What an LLM is, and what it means for a business
For anyone who keeps hearing the term and wants a clear picture of it. Technical words are explained the first time they appear.
The short answer
An LLM (Large Language Model) is a computer program trained on a huge amount of text. It learns patterns in how language fits together, so it can read and write in a way that sounds human.
Its basic job is simple: given some text, predict what word should come next. Do that well enough, billions of times over, and you get something that can answer questions, summarise documents, draft emails and write code.
How it works, in three steps
It reads an enormous amount of text
Books, websites, code and more. The model adjusts billions of internal settings (called parameters) until it gets good at predicting what comes next. This takes months and very large amounts of computing power.
It writes one piece at a time
When you type a question, it generates the reply a small piece at a time (a word or part of a word, called a token), each based on everything before it.
People teach it to be helpful
Trainers rate its answers, and it is adjusted to give answers people find useful and safe. This is why chat assistants feel different from raw text prediction.
Think of it as very advanced autocomplete. It doesn't look facts up in a database by default, which is why it can sound confident and still be wrong. That is called a hallucination, and it is the main thing to check.
Where you meet one
ChatGPT, Claude, Gemini and Microsoft Copilot are all apps built on LLMs. The app is the friendly front end; the LLM is the engine behind it.
Two ways to run one
Run by a company on its servers
You send your text to the provider and get an answer back. Quick to start and usually the most capable models. Your data leaves your machine, so check the provider's terms.
Run on your own computer
Open-weight models such as Llama can be downloaded and run privately with tools like Ollama or LM Studio. Nothing leaves your machine, but the models are smaller and it needs a capable computer.
Local models suit private or offline work. Hosted models suit most everyday use.
What happens in a business
A company rarely just "turns on AI". The steps that usually matter:
- Pick where it runs. Cloud contract, private cloud, or a local model, depending on how sensitive the data is.
- Check the data terms. Find out whether the provider trains on your inputs, where data is stored, and what your contract says about it. Data protection law applies here.
- Connect it to company knowledge. A method called RAG (retrieval-augmented generation) searches your own documents first and gives the relevant parts to the model, so answers are based on your material.
- Add tools, carefully. Agents are LLMs that can take actions, such as sending an email or updating a CRM. These need approval steps and logs.
- Write simple rules. What staff may paste in, which outputs need a human check (legal, medical, financial and HR decisions in particular), and who is accountable when something goes wrong.
- Measure the result. Pick one task, measure time saved and error rate, and compare against the cost of usage. Pilot first, then scale.
A simple test
Before you trust an AI answer, ask: would I let a smart new colleague send this without checking it? If not, check it first. The model is fast and fluent, but it does not know when it is wrong.
Next, learn how to get useful answers: How to ask AI for what you need.
Want the exact meaning of a word? See the Jargon Decoder.