Get accurate answers

The bot’s answers are only as good as what you teach it and how you set it up. This page explains how the bot picks an answer, so that when one is wrong, you know what to change.

How the bot decides what to say

For each customer message, the bot:

1. Reads the conversation. It sees the new message and up to the last 10 messages of the chat from the last 24 hours, both yours and the customer’s.
2. Works out what the customer wants. For example a general question, a product, an order or a booking.
3. Searches your knowledge. For a general question, it searches your entries in the category Other. For a product, it searches the in-stock products synced from your store. If the bot uses only selected documents, it searches only those.
4. Writes the answer from what it found. Your instructions, the bot’s Behaviour and your company details shape how it says it. It’s told to use only what it found, never its general knowledge.
5. Says when it doesn’t know. If nothing in your knowledge answers the question, it sends your Fallback Message. If the bot has no fallback message switched on, the customer gets I don’t have information on that topic.

When the bot shows products from your store, it writes the answer from the product details only. Questions about an order or a booking are handled separately and don’t use your knowledge. See Products, orders and bookings.

Stop made-up answers

A made-up answer, which AI people call a hallucination, is an answer that sounds right but isn’t true, because the AI filled a gap with a guess.

Wabot’s built-in rules already tell the bot to answer only from your knowledge, and to say it doesn’t know rather than guess. Most wrong answers come from gaps or mix-ups in the knowledge itself. To keep answers accurate:

• Cover what customers really ask. Look through your inbox for the most common questions, and make sure each one has an entry.
• Spell out the details. Delivery costs ₹60 within 8 km is better than a small delivery charge. A vague entry leads to a vague or wrong answer.
• Keep one version of each fact. When a price changes, edit the entry instead of adding a second one. Remove website pages that show old prices.
• Use the category Other for everything you add yourself. See Which category to choose.
• Write a Fallback Message. It gives the bot an honest answer when it doesn’t know, and tells the customer how to reach you.
• Don’t push the bot to always answer. Instructions such as “never say you don’t know” or “always give a price” make guesses more likely. Tell the bot what to do when it isn’t sure instead.
• Limit what each bot can use. If a bot should only talk about one service, select only those documents in its Data Configuration.
• Test with hard questions. Ask about things that aren’t in your knowledge, and check that the bot says so.

Tone, length and language

Tone

• The bot’s Behaviour, such as Friendly or Professional, sets its overall tone.
• Its Department, such as Customer Support or Sales, sets the role it plays.
• Your Instructions can fine-tune both, for example Use simple words and short sentences or Sound like a friendly shop assistant.
• If no AI bot is set up for the number, the name, role and behaviour come from the number’s agent identity.

Length and format

• Answers are kept under 1,000 characters, about 150 words.
• Answers are plain text, with short sentences on separate lines. There’s no bold text, no headings and no bullet lists, except in lists of products from your store.
• The bot adds a link only when the customer asks for a link, for more details or for your website. Links are shortened.
• The bot greets the customer only if the customer greeted first.
• While the bot is writing, the customer sees typing….

To send a menu, a price list or a booking link anyway, attach it to the knowledge entry. See Attachments.

Language

The bot replies in English, even when the customer writes in another language. There’s no language setting in the app, and you shouldn’t rely on your instructions to change it. If most of your customers write in another language, contact support.

Fixed messages such as I don’t have information on that topic. are always in English. Your Fallback Message is sent exactly as you write it, so you can write it in your customers’ languages.

Memory

The bot remembers up to the last 10 messages of a chat from the last 24 hours. If a customer comes back after two days, the bot starts fresh. In a long chat, earlier details drop out, and the customer may need to repeat them.

Test before customers see it

Test in two ways: with Test bot for quick checks, then from your own phone for a real chat.

Test bot

On Settings → Meta Platforms, click Test bot next to the number. A chat window opens. Type a question in Type a test message and send it.

• Test bot needs Bot Response on for the number. While it’s off, the button shows Bot disabled.
• It uses the number’s Default AI Bot. With no default bot, it uses the number’s agent identity.
• It works when the bot is in Test or Live mode. You don’t need to add test numbers for it.

Test bot is a little different from a real chat:

• It doesn’t remember your earlier test messages. Each question is answered on its own.
• Workflows don’t run.
• It may show I don’t have information on that topic. where a customer would get your fallback message.
• Its replies aren’t charged, but your wallet still needs credit, unless you use your own key.
• The test chat is kept only in your browser. Click the clear icon (Clear chat history) to start again.

Test from your own phone

A real chat shows what customers will see, including memory, workflows and your fallback message.

1. In AI bots, set the bot’s Mode to Test and add your number under Test Phone Numbers.
2. Check that the bot is the Default AI Bot and that Bot Response is on for the number.
3. From your phone, send WhatsApp messages to your business number, the way a customer would.
4. Look at the chat in Wabot’s inbox too. Replies from an AI bot are marked AI agent.

Good things to try:

• The 10 questions customers ask most.
• A question in two parts, such as Do you deliver to Kothrud, and how much is it?
• A follow-up that depends on the last answer, such as And on Sundays?
• Something you don’t offer, to see your fallback message.
• A complaint, to check the tone.

After you go Live

• For the first week, read a few real chats in the inbox each day.
• If the bot fails to reply, the bell Updates shows an alert under Automation, such as AI agent “Maya” failed to reply.
• Each AI charge is listed on the Wallet page (Settings → Pricing → Wallet), under Wallet history → Usage.

Before you go Live

• The agent identity has your Company name and Other details (Agent Knowledge → Edit identity).
• Your knowledge covers the questions customers ask most, in the category Other.
• If you sell through Shopify or Wix, your products were synced before you added your other knowledge.
• The bot has clear Instructions and a Fallback Message that says how to reach your team.
• Default AI Bot and Enable Bot are on.
• Bot Response is on for the number.
• You’ve tested with Test bot and from your own phone.
• No workflow with the trigger Customer sends a message catches every message. It would answer instead of the bot.
• Your wallet has credit, or your own key is set up on AI providers.
• Your team knows how to take over a chat: assign it, or turn off the AI switch. See When a teammate takes over.

Then open the bot in AI bots, set Mode to Live and click Save Changes.

Note

Something doesn’t look right? See When the bot doesn’t answer right.