A chatbot invents answers because it was built to always produce one. A language model predicts plausible text; without a source to answer from and permission to say it does not know, plausible is all you have asked it for. The fix is architectural, not a better model.
Answer from a source, not from memory
An agent that answers from your own documents and your own database is a different piece of software from one improvising out of what it absorbed in training. The first can be checked, corrected by editing a document, and made to cite where an answer came from. The second cannot be corrected at all, only re-prompted and hoped over.
This is the single largest determinant of whether a business chatbot is useful, and it is decided before any model is chosen.
Give it permission to refuse
Most business chatbots are never told that "I do not have that information, here is how to reach someone who does" is an acceptable output. Add it explicitly and a large class of confident nonsense disappears, because the model now has a correct answer available for the case where it has no source.
Customers are also far more forgiving of a bot that says it does not know than of one that confidently misquotes a price.
Put a person in front of anything that costs money
Quoting a price, promising a delivery date, agreeing a refund, sending anything outside the company: these go to a person for approval before they are sent. Not because the model is usually wrong, but because the occasional confident error in this category is the one that costs real money and the goodwill around it.
Log what it was asked and what it said
Without a record you cannot debug it, cannot improve it, and cannot answer a customer who says your website told them something. The log is what turns a chatbot from a black box into a system you operate.
What this does not fix
None of this makes a model good at judgement. If a question genuinely requires weighing things nobody has written down, the correct design is to route it to a person quickly rather than to answer it well.
AI Agents Automation