The plain-English AI glossary for business owners

Every industry has jargon. AI has more than most, and it gets used on you in sales pitches. Here is what the words actually mean, grouped by when you'll hear them. Bookmark this page. We link to it from everything else we write.

The basics

AI model. The trained system that produces answers. Think of it as an employee who has read most of the internet but knows nothing about your business until you tell them. Why it matters: when a vendor says "our AI," ask whose model is underneath. It's usually one of a few big ones.

LLM (large language model). The kind of AI model that reads and writes text. ChatGPT, Claude, and Gemini are all built on LLMs. Why it matters: "proprietary LLM" in a small vendor's pitch usually means they rent one of the big ones. That's fine. Claiming otherwise isn't.

Prompt. The instruction you give an AI. Nothing more mystical than that. Why it matters: anyone selling "secret prompts" is selling paper.

Token. The unit AI companies use to measure text, roughly three quarters of a word. Usage-based pricing is priced per token. Why it matters: "unlimited" offers usually have token limits in the fine print.

Context window. How much text a model can consider at once, like a desk that only fits so many papers. Why it matters: if a vendor promises "it reads all your documents," ask how, because the desk has a size limit.

Hallucination. When AI states something false with full confidence. Every model does it. The rate varies, but none are immune. Why it matters: any pitch that doesn't mention how errors get caught is incomplete.

Multimodal. A model that handles more than text, such as images, audio, or video. Why it matters: it's a real capability, not a premium magic trick.

How AI works with your data

Training vs. inference. Training is building the model, which costs millions and is done by big labs. Inference is using the model, which is what you pay for month to month. Why it matters: nobody is "training an AI" on your $200/month budget. They're configuring one.

Fine-tuning. Actually adjusting a model's internal wiring with your data. Expensive, rarely needed for small business. Why it matters: most things sold as fine-tuning are really just documents added to the context. Ask which one you're buying.

RAG (retrieval-augmented generation). The technique of fetching your documents and handing the relevant parts to the model before it answers. This is how most "AI trained on your business" products actually work. Why it matters: it's the honest answer to "how will it know my business."

Embedding. A way of converting text into numbers so a computer can measure how similar two pieces of text are. The plumbing behind search and RAG. Why it matters: it's infrastructure, not a selling point.

Vector database. Where embeddings get stored so they can be searched quickly. Why it matters: same as above. If this word is doing the selling, the pitch is thin.

PII (personally identifiable information). Names, emails, addresses, anything that identifies a person. Why it matters: before any AI project, ask where PII goes and who can see it.

Data retention. How long a provider keeps what you send them. Why it matters: it's in the policy nobody reads. Read it, or ask the vendor to point to it.

Buying and building

API. The plug that lets one piece of software talk to another. AI companies sell access to their models through APIs. Why it matters: "we integrate via API" is table stakes, not a differentiator.

Integration. Connecting tools you already use so information moves without a human retyping it. Why it matters: this is where most small-business AI value actually lives.

Automation vs. AI. Automation follows fixed rules: if this, then that. AI makes judgment calls on messy input. Why it matters: plenty of problems need boring automation, not AI. It's cheaper and more reliable. A good vendor will say so.

Chatbot. Software that converses with customers. Ranges from rigid menu bots to AI assistants that genuinely understand questions. Why it matters: the word covers both. Ask which one is being sold.

Copilot. An AI assistant that helps a person do their work rather than replacing them. Why it matters: this is the realistic model for most small-business AI today.

Agent. Software that uses AI to take multi-step actions on its own, like researching, then drafting, then sending. The industry's most hyped word right now, and its meaning is still shifting. Honest people disagree about what counts as one. Why it matters: when you hear "agent," ask exactly what it does without a human involved and what happens when it gets something wrong.

Open-source model. A model whose weights are published so anyone can run it on their own machines. Why it matters: relevant if you have strict privacy needs. Otherwise mostly a technical detail.

Local / on-device AI. AI that runs on your own hardware instead of a provider's servers. Your data never leaves the building. Why it matters: privacy upside, capability tradeoff. Good vendors explain the tradeoff.

Risk and control

Guardrails. Rules that constrain what an AI system will say or do, like blocking topics or requiring approval before sending. Why it matters: "what guardrails?" is one of the best questions to ask any vendor.

System prompt. The standing instructions a builder gives an AI before customers ever talk to it, defining its role, tone, and limits. Why it matters: much of what you pay a builder for is getting this layer right.

Inference cost. The ongoing cost of running AI, billed per use. Why it matters: a quote that covers building but not running is half a quote.

If you hit a term in a vendor pitch that isn't here, email us. We'll answer, and probably add it to the list.