1. Start with the questions you already get
Open your team's WhatsApp and email from last month and list the questions. Not the ones you think customers ask; the ones they do. For most businesses ten questions cover well over half the messages:
- Where is my order? When will it arrive?
- What are your timings? Where are you?
- How much does it cost? Is there a discount?
- How do I return or cancel?
- Can I speak to someone?
Write the answer to each one the way your best support person would say it. Short, specific, with the number or link the customer needs. This document is most of the work, and it is what the bot will be good or bad at.
2. Decide what the bot must never answer
This is the step most teams skip. Write down the topics that always go to a person:
- Anything about money already paid: refunds, double charges, failed payments.
- Complaints and anyone who sounds upset.
- Anything that needs judgement or a promise: custom pricing, bulk orders, exceptions to policy.
- In health, finance or legal work: any advice at all.
A bot that confidently gets a refund question wrong does more damage than having no bot.
3. Choose buttons or AI for each job
There are two kinds of automation, and good bots use both.
| Flow with buttons | AI agent | |
|---|---|---|
| Best for | Known steps: track an order, book a slot, pick a department | Open questions answered from your documents |
| Answers | Exactly what you wrote | Written fresh each time from your knowledge |
| Risk | Feels rigid if overused | Can be wrong if your knowledge is thin |
In NxAi the flow builder handles the first kind and the AI agent the second. The agent answers only from the knowledge you give it, and the topics you mark for hand-off always go to a person.
4. Design the hand-off before anything else
Customers forgive a bot that says "let me get someone". They do not forgive a loop. Three rules:
- Always offer a way out. "Talk to a person" should work at any point, typed or tapped.
- Send it to the right person. A billing question should reach billing, not whoever is free. NxAi does this with chat topics: the AI picks the topic, never the person, and the topic decides the team.
- Pass the context. The teammate should see the whole chat and not have to ask "how can I help?" again.
Outside working hours, the bot should say when someone will reply: "Our team is back at 10 am and will pick this up first."
My payment went through but the order says failed
Sorry about that. Payment issues need a person to check, so I've passed this to our billing team. They usually reply within 30 minutes.
Hi, Neha from billing. I can see the payment. I'm creating your order now.
5. Write like a person on WhatsApp
- Two or three short lines, not a paragraph.
- Answer first, then offer the next step.
- Buttons for choices, so nobody has to type "option 2".
- Reply in the customer's language. Many Indian customers switch between Hindi and English in one message.
6. Read real chats every week
For the first month, read a sample of chats every week and ask three questions. Where did customers ask for a person? Where did the bot give a wrong or vague answer? Which question came up that is not in your knowledge yet? Each answer is one edit to the knowledge or the hand-off rules. That weekly hour is what turns a decent bot into a good one.
Rules to know before you go live
- The bot can reply freely for 24 hours after a customer's last message. After that, the first message must be an approved template.
- Meta expects a clear route to a human. A bot with no exit is a policy problem as well as a bad experience.
- Do not ask for card numbers, PINs or OTPs in chat, whatever the bot is for.


