What today's AI models are

When people say "AI" at work today, they usually mean tools built on large language models (LLMs): the technology behind AI chat assistants such as ChatGPT, Claude, Gemini and Microsoft Copilot. This lesson explains, without maths, what these models are and why they behave the way they do. Everything else in the course builds on it.

What a large language model is

A large language model is a computer program trained on a very large amount of text: books, websites, articles, code and more. During training it learns patterns in language: which words and ideas tend to follow which, how arguments are structured, how a polite email differs from a contract.

When you give it some text, called a prompt, it writes a response by predicting, piece by piece, what is likely to come next, given everything in the conversation so far. The pieces are called tokens: roughly words or parts of words.

That simple mechanism, at enormous scale, turns out to be capable of a lot: summarising, drafting, translating, explaining, classifying, extracting information and writing code.

Four things that follow from how they work

1. They produce plausible text, not checked facts. A model generates what sounds right. Usually that is also correct, but not always, and when it's wrong it can be wrong fluently and confidently. This is called hallucination, and you'll learn to manage it in this course.

2. They only know what they were trained on, plus what you give them. A model's built-in knowledge stops at a point in time, and it knows nothing about your business unless you tell it or connect it to your documents. Some assistants can search the web or read files you upload, which helps, but the answer is only as good as those sources.

3. They work within a context window. A model can only consider a limited amount of text at once: your instructions, any documents you add and the conversation so far. The limits are large now, but very long conversations or documents can still exceed them, and details buried in a long input can be overlooked.

4. The same prompt can give different answers. There is some randomness in how responses are generated. That's useful for brainstorming and awkward when you need consistent results, which is why business workflows need clear instructions, examples and checks.

Models, assistants and integrations

• The model is the underlying engine.
• The assistant is the product you use, such as a chat app, which adds features like file uploads, web search or memory.
• Integrations put a model inside tools you already use: email, documents, spreadsheets, customer support software and automation platforms.

Products change quickly. Features that exist today may be renamed, moved or removed next month. This course focuses on skills that stay useful whichever product you use.

A worked example

Ask an assistant: "Write a two-sentence description of a family-run bakery for its website." You'll get something fluent in seconds, and it will probably invent details, perhaps a founding year or a speciality the bakery doesn't make. The model isn't lying. It is producing a plausible description of a bakery, because it has no information about this one.

Give it the facts instead: the name, what they bake, when they opened, what customers praise in reviews. Now the draft is useful, and you can check every claim against what you provided.

How this course works

There are five sections: what AI can and can't do; prompting and checking the work; everyday business uses; automating workflows; and using and selling AI responsibly. Each lesson ends with an exercise. Use any AI assistant you have access to, and never paste confidential, client or personal information into it while you practise. The final project has you design and document an AI-assisted workflow for a local business, which our staff review.

Exercise

Ask an AI assistant to describe a real local business you know, giving it only the business's name. Note every detail it gets wrong or invents. Then ask again, this time giving it five true facts, and compare the two answers.