Glossary / Chatbot Fallback
Chatbot Fallback
A chatbot fallback is the moment an AI agent's confidence drops below threshold and it transfers to a human — the design of this handoff determines whether AI support feels seamless or frustrating.
Full definition
A chatbot fallback (also called escalation or human handoff) is triggered when an AI agent encounters a query it cannot answer with sufficient confidence, when a customer explicitly asks for a human, or when a conversation has hit a defined decision point (negotiation, complaint, refund above a threshold) where human judgment is required. The quality of the fallback experience — not the AI's solo performance — is often what customers remember most. A well-designed fallback passes the full conversation transcript to the human agent, includes an AI-generated summary of what was already established (name, issue, sentiment, any account data retrieved), and does not make the customer repeat themselves. NimbleBiz's handoff puts the full conversation context, Captured Details, and AI confidence score in the agent's view before they send their first message. A poorly designed fallback — where the human agent asks the customer to 'start from the beginning' — destroys the trust the AI built in the conversation.
Key facts
- Triggers: low confidence score, explicit customer request, high-value or complaint query, policy boundary
- Full transcript + AI summary should be visible to the human agent before they take over
- Never make the customer repeat themselves — this is the #1 reason post-handoff CSAT drops
- Fallback rate (% of conversations escalated) is a design metric, not just a failure metric — some escalation is correct and necessary
- Pair with CSAT tracking per handoff reason to identify which query types need knowledge base updates
Related terms
An AI support agent is an autonomous AI that answers customer service questions by retrieving answers from a company's knowledge base, docs, and past tickets — typically deflecting 60–80% of tier-1 tickets before they reach a human.
ReadTicket deflection is the share of customer support tickets resolved by AI, self-service, or automation before they reach a human agent. In modern AI-powered support, deflection rates of 60–80% on tier-1 tickets are typical.
ReadCSAT is the percentage of customers who rate an interaction positively — the primary check on whether AI ticket deflection is actually resolving issues or just closing conversations without solving problems.
ReadSee chatbot fallback in production
NimbleBiz.ai ships this out of the box. Start free — no credit card required.