Fazley RabbiCustomers generally welcome AI when it gives a quick, correct answer to a simple question, at any hour, in their own language. They get annoyed when it traps them in loops, pretends to be a person, gives confident wrong answers or stands between them and a human when something has gone wrong. The balance is easy to state and takes care to build: let AI handle routine questions and the first response, and make it easy and quick to reach a person for everything else.
What customers tend to welcome
Think about your own experience as a customer. Most people are happy with automation when it saves them waiting.
- Instant answers to common questions. Delivery areas, delivery time, opening hours, whether cash on delivery is available, how to pay by bKash or Nagad.
- Order status without calling. "Where is my parcel?" answered in seconds.
- Replies outside office hours. A message sent at midnight gets a useful answer, not silence until morning.
- Their own language. A reply in Bangla when they wrote in Bangla, and a sensible reply when they wrote in Banglish.
- Not repeating themselves. When a person takes over, they already know what the customer said.
What annoys them
- Loops. The customer asks something slightly unusual and the bot repeats the same menu or the same vague answer.
- No way out. There is no clear path to a human, or the "talk to a person" option leads nowhere.
- Pretending to be human. People feel deceived when they discover that "Rina from customer care" was software. Be open that the first reply is automated.
- Confident wrong answers. A wrong price or delivery promise, delivered politely, is worse than "let me check with the team".
- Cold handling of complaints. An upset customer who gets a cheerful automated reply feels ignored.
- Over-messaging. Automated follow-ups that keep coming after the customer has lost interest.
A practical split between AI and people
The table below is a starting point. Adjust it to your business and your customers.
| Situation | Who handles it |
|---|---|
| Common questions with approved answers | AI, from your answer sheet |
| Collecting order details | AI, then a person confirms |
| Order status | AI, from your order records |
| Price bargaining or special requests | Person |
| Complaints, refunds, damaged goods | Person, quickly |
| Health, legal or safety questions | Person only |
| Anything the AI is unsure about | Person, with the conversation history |
The rule underneath is simple: AI answers only from information you have approved, and a person makes any decision that costs money, affects safety or could upset the customer. The article on human review in AI workflows goes deeper into where to place that check.
Designing the handover
The moment of handing over to a person decides whether customers like the system or resent it.
- Make the exit visible. Tell customers early that they can type "person" or tap a button to reach the team.
- Hand over on signals. Anger, repeated questions, words like "refund" or "complaint", or a large order should go to a person automatically.
- Pass the context. The person should see the whole conversation, the customer's name and what they want.
- Be honest about timing. If nobody is available until morning, say so, and make sure someone actually follows up.
Tone matters, especially in Bangla
Instruct the AI to use the respectful "apni" form, short sentences and a warm, plain tone. Give it examples of good replies written by your best staff member. Test it with real messages, including rude and confusing ones, before customers see it. Bangla tone errors are easy to miss if the setup is only reviewed in English.
Keep checking after launch
Read a sample of AI conversations every week. Look for wrong answers, missed handovers and customers who gave up. Update the answer sheet whenever prices, stock or policies change. Ask your team which questions the AI handles badly, and fix those first.
A checklist before you switch it on
- Approved answer sheet written and up to date
- A clear, visible way to reach a person
- Handover rules for complaints, refunds and uncertain answers
- The first reply says it is automated
- Tested with real Bangla, Banglish and English messages
- One named person reviews conversations every week
If you sell through a Facebook page, an AI assistant for night-time messages is a sensible first step. If email is where the pressure is, consider AI-drafted replies that staff check. Not sure whether you need a bot at all? Read chatbot vs live chat.