greets and replies in. Auto follows the visitor's device language. Choices are Auto (visitor's device), English, Norsk, Svenska and Dansk.
• Fallback message — Shown when the AI is unavailable or your tenant runs out of AI credits.

When it should give up and fetch a human. Escalation eagerness: How readily the bot hands off to a human instead of answering. The number is the knowledge-base match score below which it offers to escalate. Relevant matches typically score 0.35–0.55, so the recommended 0.30 means the bot answers whenever it finds anything and only offers a human when the search comes up empty. Raise it to escalate more eagerly when matches are weak.

In practice: if visitors complain the bot waffles instead of admitting it does not know, raise it. If it hands off things it could easily have answered, lower it — and check whether the article it needed is actually marked public.

"The chatbot gives wrong answers" is nearly always a knowledge-base problem, not a bot problem. Ask it the question yourself in Live test, then look at whether an article really covers it, clearly, in public. See Write a knowledge base article.

It costs AI credits. When they run out, visitors get the Fallback message rather than an error — so write one that gives them somewhere to go.

It stays focused on your business. Greetings and questions about your products, pricing and support are welcome. Unrelated homework, general knowledge and requests to write software get a short redirect. A small topic check uses AI credits, but an unrelated request does not run the full answer.

Each conversation has a lifetime limit of 30 AI attempts. The limit includes topic checks and interrupted replies, and does not reset each hour or when the visitor refreshes the page. Once it is reached, the visitor is directed to human support. Messages to a human agent can continue. See Step into a live chat.