A customer asks for a human agent after several turns with a chatbot. The agent needs to know what happened, but reading the full transcript delays the handoff. A conversational summary compresses that history into a few useful points so the agent can start helping without making the customer wait. section 12.1.1
Start with the transcript and decide what the recipient needs to know. A handoff summary might state the user’s goal, the information already found, and what happened immediately before the request for an agent. It can be brief prose or structured fields containing key details. section 12.3 section 12.4
A claim-status handoff
A user asks about a claim. The bot finds a claim and says it was paid; the user then types “representative.” A useful summary reports the claim-status task, the payment information given, and the request for an agent. It should not assert that the claim was denied: the transcript does not establish that. section 12.1.1 section 12.3.1
To make a summary useful and reliable:
- Keep it short enough to scan, but retain the details needed to continue the conversation. section 12.1.1
- Tell an LLM which task to emphasize; examples can help establish the desired output. section 12.3.1
- For structured output, specify the required fields and value types. Test multiple transcripts and check extracted values against the source. section 12.3.2
Summarization can also serve the customer directly: after an intent is identified, a chatbot can greet the user and restate their problem before continuing with prepared content. In either setting, a plausible-sounding summary is not enough; it must remain faithful to what the user and bot actually said. section 5.4.2 section 12.3.2
