AI in 10 minutes.

A model is the engine inside an AI tool. ChatGPT, Claude, Gemini and Copilot are apps; underneath each one is a model, a very large set of learned patterns that turns what you type into a reply. When a headline announces a “new model,” it means a new engine, not necessarily a new app. Most makers offer a range, from large models that are slower and more capable to small, cheaper “mini” models that are quick and good enough for everyday jobs.

What a language model does is simpler than it sounds: it predicts what comes next. Given your message, it picks a likely next word, adds it, and repeats, very fast, until the reply is done. Everything else, from answering questions to summarizing a contract or drafting a product description, is built from that one move.

That one idea explains most of what AI gets right and wrong. It’s fluent because it has learned from a vast amount of writing. It can be confidently wrong because it produces what sounds likely rather than looking facts up. So when an answer matters, check the names, numbers and dates: the parts that sound right whether or not they are.

A model does its learning before you ever use it, in a process called training. It’s shown an enormous amount of text (web pages, books, code and more) and practices predicting the next word, adjusting itself slightly after every guess. People then coach it with example answers and ratings, so it follows instructions like a helpful assistant. The result is a very large file of numbers, called weights: patterns, not a library of documents it can look things up in.

Its training data ends at a cutoff date, and the model’s knowledge freezes there. It knows nothing that happened afterward unless the tool can search the web or you paste the news in. It also knows little or nothing about your business. Your prices, customers and ways of working probably weren’t in its reading, so you have to tell it.

One common mix-up: chatting doesn’t retrain it. The model is the same tomorrow, whatever you told it today. What can happen, depending on your plan and settings, is that the company keeps your conversations and uses them to train a future version. That’s why it pays to check your data settings before pasting in anything sensitive.

A prompt is whatever you give the AI to work from: your question, your instructions, examples, pasted material. Because the model predicts what comes next from what it’s given, a vague prompt gets the most average possible answer. Ask for “a marketing email” and you’ll get one that could have come from any business.

Give it what you’d give a capable new hire instead: who it’s for, what you want, what good looks like and the facts it needs. “Write a friendly, two-paragraph email to our regulars announcing new Saturday hours, 8 to 2, in the tone of the example below” gets you something usable. Examples of your own writing work especially well.

Treat the first answer as a draft, not a verdict. Follow up: make it shorter, warmer or more specific, or ask it to explain its choices. When a prompt works, save it where your team can find it, so nobody has to reinvent it. None of this is technical. If you can brief a colleague clearly, you can brief an AI.

The context window is everything the AI can see while it writes a reply: its instructions, the conversation so far, any files you’ve added and your latest message. Think of it as the AI’s desk. Today’s desks are often roomy enough for a long report or several documents at once, but every one has edges. Size is measured in tokens (words and pieces of words), and your plan may allow less than the model can handle.

That limit explains a lot of odd behavior. In a very long chat, early details get buried or dropped, so the AI drifts from what you agreed at the start. And a new chat starts with a clear desk: unless your tool saves notes between conversations, it won’t remember yesterday.

It also points to the fix for most “it doesn’t know my business” problems: put what it needs on the desk. Paste the price list, attach the policy, or save standing instructions once so they’re included every time. Just remember the flip side. Whatever goes into the window has been shared with the service, so mind what you paste.

A chat assistant answers. An agent acts. Give an agent a goal and access to some tools, such as your inbox, calendar, a browser or a spreadsheet, and it plans steps, takes them, checks the result and carries on until the job is done or it needs you. Under the hood it’s the same kind of model you chat with, given tools and permission to use them in a loop.

That’s why agents are one of the big AI stories of 2026: sorting mail, booking appointments, researching suppliers, preparing orders. It’s also why they need more care. A mistake in a chat is a bad paragraph you can ignore. A mistake by an agent can be an email sent, a booking made or money spent.

So the rule is simple. Agents handle the routine steps, and a person approves anything sent to a customer, anything involving money or contracts, and anything that can’t be undone. Start with one task and one tool, keep the approval step on, and give it more room only once it has earned it. And check whether you need an agent at all: a plain automation that runs the same steps every time is often cheaper and easier to trust.