LLM (large language model)

A large language model does not work with words but with tokens. A token is a piece of text: a short word, part of a longer word, a digit or punctuation. When answering, the model works token by token and at every step calculates which token is most likely to follow everything written so far, including your prompt.

Most current LLMs are built on the transformer architecture, which Google researchers described in June 2017 in the paper "Attention Is All You Need". Its key component is the attention mechanism: at every step the model weighs which parts of the text matter for the next word, even if they appeared many sentences earlier. This lets it keep track of context in long texts and makes it efficient to train on huge amounts of data.

The practical consequence: an LLM does not "look up" facts, it generates text that statistically matches what it has learned. That is why the material and instructions you give it matter so much.

1. Pre-training: the model learns to predict the next token on a vast amount of text from the web, books and code. This gives it knowledge of language, facts and style. It is the most demanding phase and requires thousands of specialised chips.
2. Fine-tuning: on a smaller set of high-quality examples, the model learns to follow instructions and hold a conversation rather than just continue text.
3. Human feedback: reviewers compare answers and the model learns to prefer helpful, truthful and safe responses. This method is known as RLHF (reinforcement learning from human feedback).

A model's knowledge ends at the date its training data was finalised. It knows nothing about later events unless you provide them in the prompt or the tool using it can search the web.

According to OpenAI, one token in English is roughly four characters, or about three quarters of a word. Other languages usually need more tokens for the same text. Tokens matter for two reasons:

• Cost: when you connect through an API, you pay for the number of input and output tokens. We cover pricing and calculations in our OpenAI API guide.
• Context window: the maximum amount of text the model can work with at once in a conversation (prompt, attached documents and answer). It works like short-term memory. According to vendor documentation, some current models handle a million tokens or more, enough for entire contracts, manuals or large codebases.

Whatever does not fit into the context window, the model cannot see. In long conversations it may "forget" earlier instructions, so it is worth starting a new conversation with a short summary.

The LLM is the model itself, the "brain" that processes and generates text. Applications are built on top of it: an AI assistant helps a specific person with their work, a chatbot talks to a company's customers on its website, and an AI agent completes multi-step tasks on its own using other tools (more in our article on AI agents for business). One language model can power dozens of different products. We compare how specific tools built on different models differ in practice in ChatGPT vs Claude vs Gemini and cover one of them in detail in our Claude AI guide.

A large language model is strong wherever work involves text:

• writing, rewriting and translating text,
• summarising long documents and meeting transcripts,
• sorting and extracting data from emails, invoices or forms,
• answering questions from the material you provide,
• suggesting and reviewing code.

Limits you need to account for:

• Hallucinations: the model sometimes produces a convincing but false answer, such as a non-existent citation or figure. Facts must be checked.
• Outdated knowledge: without access to the web or your data, the model knows nothing about recent news or your business.
• Calculations: a pure language model is unreliable at arithmetic; precise calculations belong in a spreadsheet or code, which the model can write.
• Prompt quality: a generic prompt leads to a generic answer. Our guide How to write prompts shows how to write specific instructions.

• Ready-made app: the team works directly in an AI assistant on a business plan. The fastest start with no development.
• API integration: the LLM is built into your own processes, for example an online shop, CRM or an automation that sorts enquiries.
• Model on top of company data: for each question, the model receives relevant passages from internal documents and answers from them, often with a link to the source. This approach is called RAG (retrieval-augmented generation) and is usually cheaper than retraining a model.
• Open model on your own server: some language models can be downloaded (so-called open-weight models). Data never leaves the company, but the company carries the cost of hardware and maintenance.

Whichever option you choose, check the data processing terms: whether the provider uses inputs for training, where data is stored and for how long. When personal data is involved, GDPR rules apply.

The EU Artificial Intelligence Act (AI Act) treats large language models as general-purpose AI models. Since 2 August 2025, their providers must, among other things, keep technical documentation, have a copyright policy and publish a summary of the content used for training. These obligations apply to model developers, not to businesses that simply use a model. However, if AI talks directly to your customers, a transparency obligation applies from 2 August 2026: people must know they are interacting with AI unless it is obvious from the context. This overview is not legal advice.

Your team can start using a ready-made AI assistant on its own. An expert pays off when the language model needs to work with your data, connect to your CRM, online shop or accounting, answer customers, or when you are deciding which model and deployment option makes sense for cost and security. We build projects like these, from model selection to integration with company systems, as part of our AI solutions for business.