LLM. The engine inside AI assistants, explained without jargon.

An LLM (large language model) is a program trained on huge amounts of text to predict the next piece of text, which makes it able to write, summarise, translate and answer in natural language. It's the engine inside assistants such as ChatGPT, Claude and Gemini.

This entry is part of the AI and automation glossary, where every term has a short definition. Here the definition goes further: how an LLM works, the four words you need to understand it, what it can do in a small business, where it goes wrong and what European law says about it.

What an LLM is

An LLM is an artificial intelligence model trained on text, with billions of parameters, that produces new text from what it receives. “Large” refers to size: the amount of text used to train it and the number of parameters, the internal values adjusted during training.

The architecture almost every LLM is based on is the transformer, proposed in June 2017 in a paper by Ashish Vaswani and colleagues titled “Attention Is All You Need”. The key mechanism is attention: the model decides, word by word, which other parts of the text to look at.

How it works, in four words

Four words are enough to understand what goes on inside an LLM, and each has a practical consequence for a business using it: what it costs, what it knows, what it keeps in mind during a task, and how repeatable its answers are.

Word	What it means	Why it matters to you
Tokenthe unit of text	The piece of text the model reads and writes: a word, part of a word, a symbol.	Usage is paid for by the token, and the length of a document is measured in tokens.
Trainingwhen it learns	The phase in which the model learns from text by adjusting its parameters. It happens beforehand, once, and it's expensive.	The model doesn't know what happened afterwards, or anything about your documents.
Context windowwhat it keeps in mind	How much text the model takes into account at once: the question, attached documents, the conversation.	Whatever isn't in the window doesn't exist as far as the model is concerned.
Temperaturehow much it varies	The setting that controls how predictable or varied the answer is.	For business work it's kept low: less imagination, more repeatable answers.

What it can do in an SME

According to Istat, among Italian businesses using artificial intelligence in 2025, 70.8% use it to extract information from text documents and 59.1% use generative AI, whether written, spoken or visual. Those are the two jobs where an LLM delivers most: reading and writing.

In practice, an LLM reads an email in any language and understands what it's asking, extracts the data from an order or invoice, drafts a reply, a product sheet or a quote, summarises a long document, and translates while keeping the tone. Deciding whether the draft is right stays with a person.

Where it goes wrong

An LLM predicts probable text, and probable text can be false. When it doesn't know the answer, it tends to produce a plausible one anyway: that's a hallucination. In an article of 5 September 2025, OpenAI traces it to training and evaluation procedures that reward guessing over admitting uncertainty.

The second limit is the date. The model knows the world only up to its training, and doesn't know your documents. Both limits are kept in check by giving the model the right sources at the moment of the question, using the technique called RAG, and by putting a person wherever a mistake is costly: the human in the loop.

LLM, chatbot and agent

The LLM is the engine; a chatbot and an agent are two ways of using it. A chatbot puts the model in a conversation and produces answers. An agent gives it tools to act with, such as reading an inbox or writing to business software. The difference between the two is explained on the page AI agent or chatbot.

What the AI Act says

The AI Act, in Article 3, point 63, defines a “general-purpose AI model”: a model trained “with a large amount of data using self-supervision at scale”, displaying significant generality and capable of performing a wide range of distinct tasks. Large LLMs fall under it, and the main obligations fall on the companies that develop them.

For a business using them, two rules already in force matter most. Staff training, under Article 4, from 2 February 2025, explained on the page about the training obligation. And transparency towards people talking to a system, under Article 50, from 2 August 2026.

How Itria uses it

Itria uses LLMs inside systems with rules written in advance, never on their own. In Inbox AI the model reads enquiries and prepares a draft: on a test bench on 4 September 2026, with 100 enquiries across 4 channels and 5 languages, the draft was ready in a median of 7.290 seconds, in the right language in 97 cases out of 100.

The same models prepare the multilingual drafts and extract data in document entry. In every case, anything that commits the business goes through a person before sending, and the list of models used is available on request, as stated on the AI transparency page.

Related terms

AI hallucinations

A false statement made with confidence by a model. The best-known limit of LLMs, and the easiest to keep in check.

RAG

The technique that gives the model the right documents at the moment of the question, so it answers from them.

Prompt

The written instruction given to the model. The quality of the answer depends a lot on how it's written.

Generative artificial intelligence

The family of systems that produce new text, images or sound. LLMs are the part that writes.

Questions and answers