Guides / AI glossary

Guide · Reference

The Plain-English AI Glossary

01The basics

AI (Artificial Intelligence)

The umbrella term for software that does things which normally need human intelligence — understanding language, recognising images, making predictions. Today, when someone says "AI" they almost always mean the generative kind below.

Machine learning

Software that learns patterns from examples rather than following hand-written rules. The engine room of all modern AI.

Generative AI

AI that creates things — text, images, code, audio — rather than just sorting or scoring them. ChatGPT writing your email is generative AI; your bank flagging a dodgy transaction is the older kind.

LLM (Large Language Model)

The technology behind ChatGPT and its rivals: a program trained on vast amounts of text to predict what words come next, which in practice lets it draft, summarise and converse. Our AI basics guide explains how.

Model

One specific trained AI — the thing that actually does the work. Providers offer several models at different speed/capability/price points, like engine options in the same car.

ChatGPT, Claude, Gemini, Copilot

Brand names, not technologies — products from OpenAI, Anthropic, Google and Microsoft respectively, all built on LLMs. Hoover vs Dyson, not hoover vs washing machine.

02Using the tools

Prompt

Whatever you type into an AI tool. The single biggest factor in the quality of what comes back — think of it as a briefing, not a search query.

Prompt engineering

The (overdressed) name for writing good prompts: giving the AI a role, audience, tone and constraints. A skill learned in hours, not years.

System prompt

Standing instructions a tool follows before it reads anything you type — "you are a helpful assistant for a plumbing firm; never promise prices." How custom AI assistants get their personality and rules.

Context window

The AI's working memory for your conversation. Everything you've said has to fit in it; very long chats can push early material out, which is why the AI sometimes "forgets".

Token

The chunks of text (roughly word-fragments) an AI actually reads and writes — about 750 words per 1,000 tokens. Mostly invisible to you, but it's the unit AI usage is measured and billed in.

Hallucination

When an AI states something false with complete confidence — an invented statistic, a non-existent court case. Not a glitch but a built-in tendency, and the reason a human checks anything that matters.

Training / training data

The process (and the mountain of text) used to build a model in the first place. Important for you because some free tools use what you type as future training data — see the data section below.

Knowledge cut-off

The date a model's training data ends. It knows nothing after that date unless it searches the web — so ask for sources on anything recent.

Fine-tuning

Taking an existing model and giving it extra training on specific material so it gets better at one job. Rarely needed for small businesses — good prompts and RAG (below) usually get there far more cheaply.

Multimodal

A model that handles more than text — it can look at images, listen to audio, or read documents you upload. Most of the big tools now are.

03Beyond the chatbox

Agent

An AI given tools and permission to do multi-step tasks — check the calendar, send the email, update the spreadsheet — rather than just answer questions. We've a whole guide: What Is an AI Agent?

Automation / workflow

A chain of automatic steps triggered by an event: "when a form is submitted → add to spreadsheet → send confirmation." Works with or without AI in the chain; the AI adds judgement to the plumbing.

n8n, Make, Zapier

The popular workflow platforms — the plumbing kits used to build automations and agents by connecting your existing apps together.

API

The connector that lets one piece of software talk to another — how your booking system talks to your calendar, and how custom tools talk to AI models. If a quote mentions "API access", it means software talking to software, no human in the chatbox.

Integration

Any connection between two systems so data flows without retyping. "Does it integrate with Xero?" is asking "will these two talk to each other?"

RAG (Retrieval-Augmented Generation)

A technique where the AI looks things up in your documents before answering, instead of relying on its general training. How "chat with your own company handbook" tools work — and the standard fix for hallucinated facts.

Chatbot

Any program you converse with. The word covers everything from the rigid button-menu bots on retail websites to full LLM assistants — so when someone sells you "a chatbot", ask which kind.

04Data & safety

Personal data

Anything that can identify a living person — names, emails, phone numbers, addresses. The stuff GDPR cares about, and the stuff that shouldn't go into free AI tools. Full picture: Is AI Safe to Use With My Business Data?

GDPR / UK GDPR

The data protection law. It says nothing about AI specifically — but if personal data goes into an AI tool, the law applies exactly as it would for any other software.

Data processing agreement (DPA)

The contract in which a provider commits to handling your data properly on your behalf. Business-tier AI tools include one; free consumer tiers generally don't — which is most of the safety difference.

Training opt-out

A setting (or contractual promise) that stops a provider using your conversations to train future models. Standard in business tiers; worth checking in free ones.

Shadow AI

Staff using AI tools privately, without telling anyone — usually on personal phones. You carry the data risk without the open benefit; the fix is permission with rules, covered in our adoption guide.

Anonymisation

Stripping identifying details out before text goes into a tool ("draft a payment reminder" needs no customer name). The cheapest data protection there is.

Human in the loop

A design where a person approves the AI's work before it takes effect — the email queued for sign-off rather than sent. The hallmark of automation built by grown-ups.

Guardrails

The limits built around an AI system: what it may access, what it may do without approval, what it must refuse. The agent equivalent of "no signing authority".

05Buying & running

Free / Plus / Team / Enterprise tiers

The typical product ladder for AI tools. The jump that matters for business isn't features — it's the data handling: Team and above is usually where the no-training commitments and DPAs live.

Open-source model

A model whose files are published for anyone to download and run — meaning it can run on hardware you control, with data never leaving the building.

Self-hosted / local AI

Running a model on your own computer or server instead of a provider's cloud. More effort, total data control — the option for the especially cautious.

Cloud

Someone else's computers, rented over the internet. Where almost all AI actually runs, because the hardware is expensive.

GPU

The specialist chip AI runs on (originally built for video games). Mentioned here only so the headlines about chip shortages make sense — you'll likely never buy one.

SaaS (Software as a Service)

Software you rent monthly and use in a browser rather than buy outright — the model nearly every AI tool follows.