Almost every AI tool a small business meets today is built on a large language model, and understanding one simple thing about them saves a lot of trouble. They predict likely text. They do not look up answers in a database of facts. Once that clicks, most of their odd behaviour stops being mysterious and starts being predictable.
It predicts text, it does not look things up
A language model is trained on a very large amount of text and learns which words tend to follow which. When you ask it something, it produces the continuation that seems most likely, one piece at a time. Unless the tool has been specifically connected to a search engine or to your own documents, nothing is being looked up anywhere.
Fluent is not the same as correct
The model is optimised to produce text that reads well, not text that is true. A wrong answer is written in exactly the same confident, tidy prose as a right one. There is no wobble in the tone to warn you, so confidence tells you nothing about accuracy.
It has no memory by default
Within one conversation it can see what has already been said. Start a new chat and it knows nothing about the last one, unless the product has added a memory feature on top. It does not remember your business, your customers, or the corrections you made yesterday.
Its knowledge stops at a fixed point
Training happens on data gathered up to a certain date, so anything after that is simply not in the model. It will not always tell you when a question falls outside what it was trained on. For prices, rules, or news, check a current source rather than trusting the answer.
The same question can give different answers
Most tools pick from among the likely next words with a degree of randomness, which is why asking twice can produce two different replies. That is normal behaviour, not a fault. It also means you cannot assume an answer is reliable just because it came out cleanly the first time.
Common mistakes
- Treating the model as a search engine and assuming it has checked a source.
- Reading confident, well written prose as a sign the content is accurate.
- Expecting it to remember your business or your earlier corrections in a new chat.
- Asking about recent prices, rules, or events without checking a current source.
- Being surprised or annoyed that the same prompt gives a slightly different answer.
- Assuming an answer applies to the UK when the model was trained mostly on other markets.
Check yourself
0/41.You ask an AI chatbot for the current rate of a UK tax and it replies instantly with a specific figure. What should you assume?
2.You correct a mistake in a chat on Monday. On Wednesday you open a fresh chat and it makes the same mistake. Why?
3.You send the same prompt twice and get two noticeably different drafts. What does this tell you?
4.Which statement best describes what a large language model is doing when it answers you?
This is general guidance, not a substitute for advice on your specific setup. Want a hand putting it into practice? Talk to us or see our care plans.