When a customer asks at 10 pm “what are your opening hours?”, or ten people a day repeat the same pricing question, answering by hand eats your time. A Telegram or WhatsApp chatbot can take that part, as long as you start the right way.
Which questions should AI answer?
A simple rule: give repeated questions with clear answers to the chatbot. Questions that need a decision, an exception or empathy stay with a person.
| Type of question | Who answers? | Example |
|---|---|---|
| Fixed information | Chatbot | Opening hours, address, how to order |
| Repeated question about a service | Chatbot (from your text) | What does this service include? How long does it take? |
| Simple follow-up | Chatbot or automation | Where is my order? |
| Complaints and refunds | Person | The goods arrived damaged |
| Exceptions and special prices | Person | Can you give a discount for a bulk order? |
- Customer message
It arrives on Telegram or WhatsApp.
- Search your knowledge
The chatbot finds the answer only in your own texts.
- Answer or “I don’t know”
If sure it answers; otherwise it says plainly that it doesn’t know.
- Handoff to a person
Complaints, money and exceptions reach you with a chat summary.
How do you make sure it doesn’t say wrong things?
- Answers should come only from your own texts. See how it works in how a RAG chatbot works.
- When it doesn’t know, it should say “I don’t know” and pass the customer to you instead of guessing.
- Limit its topic so it ignores unrelated questions and nobody can pull it off course. One known risk is tricky instructions hidden in a user message (see the OWASP list in the sources).
- It must never ask for private customer data (cards, passwords, ID numbers).
- If you are unsure which kind of chatbot fits, read rule-based chatbot or language model.
When should a person step in?
- The customer clearly asks for a person.
- The chatbot fails to give a confident answer twice in a row.
- The message is a complaint, about money, or otherwise sensitive.
- A summary of the chat goes with the customer so they don’t repeat everything.
Solo, the assistant on this site, is an example of the same idea: it answers only about our services and, for a direct conversation, sends the customer to Saman on Telegram and by phone.
How do you know it works?
- First-reply time, before and after the chatbot.
- The share of questions answered without a person.
- The list of unanswered questions; read it weekly and add the answers to the chatbot’s knowledge.
- Customer satisfaction; one short question at the end of a chat is enough.
Measuring and handing off to a person matches NIST’s AI risk-management guidance: a system that makes decisions needs human oversight. For costs, see the cost of a chatbot and an AI model.
For clinics and medical offices the limits are stricter; read chatbots for clinics.
To build a sales-reporting routine, also see the automatic weekly sales report.
If you run a greengrocer’s or fruit shop, AI for a greengrocer: today’s price, weight and orders takes the same view to daily prices and orders by weight.
Before automating, measure your current reply speed: see the 15-minute clinic reply-speed test.
Frequently asked questions
Can AI replace all customer support?
No. It handles repeated, clear questions well, but complaints, money and exceptions should stay with a person.
What happens if the chatbot doesn’t know the answer?
It should say so plainly and pass the customer to you or your support team instead of guessing.
What do I need to start?
A list of repeated questions with correct answers and a channel such as Telegram or WhatsApp. The rest can be built in stages.
How do I prevent abuse of the chatbot?
Limit its topic, never let it ask for private data, cap the number of messages and have it reject suspicious messages.