An LLM predicts text piece by piece from what is in its context — the instruction, the documents and the tool results sent along with it. It knows nothing about your business that is not in there, and it can state something wrong just as convincingly as something right. An LLM is never better than the context it gets, which is why its value is decided more often by data, tools and guardrails than by the choice of model. A good introduction is on Wikipedia.
In a product the model is connected to real systems through APIs and the Model Context Protocol; in a repository it works as a coding agent and reads an AGENTS.md to learn the rules. LLM-based features need to be tested, measured and improved in real use — just like any other feature.