The mistakes that annoy customers are pretending to be human, guessing at answers, trapping people with no way to reach a person, repeating stale facts, pestering with follow ups, and giving advice it should not. Prevent them by telling people it is an AI, limiting it to approved answers, handing over with the full chat, and testing before launch.
Many people have been stuck in a bad chatbot, and they remember it. If your assistant is the first thing a new customer meets, one of these mistakes can lose the sale before you know there was one.
1. Pretending to be human
A customer who finds out later that they were talking to software feels tricked, even if the answers were fine. Say it in the first message: this is an AI assistant, and a person is available on request.
2. Guessing
A made up price, a promised discount or an invented policy is worse than no answer. Limit the assistant to answers you have approved. When it does not know, it should say so and tell the customer that a person will reply, then pass the chat on.
3. Leaving customers stuck
If someone asks for a human, they get one. The handover should carry the full chat history, so the customer never has to explain their problem twice. Also decide who picks up, and how fast. A handover nobody sees is a dead end.
4. Repeating stale facts
An assistant repeats whatever it was last told. If your hours, prices or offers change and the answers do not, every customer who asks gets the wrong thing. Update answers the same day, and read our post on running an assistant week to week for a routine that makes this automatic.
5. Writing like a brochure
Long, stiff replies are tiring on a phone. Write answers the way you would text a customer: short, plain, in your own voice. A good test is to read the reply aloud. If you would never say it, rewrite it.
6. Pestering
Follow ups help only when they are limited. Set a maximum number of messages, stop the moment the customer replies or books, make opting out a single word, and approve the wording yourself. Our post on why enquiries go cold shows a routine that stays polite.
7. Giving advice it should not
An assistant for a clinic must not comment on symptoms, and one for an accountant must not decide what is legal. Set hard limits on medical, legal and financial advice and send those questions to a person. In the same way, it should send marketing only to people who have opted in.
8. Skipping the test
Most of these mistakes show up in the first hour of real testing and never in a demo. That is why we test every assistant against at least 50 tricky questions before launch, with the owner testing first, and then read every conversation for the first seven days.
How do you test before you launch?
Try questions that customers really send, and especially the awkward ones:
- Misspelled and half finished messages.
- Questions about something you do not offer.
- Someone who is angry or upset.
- A request to bargain on price.
- A question that needs a person, such as a complaint.
- A direct request to speak to a human.
Fix every wrong answer, then ask the same questions again.
Frequently asked questions
Should an assistant say it is an AI?
Yes. Say it in the first message, and let the customer ask for a person at any time.
Can an AI assistant still give a wrong answer?
Yes, which is why it should only use approved answers, hand over when unsure, and have its conversations reviewed.
How many questions should I test?
We use at least 50 before launch. Include awkward ones, not just the common ones.
Who should read the conversations?
Someone with the authority to fix an answer. In the first week that should be every conversation.
