Skip to Content
📚 Documentation is in active development — pages marked TODO will fill out over the next few weeks.
Chat AIOverview

Chat AI agent

/dashboard/chat-ai — configure a per-client AI agent that auto-replies to inbound text when the conversation is in AI mode.

What you configure

  • Persona name — what the agent calls itself (“Maya”, “Acme’s assistant”)
  • Persona tone — friendly / formal / casual / professional
  • Business description — markdown blob folded into the system prompt
  • FAQ knowledge — markdown FAQ block; the agent uses this for retrieval-ish Q&A
  • Guardrails — hard rules as bullet points (“Never quote prices”, “Don’t make promises about timelines”, “Always greet the customer by name if known”)
  • Auto-handover phrases — substring matches that flip the conversation to HUMAN mode (“agent se baat”, “talk to a human”, “real person”)
  • Confidence threshold — 0..1. Replies below this are suppressed; conversation flips to HUMAN.
  • Reply delay — min + max milliseconds. The agent waits a randomised interval before sending so it doesn’t feel bot-fast.
  • Max replies per conversation — safety cap to prevent infinite back-and-forth.

Chat modes per conversation

ModeBehaviour
HUMANAI silent. Agent replies manually.
AIEvery text inbound triggers an AI reply.
FLOWA flow is running on this conversation; AI defers.
HYBRIDAI suggests a reply (shown in the composer); agent reviews + sends.

Toggle via the conversation header in the inbox.

How the AI knows what to say

System prompt is assembled from:

  1. Persona name + tone (gives a voice)
  2. Business description (gives context)
  3. FAQ knowledge (gives canned answers)
  4. Guardrails (gives don’ts)
  5. Recent conversation history (last ~10 turns)

User prompt is the customer’s latest inbound text.

Auto-handover

Two paths to handover:

  1. Substring match on configured auto-handover phrases → instant HUMAN flip
  2. Low confidence — AI returns reply but the confidence score is below threshold → suppress + flip HUMAN

In both cases the conversation gets a system note: “Handed off to a human agent.”

Last updated on