Large Language Model (LLM)

What an LLM is

A large language model is an AI system trained on enormous quantities of text to predict what words should come next. ChatGPT, Claude, and Gemini are all built on one.

That description sounds reductive, and it is worth sitting with rather than dismissing. Predicting the next word well enough, across enough text, produces something that can summarize a contract, draft an estimate, or explain a regulation. But the underlying mechanism is still generating plausible language — and plausible is not the same as true. Nearly every practical rule for using these tools safely follows from that one fact.

What they are genuinely good at

For a small business, the reliable wins are consistent:

• Drafting from your material — turning notes into a quote, a job into a service description, a complaint into a measured reply.
• Reformatting and summarizing — long email threads into a decision, a spec sheet into plain English.
• First drafts of repetitive writing — listings, descriptions, follow-ups where you edit rather than start blank.
• Explaining unfamiliar territory — a contract clause, an error message, an industry term.

The pattern: they are strongest when you supply the facts and ask for shape, and weakest when you ask them to supply the facts.

Where they fail

• Hallucination. Confident invention of statistics, citations, prices, part numbers, and regulations. This is the failure mode that causes real damage, because the output looks exactly like the correct output.
• No inherent access to your business. A model knows nothing about your pricing, inventory, or customers unless something puts that information in front of it. See retrieval-augmented generation.
• Training cutoffs. Models are trained up to a date. Anything after it is unknown unless the tool can search.
• Confident wrongness at the edges. Accuracy degrades on niche, local, and regulatory specifics — exactly where a small business’s questions tend to live.
• Prompt injection. Content the model reads can carry hidden instructions, which matters the moment you automate anything that processes incoming email or documents.

Practical rules that hold up

1. Never publish or send unreviewed output where a factual error costs money or credibility. Human-in-the-loop approval is the control that makes the rest safe.
2. Give it the facts rather than asking for them. Paste your actual pricing and specs; do not trust its recollection.
3. Verify every number, date, citation, and regulation independently. This is where hallucinations concentrate.
4. Assume anything you paste may be retained unless you are on a plan that states otherwise in writing. Customer data and credentials deserve that caution.
5. Judge it against the realistic alternative. A drafted reply you edit in two minutes beats one you write in fifteen — and beats the one you never sent because you were busy.

The framing that keeps you out of trouble

Treat an LLM as a fast, capable, tireless assistant with no institutional memory and no ability to tell you when it is guessing.

You would not let that person send a quote to a customer unreviewed. The same judgment, applied consistently, covers almost every AI risk a small business will actually encounter.