A customer support chatbot works best on repeat, factual questions with a clear answer, such as order status, opening hours, how a process works or where to find a document. Anything involving money disputes, complaints, emotion or unusual cases should go to a person, quickly and with context.
What to automate first
Start by reading a month of real support messages and sorting them. You will usually find that a small number of questions make up most of the volume.
| Good for a chatbot | Better for a person |
|---|---|
| "Where is my order or booking?" | Refund or compensation disputes |
| "What are your hours or locations?" | Upset or distressed customers |
| "How do I change my details?" | Complex, multi-step problems |
| "What documents do I need?" | Anything legally sensitive |
| Questions answered in your FAQ | Exceptions to policy |
Design the hand-off before the bot
The failure that customers remember is being trapped. Design the exit first:
- Offer a clear "talk to a person" option at any point.
- Trigger a hand-off automatically on frustration signals (repeated rephrasing, strong language) or low confidence.
- Pass the conversation summary to the agent so the customer does not repeat themselves.
- Say honestly what hours a human is available.
Keep it accurate
The bot should answer from your own content, not from general knowledge. If it does not know, it should say so and offer a hand-off, rather than guess. Read real conversations every week and add the missing answers.
See also our guide on what an AI chatbot should never say.
Measure what matters
- Resolution without a human, but only counting cases where the customer did not come back.
- Hand-off rate and reason, to find gaps in your content.
- Customer feedback on the conversation, even a simple thumbs up or down.