Conversation summary at close

When an agent closes a chat conversation, this feature generates a small summary of what happened and stores it on the conversation. Useful for skimming the inbox after the fact and for reports.

How it works

Closing a chat triggers a server-side Claude Haiku call with the full transcript. The model returns:

• Summary — 2–3 sentences on what the conversation was about and how it ended.
• Resolution — a one-line "what we did" or "what's pending".
• Tags — up to 3 suggested tags (free-form text).

All three are saved on the ChatConversation row and surface in the closed-conversation sidebar as a violet card.

The agent doesn't trigger it — it just happens on close. The chat itself is closed immediately; the summary writes a couple of seconds later. Refresh once and it's there.

Where you see the summary

• The chat detail page sidebar, on closed conversations.
• The inbox list — closed chats can be filtered to those with summaries.
• (Future) CSV exports of chats.

What it's good for

• Postmortem — quickly see what 50 chats from yesterday were about.
• Pattern spotting — if the same resolution string appears 30 times, you have a documentation gap.
• Manager review — instead of reading transcripts.

Privacy

The model sees the full chat content, including the visitor's name and email if the chat collected them. If your customers wouldn't want their chats summarized by a third-party LLM, disable this feature.

Cost

Typical: 2000 input tokens + 100 output tokens. Around $0.003 per closed chat. Even at high volume this stays small.

Off-switch

Settings → AI → uncheck Conversation summary at close. New closes won't generate summaries; existing summaries stay in place.