Intents
Top 5 intents often cover 40%+ of chat volume. Map them, train them, build flows per intent.
Flows
Conversational flow: greet, disambiguate, provide answer or collect info, offer handoff. Keep flows shallow — deep nesting frustrates users.
LLM Answers
Freddy Self Service answers unstructured questions by searching KB. Pair with intents — bot uses flows for structured, LLM for edge cases.
Handoff
Escalate to human on low confidence, specific keyword (‘agent’, ‘human’), or after 2 failed attempts. Carry context.
Measuring
Deflection rate, CSAT on bot interactions, handoff rate by intent. Iterate. A 40% deflection rate on supported intents is reasonable; 70% is excellent.