The AI assistant is who talks with your customers first over WhatsApp, Telegram and your website’s chat (Manual 16 · Conversations). It doesn’t make up what it knows out of nowhere: it replies with the “business memory” you’ve given it — the documents you upload to the Knowledge base (Manual 18) — and, if your agency has them connected, with your agent’s external tools (MCP servers, Manual 26). If an agent has no document or tool assigned, it doesn’t see any of that: it replies only with its general instructions (Manual 23 · AI agents, section 6).
2.1. It can be wrong: it isn’t an absolute source of truth
Section titled “2.1. It can be wrong: it isn’t an absolute source of truth”The assistant generates its replies by probability, not certainty: it can combine two facts incorrectly, fail to find the exact answer, or state something incorrect with total confidence. Like anyone on your team, it can make mistakes — the difference is that its mistakes can sound very convincing. Never treat one of its replies as a verified fact without checking, especially on something sensitive: prices, legal deadlines, contractual terms, or any firm commitment to a customer. If something the assistant said seems doubtful to you, review it yourself in the conversation thread (Manual 16, section 3) before taking it as correct.
2.2. Limits by design
Section titled “2.2. Limits by design”- It doesn’t make up prices or policies that aren’t in its documents. When it can’t find the answer, it replies with the fallback message your agency configured for that agent (Manual 23, section 3) — never with a made-up fact.
- It hands off to a person when appropriate. If the customer explicitly asks to talk to someone, or an escalation rule your administrator configured fires (keywords, unresolved turns…), it stops replying and the conversation moves to wait for your team (Manual 23, section 4; section 5 of this manual).