When the bot becomes the problem
A customer writes: "Do you have this in black, size medium?" Instead of "Yes, in stock", they get: "Welcome! For orders press 1, for opening hours press 2, for a representative press 3." They press 1, the bot asks what they want again, and they leave to ask someone else.
Chatbots are supposed to save time. Set up badly, they cost you customers instead. This guide covers what goes wrong, with real cases, the six mistakes behind most chatbot problems, and how to avoid them.
Can AI chatbots make mistakes?
Yes, and there are two different kinds of mistake, with different fixes:
- Menu bots (menus and flow charts) make mistakes of understanding. They only recognize what someone planned for, so anything phrased differently gets "Sorry, I didn't understand."
- AI chatbots that write free text make mistakes of fact. AI chat tools are built to sound fluent, not to check facts. If nothing limits them to your information, they can fill a gap with something that sounds right but isn't: a policy you don't have, or a discount you never offered.
Three real chatbot failures, and what they teach
These are well-known cases, but the lessons apply to a five-person business just as much.
Air Canada: the bot invented a policy, and the company paid. In 2024, a Canadian tribunal ruled on a case in which Air Canada's website chatbot told a customer he could apply for a bereavement discount after his trip. The airline's actual policy didn't allow that. Air Canada argued, in effect, that it wasn't responsible for what its chatbot said. The tribunal disagreed: a company is responsible for all the information on its website, whether it comes from a static page or a chatbot, and it ordered the airline to pay him damages (Moffatt v. Air Canada, 2024 BCCRT 149, as reported by the BBC).
Lesson: whatever your bot tells a customer, the customer hears it from you.
DPD: an update broke the bot's behavior. In January 2024, after a system update, the parcel company DPD's chat assistant swore at a customer and criticized the company when he prompted it to. DPD switched off the AI part of the chat and said the error had followed the update (BBC).
Lesson: test again whenever something changes, and keep a way for people to reach a human.
A car dealership's bot "agreed" to sell a car for $1. In December 2023, a user told a US dealership's AI chatbot to agree with everything the customer said, then offered $1 for a new SUV. The bot replied that it was a deal. No car was sold, but the screenshots spread widely (Gizmodo).
Lesson: a bot should never be able to make commitments, such as prices or discounts, that you didn't give it.
Mistake 1: Making customers fit your menu
People don't write in menu options. They ask two things at once, make typos, and change their mind halfway through. A menu bot can't follow that, so the customer has to do the work of fitting the bot.
Instead: let customers write normally, and make sure whatever answers them can handle "how much is it and can I pick it up today?" in one go.
Mistake 2: Letting the bot guess
This is the costly one, as Air Canada found out. Imagine a customer asks "Do you deliver to the coast for free?" and an AI bot with no limits answers "Yes, free delivery!" because it sounds helpful. Now you either honor a promise you never made or disappoint a customer.
Instead: the bot should answer only from information you gave it. When it doesn't know, it should say so and pass the question to you. It shouldn't refuse on its own initiative either: "we don't do that" is also a fact, and it needs a source like any other.
Mistake 3: A bad fallback message, or none
The fallback message is what the bot says when it can't help. "Sorry, I didn't understand your request" on repeat is how customers learn to hate a bot.
A good fallback:
- admits it plainly: "Good question, let me check that for you";
- says what happens next: "I'll get back to you shortly" (and then someone actually does);
- makes sure a person sees it: the owner gets the question with enough context to answer quickly;
- never loops: one fallback, then a handoff, not the same menu again.
Mistake 4: No way to reach a person
Some customers want a human: someone angry, a large order, a sensitive question. If they type "I want to talk to someone" and get the menu again, you've probably lost them. In a 2026 Gartner survey of 3,566 customers, 87% said it is essential for companies using generative AI in customer service to offer an option to reach a human agent (Gartner). On WhatsApp it's also a rule: the WhatsApp Business Messaging Policy says businesses that automate replies must have "prompt, clear, and direct escalation paths".
Instead: a request for a person should always reach a person. So should signs of impatience ("hello??"). Decide in advance which topics always come to you, such as complaints, refunds or custom quotes.
Mistake 5: Forgetting the conversation
Customer: Can I get the price list?
Bot: sends the price list
Customer: Does that include VAT?
Bot: Sorry, I didn't understand. For orders, press 1.
A bot that treats every message as a new conversation forces the customer to start over. Whatever answers your customers needs to read the conversation so far, not just the last line.
Mistake 6: Set and forget
Prices change, hours change, you add a service. A bot built once and never updated starts giving old answers, often without anyone noticing.
Instead: updating should be as easy as telling an employee. If changing an answer means opening an editor or calling whoever built the flow, it won't happen. And after any change, test it again (see DPD above).
How to avoid all six: a short checklist
- Write down your real answers. Go through the last two weeks of WhatsApp, list the questions that come up again and again, and write the answer you'd want a good employee to give. Be honest: if delivery sometimes takes five days, write five. (How to do this step by step.)
- Decide what always comes to you. Complaints, refunds, unusual discounts, anything medical or legal. Write the list down.
- Choose a tool that answers from your information and hands off when it doesn't know, rather than one you have to script or one that improvises.
- Test it with your own questions before any customer sees a reply, including awkward ones and attempts to talk it into a discount.
- Review regularly. Read a sample of conversations each week and fix any answer that customers still ask follow-up questions about.
How Answr handles these problems
Answr is an assistant on your WhatsApp, designed around exactly these risks:
- No menus or flows. Customers write normally, in their own language, and Answr replies in that language. (Flow builders and learning assistants compared.)
- It answers from what you've taught it, and says when it doesn't know. You teach it in a WhatsApp chat: type answers, send files such as your price list, or send a link to your website. When something isn't covered, it's built to check with you rather than guess or refuse on its own. It tells the customer it will check and sends you a short summary on WhatsApp. (How Answr avoids guessing.)
- It doesn't invent offers. It only offers what you've taught it: no made-up discounts, deadlines or products.
- People can reach you. It passes the conversation to you when a customer asks for a person or seems impatient, and you can name topics that always come straight to you.
- It answers the bot question plainly. If a customer asks whether they're talking to a bot, it says it's an automated assistant and carries on helping. You can also set it to introduce itself that way from the first reply.
- Updating is a message. Tell it the new price or hours in the chat. When you reply to a customer yourself, it can pick up your answer for next time, and it asks you first when it isn't sure.
- Nothing goes live until you say so, so you can test it first. Start with automated WhatsApp replies.
It's still only as accurate as what you've taught it. If your information is out of date, its answers will be too, which is why the weekly review above still matters.
When not to automate at all
Some conversations shouldn't be handled by any bot:
- Emergencies and serious complaints, where someone needs a human response now.
- Medical, legal or financial advice that depends on the person's situation. Automation can handle the logistics (hours, booking, directions), not the advice.
- Anyone who has asked for a person. Making them argue with a bot only makes things worse.
Frequently asked questions
Why do chatbots fail so often?
Usually because they were built as menus that people have to fit into, or as free-text AI with no limit on what it can claim. What works is the combination: understand natural language, but only state facts that came from the business, and hand off the rest.
What should a chatbot fallback message say?
That it will check, what happens next, and nothing else. Then make sure a person actually follows up. A fallback that loops back to the menu is worse than none.
Can I stop an AI assistant from inventing things about my business?
Choose one built to answer from the information you provide and to hand off when something isn't covered. Then test it with questions it shouldn't be able to answer, and with attempts to talk it into a deal. With Answr, those questions are meant to come to you instead of getting a made-up answer.
Will customers be annoyed that it isn't a person?
A correct answer quickly, plus a person when needed, matters more than whether the first reply came from a human.
Want an assistant that answers your customers on WhatsApp from what you've taught it? Start with a free first month, no credit card needed.
