What is an LLM? The brain behind ChatGPT, in plain language

In short: a language model keeps predicting what most likely comes next.

1. 01Your text is cut into tokens. A token is a piece of text, often part of a word. That's what the model calculates with.
2. 02The model looks at everything it has at that moment: your question, the documents you provide and the conversation so far.
3. 03It predicts the next token, then the next one, and so on, until the answer is complete.

The reason it works so well is the training. The model has seen so much text that it has picked up patterns in language, knowledge and reasoning. After that, it was trained further to answer helpfully and safely.

The key thing to remember: a language model predicts language, it doesn't know anything for certain. That's why a wrong answer sounds just as convincing as a right one.

• Token. The piece of text the model calculates with. The price of AI services is also per token: you pay per million tokens in and out. An average business email is a few hundred tokens.
• Context window. How much text the model can take in at once. If something no longer fits, it forgets the beginning, or you have to split it up.
• Reasoning model. A model that first thinks in intermediate steps. Better at calculating, planning and difficult questions, but slower and more expensive. For simple work, a faster model is often enough.
• Open-source model. A model you can download and run yourself, for example on your own server. More control over your data, but you also take care of the computing power and management yourself.

Good at: writing and rewriting, summarizing, translating, extracting information from text (an invoice, a CV, an email), answering questions from documents you provide, writing code and working out ideas.

Less good at, without help: current facts (the model only knows what it saw during training), exact calculation, and anything where a single mistake has major consequences. It can also make up things that sound real. That's called a hallucination.

The solution is rarely a smarter model, but a smarter setup: let it answer from your own documents with the source included (RAG), give it access to the right systems, and have a human approve what matters.

• An LLM is the model: it can read and write, but doesn't do anything by itself.
• A chatbot or assistant is an LLM with a conversation around it. You ask a question, you get an answer.
• An agent is an LLM that gets a goal and takes steps by itself: reading an email, looking up an order in the ERP, preparing an answer. For that it needs access to systems, for example through MCP or a workflow.

The further you go, the more it delivers, and the more important the agreements about permissions and control become. Also read what is an AI agent.

There is no best model. There is a best model for a task. What counts for us in practice:

• Quality on your work. A model that scores well on a general test isn't necessarily the best at your quotes or customer emails. Test on your own examples.
• Speed and price. For simple, frequent work, a faster and cheaper model is often better than the most expensive one.
• Where the data goes. Business versions with a data processing agreement, and where needed in the EU or within your own environment.
• What you already use. If your team works in Microsoft 365, Copilot is the obvious choice; for your own applications, you choose per task.

We test models ourselves on ordinary office work and publish the results in our benchmarks. There you can see which model fits which work and what it costs. A comparison of the well-known apps is in ChatGPT, Copilot or Claude.

That depends on the version and the agreements. Business versions from the large providers don't train on customer data by default, and work with a data processing agreement. Free consumer versions don't offer that certainty.

In practice: record which data may go into which tool, use business accounts, and run sensitive applications where you have control over the data. More in using AI safely.