The AI Jargon Decoder

Every term you keep hearing in meetings, articles and sales pitches — in one honest sentence each. Bookmark it, share it, wave it at the next person who says "leverage".

AI

Software that produces answers, text or decisions that used to need a person. Not magic, not conscious — very good pattern-matching.

LLM (Large Language Model)

The engine behind tools like ChatGPT: a program trained on enormous amounts of text so it can read and write like a person.

Model

One specific trained AI. Models come in sizes — bigger is usually cleverer, slower and hungrier for hardware.

Training

The one-off process of teaching a model from data. You almost certainly never need to do this — using a model is not training one.

Prompt

Whatever you type at an AI. The whole "skill" of prompting is mostly just asking clearly.

Hallucination

When an AI states something false with total confidence. The industry's polite word for making things up. Good systems cite their sources so you can check.

The cloud

Someone else's computer. Convenient, capable — and your data is on it.

On-premises (on-prem)

The opposite: the software runs on hardware in your building, so your data never leaves your network.

Open-weight models

AI models anyone can download and run on their own hardware — no account with a tech giant, no data sent to one.

Inference

The act of a model answering. When someone sells "inference", they're selling the running of AI, not the building of it.

GPU

The graphics chip that happens to be superb at AI maths. The reason AI hardware is expensive and gamers are annoyed.

RAG (Retrieval-Augmented Generation)

A technique where the AI looks things up in your documents first, then answers from what it found — so answers come from your files, not its imagination.

Vector database

A filing system that stores meaning rather than words, so "invoice" also finds "bill". Plumbing — you'll never touch it.

Fine-tuning

Adjusting an existing model with your own examples. Occasionally useful, usually oversold — most businesses just need RAG (see above).

Agent

An AI given permission to take actions — send the email, update the record — rather than just suggest them. Powerful; deploy with care.

Token

The chunks AI reads and writes text in (roughly ¾ of a word). Cloud AI is often billed per token — which is how usage bills surprise people.

Per-seat pricing

You pay for every person, every month, forever. Every new hire raises the bill. The alternative is owning the system outright.

Air-gapped

A computer physically disconnected from the internet. The maximum-security end of "your data never leaves the building".

Automation

Software doing a repetitive task by fixed rules — no AI required. Honest advisors will tell you when this is all you need, because it's usually cheaper.

Shadow AI

Staff quietly using free AI tools with company information — pasting quotes, emails and contracts into websites that may keep them. Most businesses have it; few have decided anything about it.