Automation & AI
AI agents
An AI trained on your own knowledge base, replying to routine questions and handing off the rest.
Setting it up
Under AI Agents, choose a provider and model, and write a system prompt describing your business — what you sell, your policies, your tone. A master switch turns the agent on or off for the whole account, and a separate toggle controls whether it auto-replies to customers versus only being available for drafting.
Knowledge base
Add documents — FAQs, policies, product details — and the agent draws on them when answering. Each document is broken into smaller chunks and, when configured with an embeddings key, searched semantically (by meaning, not just keyword) so a differently-worded question can still find the right answer; a keyword-based search fills in alongside it either way.
The playground
Test the agent in a private chat before it ever talks to a real customer — same prompt, same knowledge base, same model, no messages sent anywhere. Useful for checking a prompt change or a new knowledge document actually changes how it answers before you trust it with live conversations.
When it actually replies
Auto-reply only fires when nothing else already has a claim on the conversation:
- A human teammate isn’t already assigned to the thread.
- No automation is already set to answer this message.
- The conversation hasn’t already hit its per-conversation reply cap (configurable, so the agent can’t loop indefinitely on one thread) — unless it’s been explicitly assigned to that conversation, which lifts the cap.
It also waits a few seconds after a customer stops typing before replying, so a burst of several quick messages gets answered once, not once per message.
Handing off to a human
When the agent can’t confidently help — a question outside its knowledge base, something that needs a judgment call — it hands the conversation off rather than guessing. The conversation then surfaces as unassigned in the inbox for your team to pick up.
Assigning the AI to one conversation
Beyond the account-wide auto-reply setting, you can assign your AI agent to a specific conversation directly from the inbox — the same way you’d assign a teammate or a Flow. Only one of the three can own a conversation at a time.
| Configurable | Notes |
|---|---|
| Provider & model | OpenAI, Anthropic, or OpenRouter — pick the one you have an API key for. |
| System prompt | Your business context — what shapes every reply. |
| Auto-reply cap | Max automatic replies per conversation before it stands down and waits for a human. |
| Knowledge base | Documents the agent can draw on when answering. |
Drafting instead of auto-sending
