How AI language models work

AI tools can now write emails, summarise reports and answer questions in seconds. For a lot of small businesses they've gone from novelty to everyday tool almost overnight. But what's actually happening when you type a question into one?

It predicts, it doesn't look things up

The tools behind ChatGPT, Microsoft Copilot and Google Gemini are called large language models. A language model is trained on an enormous amount of text: books, websites, articles, code. From all that reading it learns patterns, mainly which words tend to follow which.

When you ask it something, it isn't searching a database for the right answer. It predicts a likely next word, then the next one, and keeps going one piece at a time. Trained on enough text, those predictions become remarkably good. Good enough to draft a proposal, explain a tricky concept or work through a problem step by step.

Why it can be confidently wrong

That strength is also the catch. Because the model is predicting what sounds right, it can produce an answer that reads perfectly and is still wrong. A made-up statistic, a law that doesn't exist, a reference to a report nobody wrote. This is called a hallucination, and it's delivered with the same confidence as everything else.

So treat AI output as a first draft, not a final answer. Check facts, figures, names and anything important before you rely on it or send it to a client.

Be careful what you paste in

The second risk is what goes into the tool. Anything you type or paste may be stored, reviewed or used by the provider, depending on the product and your settings. Client details, staff records, financial data and passwords shouldn't go into AI tools your business hasn't approved.

Business versions of these tools, like Microsoft 365 Copilot, usually come with stronger data protections than free consumer accounts. It's worth knowing which one your team is using.

Set some simple rules

You don't need a 40-page policy. A one-page AI guideline covering a few points goes a long way.

• Which AI tools are approved for work
• What information must never be entered
• That important output gets checked by a person
• Who to ask when unsure

The bottom line

Used well, with clear rules, AI can take a lot of busy work off your team: first drafts, summaries, meeting notes and research starting points. Used carelessly, it can leak data or put confident mistakes in front of clients. A little structure lets you get the benefit without the risk.

If you'd like help choosing the right AI tools or writing a simple policy for your team, Nexcentric can help. We also have some useful free tools at nexcentric.com.au/tools.