The vocabulary, without the jargon.

01

Large language model (LLM)

The engine behind tools like ChatGPT or Claude. It predicts text extremely well, which makes it good at summarising, drafting, classifying and extracting — and unreliable at arithmetic and at knowing what it doesn't know.

02

Prompt

The instruction you give the model. In a business system a prompt is not something you type each time — it's a versioned piece of configuration that should be tested like code.

03

Agent

A model that has been given tools and permission to use them in sequence — read a mailbox, look something up, write a record. Useful when the task and boundaries are tightly defined; unpredictable when they are not.

04

RAG (retrieval-augmented generation)

Giving the model your own documents to answer from, instead of hoping it memorised them. This is what turns a generic assistant into one that knows your products, contracts or procedures.

05

Fine-tuning

Further training a model on your own examples. Occasionally the right answer, usually not: retrieval and a better prompt solve most problems for a fraction of the cost and effort.

06

Hallucination

A confident, fluent, wrong answer. It's a property of the technology, not a bug to be patched — which is why a serious system is designed around checks, sources and human review rather than trust.

07

Token

The unit models read and write in, and the unit you are billed in. Roughly three quarters of a word. It matters because it is the difference between a system that costs €30 a month and one that costs €3,000.

08

EU AI Act

European legislation that classifies AI use by risk and attaches obligations to each level. Most business use is low-risk, where the duties are transparency and AI literacy. Some uses — recruitment screening, credit scoring, biometrics — are high-risk and carry real requirements.

09

Human in the loop

A deliberate checkpoint where a person approves, corrects or overrides what the system produced. The cheapest way to make an automated process safe to switch on.

10

MCP (Model Context Protocol)

An open standard for connecting AI models to tools and data sources. It matters because it makes those connections reusable across models and vendors, instead of rebuilt for each one.

Anything here you'd like explained against your own situation? That's exactly what a first conversation is for.