12 Conversational summarization for smooth handoff
This chapter covers
- Defining elements of an effective conversation summary
- Instrumenting your conversational AI to enhance summarization
- Summarizing a chat transcript into prose with LLMs
- Extracting structured details from a chat transcript with LLMs
Conversational AI builders would love it if their systems contained all user conversations. But for most use cases, some percentage of users will end their interaction with a human and not your bot. Conversational AI is designed to handle the easily automated conversations and direct the higher-value or more challenging ones to human agents. Users who want to self-service may be frustrated by “failing” with the conversational AI, so it’s important to give that human agent the best start possible at handling the call after a transfer.
The two simplest handoff methods are also the least satisfactory. We can transfer the conversation “blind” to the human agent and have them ask again for all the information they need. Or we can pass the agent the full conversational transcript and ask them to search it for the information they need (while the user is waiting!). It’s better to give the human agent a targeted summary of the conversation to date so they can quickly pick up where the conversational AI left off.