A language model is not a search engine, a database, or a brain. It is a very large statistical machine trained on enormous amounts of text to answer one question about language: given everything that came before this word, what word likely comes next? That sentence sounds too simple to matter, which is exactly why the technology surprises people.
Because predicting the next word across billions of examples forces the model to learn grammar, facts, formats, tone and reasoning patterns, it can draft your follow-up email, summarize a contract, restate a customer complaint, or answer a question in fluent sentences. It does all of this without a database behind it. Which is where the misunderstandings start.
A model without your business data behind it is a well-read stranger. It knows what a receptionist does. It does not know your hours, your pricing, or that you do not service that ZIP code. That is why a real deployment pairs the model with your approved content: your pages, your price list, your policies, your calendar rules. The model supplies language ability. Your data supplies truth.
Second misunderstanding: it never guesses on purpose and never knows on purpose. Both are statistics. That is fine for a first draft you will review. It is not fine for a customer-facing promise with no guardrails, which is why well-built assistants are restricted to answering only from approved material and escalate the rest to a human.
Third: it is not watching you. A hosted provider stores and reads what you send unless your contract says otherwise. A local model on your own hardware sees nothing leave your building. Same words, very different arrangements.
None of this is an argument against the technology. It is an argument for knowing what you are buying. The useful framing for small businesses: a language model is an engine, and everything you have read about AI products is really about the car built around that engine: whose data loaded the map, whose servers run it, and what the speed limiter allows.