Fallback is a scenario when the chatbot does not understand or cannot answer — the customer asks an out-of-range question, mistypes it, or the question requires a human. Good fallback: short apology, suggest another question or switch people — don't repeat the machine, don't "stuck" the customer in the bot. Zenify recorded 98.4% AI accuracy — meaning there is still nearly 2% that need fallback to be properly processed (approved).
TL;DR
- Fallback is the way out when chatbots don't understand.
- Short apology + suggest other options + switch people.
- After 2 times of not understanding → switch agent immediately.
- Do not repeat the same mechanical questions.
- Good fallback keeps customers, bad fallback loses customers.
When is fallback needed?
Questions outside the scope of knowledge base (see Knowledge base chatbot).
Customers type incorrectly, lack information, strange abbreviations.
Guest requests to meet people, complaints, sensitive situations.
"Uncertain" Chatbot — it's better to switch people than to guess wrong.
Good fallback principles
Short apology: "I don't understand what you mean."
Suggested solution: re-show menu or common questions.
There is a way out: always let guests choose "meet the staff".
Repetition limit: Maximum 2 fallbacks → switch people.
Do not "stuck" customers: The more stuck the customer, the more upset they become → the more they want to go.
Fallback sample
Mild fallback (1st time)
Bot: "I don't understand what you mean. Please help me choose again in the menu below, or retype the question."
Heavy Fallback (2nd time — shifting body)
Bot: "Sorry, I can't support this question. I'll connect a live support staff in a moment."
Guests want to meet you
Bot: "Yes, I'll connect to a live support staff. Please wait a moment."
How many times to fallback before switching people?
Maximum 2 times — then switch agents (see When you need someone).
1 time allows customers to correct mistakes or choose again.
Twice is enough to know the chatbot doesn't understand — keeping it any longer will only annoy the customer.
Switch people with context: agent sees conversation history so he doesn't ask again from the beginning.
Lỗi fallback thường gặp
Error | Consequences
Repeat "sorry, I don't understand" infinitely | Customers are upset and leave
There is no escape route | Guest "stuck" in bot
Fallback to death (no agent on duty) | Customers are tired of waiting
Sorry for the length | Readers are tired, still don't know what to do
Do not log fallback sentences | Don't know which sentence needs to be added
Fallback is precious data
Each fallback question is a signal: a question the customer needs that you don't have.
Periodically collect → add to knowledge base → increase containment.
Fallback is not "failure" — it is where the chatbot learns (see Containment rate).