How to Add an AI Chatbot to Your Website That Doesn't Make Things Up
An AI chatbot for your website can answer the same questions your team answers every day, instantly and at any hour. But everyone has seen a chatbot confidently invent a refund policy or a delivery date. This guide covers how to set one up so it helps customers instead of creating problems for your team.
Why chatbots make things up
AI models are trained to produce fluent, plausible text. If a chatbot doesn't have the facts it needs, it may fill the gap with something that sounds right. That's harmless in a brainstorm and a real problem when a customer asks whether they can return an opened item.
The fix isn't a smarter model. It's giving the model the right facts and clear rules about what to do when it doesn't have them.
Rule 1: Answer only from your own content
A reliable website chatbot answers from your help articles: shipping, returns, pricing, setup guides. When a visitor asks something, it looks for the answer there and replies in a few sentences, ideally linking the article it used.
This has a useful side effect: the better your help center, the better your chatbot. Write a short article for every question your team answers more than once a week. You'll help customers who read articles, and the AI gets smarter at the same time.
Rule 2: Never invent prices, policies or order details
Tell the assistant plainly that it must not state anything about prices, deadlines, policies or orders unless your articles say it. When it doesn't know, it should say so and offer to bring in a person. A good tool builds these instructions in, so you don't have to remember to write them.
Rule 3: Hand over to a person at the right moments
The chatbot should pass the conversation to your team when:
- the visitor asks for a person,
- the question isn't covered by your articles,
- the visitor needs something done on their account or order, like a refund or a change, or
- the visitor is upset.
The hand-over should include a one-line note for the agent saying why, so they don't have to read the whole conversation first. If nobody is online, the visitor should be told when to expect an answer.
Rule 4: Let your team take over any time
Your agents should be able to see AI conversations as they happen and step in. The moment an agent joins, the AI should stop. This keeps people in control and lets you spot answers you'd like to improve.
Choosing the AI model
You don't have to pick one company. Our AI chatbot works with Claude, ChatGPT (OpenAI), Google Gemini, DeepSeek, Mistral, Grok and hundreds more models through OpenRouter. You bring your own API key, so you pay the AI provider directly and control the cost.
A few practical tips:
- Bigger models give better answers. They usually cost more per reply.
- Smaller "mini" or "flash" models are fast and cheap. They're often good enough for straightforward FAQs.
- On phone calls, speed matters most. Pick the quickest model for a voice agent.
Answer first, or only as backup?
There are two common ways to run an AI assistant:
- AI first: the assistant answers every new chat, and your team only gets the ones it hands over. Best when most questions are covered by your articles.
- Backup: your team answers as usual, and the AI only steps in when nobody is free: nights, weekends and busy hours. A gentle way to start.
A simple rollout plan
- Write or update help articles for your ten most common questions.
- Turn the assistant on in backup mode for a week.
- Read its conversations. Where it handed over, ask whether an article would have answered it, then write that article.
- When you're happy with the answers, switch to AI first.
The result
Set up this way, an AI chatbot answers the easy questions instantly, admits what it doesn't know, and gives your team more time for the conversations that need a person. You can try it free on your own website.