The fastest way to lose trust in AI is a chatbot that confidently tells a customer something untrue: a price you don't charge, a refund policy you don't have, an appointment slot that doesn't exist.
The good news is that this is a solvable problem. Accurate chatbots aren't lucky. They're designed with guardrails.
Why chatbots get things wrong
Large language models are built to produce fluent, plausible text. If they don't have the right information, they'll often produce a plausible guess rather than saying "I don't know". Most chatbot mistakes come from one of three places:
- No source of truth. The bot is answering from general knowledge instead of your business's information.
- No boundaries. Nobody told it what it's not allowed to answer.
- No exit. It has no way to hand over to a person, so it improvises.
The guardrails that matter
Answer only from approved information
Give the chatbot your own material, such as services, policies, FAQs and opening hours, and instruct it to answer only from that. If the answer isn't in the material, it should say so and offer to connect the customer with someone who knows.
Write down what it must never do
Be explicit. A typical list:
- Never quote a price unless it's in the approved price list.
- Never promise delivery dates, availability or outcomes.
- Never give legal, medical or financial advice.
- Never make commitments on behalf of the business.
Hand over gracefully
The handover is a feature, not a failure. When a question is outside its knowledge, emotional, or involves money, the bot should collect the details and pass the conversation to a person, with the context, so the customer doesn't have to repeat themselves.
Keep sensitive actions behind a person
Booking a slot is fine. Issuing a refund, changing an order or agreeing to a custom quote should go through a human checkpoint.
| Situation | What the chatbot should do |
|---|---|
| Common question with an approved answer | Answer it directly |
| Question not covered by your material | Say so, offer a handover |
| Price, refund or contract question | Collect details, pass to a person |
| Upset or urgent customer | Hand over straight away |
Test it like a sceptical customer
Before launch, try to break it. Ask questions slightly outside its knowledge, ask for discounts, ask the same thing three different ways, and ask it to ignore its instructions. Every wrong answer you find in testing is one a customer won't see.
Review real conversations every week
After launch, read a sample of conversations each week. You'll find gaps in your information, questions you didn't expect, and places where the wording could be clearer. Each fix makes the next week better. Our AI chatbots and agents include this review as part of the setup, and the same principle (answer from approved material, guide rather than guess) shapes the AI tutor we built for Sprout Lessons.