AI Jargon Explained: 24 Terms for Business Leaders

Nobody should have to learn a vocabulary to buy business software. But the AI market has produced an unusual amount of it, some genuinely useful and some functioning mainly to make simple things sound expensive.

This is the list you need to hold your own in a supplier meeting. Twenty-four terms, what each actually means, and whether it matters to a business your size.

The fundamentals

AI (artificial intelligence). An umbrella term covering everything from the spam filter in your inbox to the thing that writes your emails. When a supplier says "AI powered", the term on its own tells you nothing. Ask what it does.

Large language model (LLM). The engine behind ChatGPT, Claude, Gemini and Copilot. It works by predicting what text should come next, over and over, extremely well. Explained properly in What is a large language model?

Generative AI. AI that produces new material, text, images, audio, code, rather than sorting or scoring existing material. The distinction matters because the older sort, which classifies and predicts, has been running quietly in business software for twenty years.

Model. The engine. GPT, Claude, Gemini and Llama are families of them. Products are built around models, and a product can change which model it uses without changing its name. This is why "we use GPT" ages badly as a specification.

Machine learning. The broader field: software that improves at a task by processing examples rather than by being explicitly programmed. Large language models are one branch of it.

Talking to the thing

Prompt. Everything you type. The instruction and the background together. The most important word on this list, because prompt quality explains most of the difference between people who find AI useful and people who do not.

Context. The background you supply so the answer applies to your business rather than to businesses in general. Who you are, who the reader is, what you sell, what good looks like. Missing context is the commonest cause of disappointing output.

Context window. How much text the model can hold in view at once, measured in tokens. Large in current models, and still finite. In a very long conversation the earliest parts fall out of view, which is why a chat that started well sometimes drifts.

Token. A chunk of text, roughly three quarters of a word. Usage and pricing are counted in these. Mostly relevant if you are paying per use through an API rather than a monthly subscription.

System prompt. Standing instructions applied to every conversation, set once behind the scenes. "You are the assistant for a plumbing firm in Mold. Never quote a price. Always offer a callback." Most of what suppliers describe as a custom AI is a good system prompt, and there is nothing wrong with that.

Temperature. A setting controlling how predictable the output is. Low for anything factual, higher for brainstorming. You will rarely see it outside developer tools.

Where it goes wrong

Hallucination. A confident, fluent, wrong answer. Not a malfunction but a consequence of how the technology works, and the single most important risk to manage. Covered in full in AI hallucinations: how to check what AI tells you.

Knowledge cut-off. The date after which the model has read nothing. Ask it about something more recent and it will either say so or, worse, answer from before the cut-off without mentioning it. Tools with web search largely work around this.

Bias. Patterns in the training data reappearing in the output. Practically, this matters most anywhere AI touches recruitment, lending, pricing by customer group, or anything else where treating people differently has legal consequences. Keep a human decision in those loops, and be able to explain the decision without referring to the tool.

Guardrails. Restrictions on what a system will do or say. Worth asking about specifically for anything customer-facing: what stops it quoting a price, promising a delivery date, or discussing a competitor?

The words on the invoice

Fine-tuning. Further training of an existing model on your own examples so it adopts a particular style or format. Genuine, occasionally useful, and far less often necessary than it is sold. A good system prompt achieves the same thing for most small businesses at a fraction of the cost.

RAG (retrieval-augmented generation). The model looks things up in your documents before answering, so it quotes your actual returns policy rather than inventing a plausible one. This is what most businesses actually want when they say they want a custom AI, and it is the sensible thing to buy.

Embeddings and vector database. The plumbing that makes the lookup above work: a way of storing documents so the system can find the passages that are relevant in meaning rather than by keyword. You do not need to understand this. You do need to know it is a normal component and not a reason for a large number on a quotation.

API. A way for two pieces of software to talk to each other without a person in the middle. Relevant when you want AI inside your own systems rather than in a browser tab. Billed per use rather than per seat.

Agent. A system that takes actions rather than just answering: sending emails, updating records, booking things. Currently applied to almost anything, so ignore the label and ask what it is allowed to do without a person approving it.

Workflow automation. Joining tools together so something happens automatically. Often more valuable than AI itself, and frequently sold alongside it. The AI is usually one step in a chain that is otherwise ordinary plumbing.

Multimodal. Handles more than text: images, audio, documents, sometimes video. Practically, it means you can photograph a handwritten note or an invoice and get something useful back.

The words your accountant and your clients will ask about

Data processing agreement (DPA). The contract governing what a supplier may do with personal data you send them. Business tiers of the major AI tools provide one; free consumer accounts do not. If you take one term from this article into a supplier meeting, take this one. See Is it safe to put company data into ChatGPT?

Training opt-out. A setting, or contractual term, preventing your conversations being used to improve the provider's models. Off by default on some consumer tiers. Check it, and record that you checked.

Data residency. Which country your data is stored and processed in. Not automatically a problem if it leaves the UK, provided the right safeguards are in place. Ask, and keep the answer with your records.

Human in the loop. A person reviews or approves before anything happens. This phrase should appear in any proposal involving customers, staff or money, and if it does not, ask why.

Four questions worth more than the whole glossary

You can be sold to fluently in all twenty-four of the above and still not know what you are buying. These four cut through it:

1. What will it do, in a sentence, without the word AI in it? If that sentence cannot be produced, there is nothing there.
2. Where does our data go, and is there a DPA? A supplier who has to check is a supplier who has not thought about it.
3. Who checks the output, and what happens when it is wrong? The answer should name a person or a step, not a percentage.
4. What breaks when the underlying model is replaced next year? It will be. Anything built on one specific model needs an answer.

Those four have not changed in three years and will still work when half the terms above have been replaced by new ones.

If you would rather learn this by watching it used than by reading a list, the free AI Breakfast Club webinar runs online every other Friday morning, and the open floor afterwards is a reasonable place to bring a supplier's proposal and ask what it actually says.