Key takeaways
- Map intents and handoffs before choosing a model
- Silence and escalation beat fake empathy
- Measure containment and CSAT - not chat volume
Users don’t hate chatbots. They hate bots that stall, hallucinate policy, or trap them in loops while a human is three clicks away. The fix is product design, not a warmer system prompt.
Intent map first
List the top intents from real tickets or call logs. For each intent, decide: automate, assist a human, or refuse. If you can’t classify 70% of volume into clear intents, you’re not ready for a bot - you’re ready for better FAQs and search.
- 01Automate: status checks, password resets, hours, shipping tracking
- 02Assist: draft replies for agents, retrieve policy snippets
- 03Refuse + hand off: billing disputes, legal, medical, anything high-stakes
“A good bot knows when to shut up and get a human.”
Handoff is a feature
- 01
Detect stuck states
Two failed clarifications, repeated “that’s not what I meant,” or any safety intent → escalate.
- 02
Pass context
Transcript, user ID, and attempted intent must land with the agent - not a blank ticket.
- 03
Set expectations
Tell the user who is next and when - “Connecting you to support” beats fake typing forever.
Knowledge that doesn’t rot
Retrieval only works if the source of truth is owned. Pick a small set of canonical docs. Version them. Block the bot from answering outside that set. Hallucinated policy is worse than “I don’t know - here’s a human.”
70%+
Volume coverable by clear intents before launch
2
Failed clarifications before human handoff
1
Canonical knowledge owner
