Chat reply suggestions
Get an AI-drafted reply to the last visitor message, grounded in your knowledge base.
In the chat reply form, the ✨ AI suggest reply button asks Claude Haiku to draft a response based on the conversation so far + the most relevant published KB articles.
How it works
When you click the button:
- We grab the last ~10 messages in the conversation as context.
- We pull the latest visitor message, run it as a MySQL FTS query against your published, public KB articles, and take the top 3 by relevance.
- We send Claude: the last 10 messages, the 3 KB articles (title + body), and a prompt asking for one suggested reply.
- Claude returns a draft and the IDs of the articles it cited.
The draft appears in a violet panel above the reply form. You can:
- Use it as-is — click Use. The draft fills the reply textarea; you edit if needed and Send.
- Translate it — if the visitor's last message wasn't English, a Use translated version option also appears (see Language translation).
- Ignore it — just keep typing your own reply. The draft doesn't auto-fill.
Citations
The panel shows which KB articles the suggestion drew from. Cited articles are public and visible in your /help — useful both for verifying the AI's claims and for double-checking that you actually have an article on this topic (if you do, maybe you should send the customer a link to it instead of typing a new answer).
Without a KB
If your KB is empty or has no matches, we still send the prompt to Claude with just the conversation context. The response will be more generic — based on what Haiku knows about your domain from training, not your specific docs. Better than nothing for new workspaces, but the value compounds as your KB grows.
Cost
Typical call: ~3000 input tokens (KB context plus message history) + ~150 output tokens. About $0.005 per call. Heavy use by a busy agent (50 suggestions a day) costs $0.25/day per agent.
When it's most useful
- Common questions you answer the same way every time but want to phrase per-customer.
- New agents who don't yet know the standard reply patterns.
- Long-running chats where you'd benefit from a quick rephrase before responding.
When it's worst
- Sensitive customer relationships where every word matters — better to write yourself than risk a tone-deaf draft.
- Topics outside your KB where the AI's draft is generic and unhelpful.
- Visitors writing in a niche language with no KB content — the suggestion will read awkwardly.
Off-switch
Settings → AI → uncheck Chat reply suggestions. The button stays visible but always returns "no suggestion".