An LLM is a large language model. It has read an enormous amount of text and got very good at guessing the next piece. That is why it can draft, summarise, translate, and answer in sentences that sound sure of themselves. It is the engine inside most of the products sold as generative AI. It is not a database of your business, and it is not a colleague who checked the file.
Prediction, not lookup
Ask it for a policy and it will write something that looks like a policy. If the real policy was not in the prompt, and was not retrieved from your own documents, the model is completing a pattern. It will invent a clause, a date, a percentage. The invention is fluent. Fluency is not evidence. This is the same limit described in what generative AI is: good at language, weak at facts it was never given.
Grounding fixes the useful case. You find the passages in your own files that bear on the question, and you hand those passages to the model with the question. That technique is RAG. The model still writes the sentences. The sentences are now about material you can open and check.
Which model is the wrong first question
Teams lose months comparing model names. The project is elsewhere. What documents may be sent? Who is allowed to see the answer? What does "wrong" look like, and who notices? A brilliant model wired to every folder in the company is a leak with good grammar. A modest model wired to the current price list, with a person checking the draft, is a tool.
Do not paste client files into a public chat box and call it a pilot. The documents left the building. You will not get them back.
When it matters on a project
Use an LLM when the slow part is reading or drafting: invoices to fields, a long thread to a summary, a first reply a person will edit. Do not use one when the rule is already a formula. If the amount is over a limit, send it to a manager. That is a workflow. It does not need a model.
The build, when you are ready, sits under AI integration: your data, your permissions, and a person still responsible for what goes out. The longer map of where this pays off is AI for business.