Home · Insights · Fundamentals
How AI chatbots work, and why most of them are bad
Most chatbots on business websites annoy people. The cause is rarely the model. It is the lack of a clear job, limits and a way out. Here is how they work and what separates a good one from a bad one.
What is actually going on inside
A modern chatbot is a language model with instructions in front of it. You type, the instructions and your message go to the model, and the model writes the most plausible reply. Nothing in that process knows whether the reply is true. It is why a chatbot can sound equally sure when it is right and when it is making something up, a pattern we described in the jagged frontier.
What makes a chatbot good or bad is therefore not the model. It is what surrounds it: what it has been told, what it is allowed to say, and where it hands over to a person.
Why most business chatbots disappoint
- No clear job. A bot that is meant to answer questions, take bookings, handle complaints and qualify leads does none of them well.
- No limits. Without rules about what it must not say, it will fill gaps with confident guesses about prices, availability and policy.
- No exit. A visitor stuck in a loop with a bot and no route to a person leaves, and remembers.
- Nobody owns it. The answers go stale and nobody is told.
The industry data points the same way on the pressure and less clearly on the results. Gartner reported in February 2026 that 91 percent of customer service leaders feel pressure from executives to implement AI. Pressure is not performance, and we have not found a source we trust that shows small-business chatbots resolving most enquiries on their own, so we do not quote one.
What a good one looks like
| Bad pattern | Good pattern |
|---|---|
| Tries to do everything | One job, stated in the first line of its instructions |
| Answers from general knowledge | Answers from your own content where you can, and says when it does not know |
| Invents specifics under pressure | Is told never to state a price, a result or a name it was not given |
| Traps the visitor | A person is always available, and the handover keeps the context |
| Left alone after launch | Someone reads real conversations each week and updates the rules |
How we do it
The assistant on our own site follows the right-hand column. It keeps to one job, will not state a price or make up a client, offers a next step once, and hands over to a person with a summary so nobody has to repeat themselves. We describe the build in how we built the chat widget.
The question to ask any vendor is not which model they use. It is: what is this bot not allowed to say, who reads its conversations, and what happens when it does not know? A confident answer to all three is worth more than a long feature list. This is how we approach AI customer service on Mallorca, done properly.
Sources, checked 9 October 2026
- Gartner press release, 18 February 2026: 91 percent of customer service leaders report pressure to implement AI in 2026 (headline figure).