Answer customer support questions with your bot

A support bot is for any business whose team answers the same questions every day: how do I reset my password, where is my order, can I get a refund, do you ship to Ireland. The bot answers those from your own help content, any time of day, on your website and the other channels you use. When it can't help, or the customer asks for a person, it passes the conversation to your team with everything the customer has already said.

Done well, the bot handles the routine questions and your team gets fewer, better-prepared conversations. It's the use case most Chat Thing customers start with, and the others on these pages build on it.

What you'll set up

1. Knowledge sources. Add your help centre or docs as a website data source, using a content selector so menus and footers are left out. Add a Notion wiki or uploaded files if that's where answers live, and a manual data source for anything unpublished, such as returns rules. Turn on automatic syncing (Standard and above) so answers stay current.
2. Instructions. Write the Prompt in General settings using the example below. See Write your bot's instructions. With many help articles, turn on Enhanced knowledge retrieval so follow-up questions find the right content (Retrieval settings).
3. Power-ups. Add Hand off a chat to a human so the bot can email your team with the customer's details and the conversation. To look up orders or accounts, add Call your API.
4. Channel. Install the chat widget on your site and help pages. Add email or WhatsApp if customers already contact you there.
5. Handoff. On Standard and above, turn on Allow human takeover of conversations so your team can reply live in the chat.
6. Analytics. Read Conversations weekly. On Pro and Enterprise, use Analytics to see common questions, topics and sentiment.

The quickest start is to let Otto build the bot from your website: it reads your site, writes the instructions and adds your help pages. See Build your first bot.

Example instructions

Copy this into Prompt and replace the details with your own:

You are the support assistant for Acme, an online homeware shop.

You help customers with orders, delivery, returns, accounts and our products.

How to answer:
- Answer only from the information provided. Never guess prices, dates or policies.
- Keep answers to two or three sentences, and use numbered steps for instructions.
- Link to the relevant help article when there is one.
- Ask one clarifying question if the request is unclear.

When you can't help:
- If the answer isn't in the information provided, say so plainly and offer to pass the conversation to the support team.
- If the customer is upset, asks for a person, or the issue involves a payment, a damaged item or a complaint, offer the handoff straight away.
- Before handing off, ask for their email address and a one-line summary of the problem.

For more starting points, see the customer support prompts in the prompt library.

Example conversations

A routine question answered from the help centre:

Customer: How long do I have to send something back?

Bot: You can return most items within 30 days of delivery, as long as they're unused. Start a return from Your orders and choose Return an item. Sale items can be returned within 14 days. Full details: Returns policy.

A problem the bot shouldn't solve on its own:

Customer: My order arrived smashed and I want my money back.

Bot: I'm sorry, that's not the delivery you should have had. I'll pass this to our support team so they can sort out a refund. What's the best email address for them to reach you on?

Customer: sam@example.com

Bot: Thanks, Sam. I've passed our conversation to the support team, and they'll reply to you by email.

Pitfalls

• No way to reach a person. A support bot without a handoff leaves stuck customers going round in circles. Add the handoff before you launch, and say in the Prompt when to offer it.
• Letting the bot guess. Without "answer only from the information provided", the bot may fill gaps with plausible but wrong policies. Keep that rule in.
• Cluttered knowledge. Scraping whole pages, including navigation and cookie banners, gives the bot noise to match against. Use a content selector and remove outdated pages. See Structure your content for good answers.
• Testing only the easy questions. Before launch, ask the questions your team really gets, including ones your content doesn't cover, and check the bot offers a person when it should. See Test your bot.
• Launching and forgetting. The first few weeks show you what's missing. Review conversations and add answers for anything the bot couldn't handle. See Improve your bot's answers.

Before you go live, run through the launch checklist.