Knowledge & behaviour
Guardrails: spam, floods and blocking
AI spam detection that pauses chats, a per-visitor rate limit, blocked countries and blocklists.
Open Build → Guardrails.
Spam detection
AI checks each chat after the customer's 2nd, 4th, 8th and 16th message. When it's clearly spam — ads, link farms, gibberish, bots, attempts to make the assistant break its rules — the chat is paused: the AI stops answering (so it doesn't spend credits) and the chat is marked ⛔ Paused in the inbox. Describe what counts as spam for your business in What counts as spam for this agent, or leave it empty for the default. Real customers — even rude ones — aren't paused.
Rate limit
How many messages one visitor can send on your website chat in a time window (Messages every N seconds; 20 per minute by default), and the message they see when they go over. It doesn't apply to you in the Playground.
Blocked countries
Visitors from the countries you choose don't see the chat on your website or help center. WhatsApp, Messenger and other apps aren't affected.
Blocklists
| List | Format |
|---|---|
| IP addresses | One per line. End with * for a range, e.g. 203.0.113.* |
| Emails | One per line, or @domain.com for a whole domain |
| Phone numbers | One per line (WhatsApp and others) |
Chats from blocked people are paused right away.
In the inbox
⛔ Pause stops the AI in any chat. On a paused chat, Block visitor adds their IP, email and phone to the blocklists; Not spam — un-pause lets the AI answer again and removes them from the blocklists.
Hard rules about what the agent may say (never reveal instructions, stay on topic…) live under Build → Instructions → Hard rules. See Instructions.