In 2022 a customer asked an airline's website chatbot how bereavement fares worked. The bot said they could buy a full-price ticket and claim a partial refund afterwards. The airline's actual policy said something different. The customer bought the tickets, applied for the refund, and was refused.
In February 2024 a tribunal in British Columbia, Canada, ruled in the customer's favour. It rejected the airline's argument that the chatbot was a separate party responsible for its own words, and found that the airline had not taken reasonable care to make the chatbot accurate. The case is Moffatt v. Air Canada, 2024 BCCRT 149 (full decision from the tribunal, PDF; summary by law firm Manatt).
That case is Canadian, and it is one tribunal's decision. It does not set the law in the US or Australia. Still, it shows what a customer's screenshot of your chatbot's answer can look like in a dispute. If you run a small business and are thinking about adding an AI chat, the useful question is how to keep it from inventing things.
I'm Ilya, a solo developer in Bali. I build AI chat for websites (AI assistant for your website). Here are five controls I'd want on any chat that talks to customers about money, rules or deadlines.
This is general information, not legal advice. Ask a lawyer how your local consumer law treats statements made by your chatbot.
Why chatbots make things up
A language model writes the most likely next words. When it has the facts, the likely words are right. When it lacks them, it still produces a fluent answer, and the answer sounds equally sure. Prices, opening hours, refund windows and delivery times are exactly the details a model can't guess, and exactly the ones customers act on.
So the fix is not a smarter model. The fix is to control what the model is allowed to say.
Control 1: answers come only from your own content
The chat should answer from a set of material you load: your price list, service pages, policies and FAQs. It should treat that material as the only source and ignore what it "remembers" from general training.
This technique is often called retrieval-augmented generation, or RAG. The chat looks up the relevant passages in your content first, then writes an answer based on them. If you can open your content and point to the sentence the answer came from, you can check the chat.
My AI assistant for websites works this way: it answers from the content you load (AI assistant for your website).
Control 2: permission to say "I don't know"
The most important instruction is the one that tells the chat what to do when the answer is not in your content. The right behavior is to say so and pass the customer to a human: your phone, email or a messenger. That is how my assistant behaves: if the answer isn't in the loaded content, it says so and can pass the conversation to you.
A chat that always has an answer is dangerous. A chat that says "I don't have that information, here is how to reach us" is boring, and it protects you.
Control 3: keep prices and rules in one place, with dates
Most wrong answers come from outdated source material. If the price list on your site says one thing and an old PDF says another, the chat may use either.
- Keep one current version of each policy.
- Put a "last updated" date on pricing and policy pages.
- Remove old PDFs from the material you load, or replace them.
- When a price changes, update the loaded content the same day.
The chat is only as accurate as the page it reads.
Control 4: test it with the questions customers really ask
Before launch, collect the real questions from your inbox, your phone log and your messengers. Ask the chat each one. Add the trick questions too: "Can I get a refund after 60 days?", "Is there a discount for students?", "Can you guarantee delivery by Friday?"
Read every answer next to the source text. Fix the content or the instructions, and ask again. I run this step before launch, with the real questions customers send.
Control 5: set limits on what the chat can promise
Write down what the chat must never do on its own: confirm a refund, promise a delivery date, quote a price outside the price list, give legal or medical advice. For those topics it should send the customer to a human.
Two practical limits help as well:
- A spending cap, so a surge of chats can't produce a surprise bill. I set one upfront and the chat won't go past it.
- Access to the conversation history. Whoever builds your chat, ask whether you can read past conversations, so you notice what customers ask and where the chat hesitates.
A real example: Rentacar Chile
Rentacar Chile is a car rental site in Chile. I added an AI chat that answers from the company's own knowledge base, so a visitor asking about renting a car gets an answer drawn from the company's information.
Two details shaped how the chat works.
- Suggested questions. The chat shows ready-made question buttons for common topics. A customer taps one, and the answer comes from the loaded content.
- Contact handoff inside the chat. If a visitor wants to book or needs a person, the chat collects a contact and sends it to the owner in Telegram. The chat doesn't try to close a deal on its own.
The chat stays within the company's information, and anything it can't settle goes to a human. For the wider picture of adding AI to a product, see my guide to adding AI features with Claude or GPT.
What I can't promise
No one can promise a chat with zero mistakes, and you should be wary of anyone who does. Mistakes can't be ruled out completely. What the controls above do is narrow the room for error and make the remaining errors visible: when something goes wrong, you can look at the source text and see why.
A chat is not worth it for everyone. If you have about twenty questions that never change, a clear FAQ page does the job. If you get five messages a day, your team can answer them. A chat starts to pay for itself when the same questions arrive in volume.
What it costs
My ready-made chat widget starts at $700 and goes live in up to 3 days if your material is ready. A custom build that connects to your booking calendar, CRM or product database takes 2 to 4 weeks. You pay for AI usage every month, in proportion to the number of conversations, and I estimate it upfront from your volume.
Payment is by bank transfer (to an Indonesian account), PayPal or crypto (USDT or USDC).
Summary
- Load your own content and make it the only source.
- Let the chat say "I don't know" and hand over to a person.
- Keep one dated version of every price and policy.
- Test with real customer questions before launch.
- Set limits on what the chat may promise and a spending cap.
Published: . By Ilya Litau, more on the About page.
Want a chat that sticks to your own content?
Tell me what your customers ask most. I'll tell you whether a chat makes sense and what it would cost.