AI has its own vocabulary, and most of it is explained either in three words or in a research paper — nothing in between. This glossary is the in-between: every term a business owner actually runs into, defined in two plain sentences, with no assumption that you’ve seen the inside of a data center. Skim it, search it, or jump to a section.
Foundations 8 terms
Artificial Intelligence AI
Software that does work we used to think required human judgment — writing, summarizing, answering questions. In business today it almost always means the newer “generative” kind that produces text, images, or code on demand.
Machine Learning ML
A way of building software by showing it thousands of examples instead of writing step-by-step rules. The system learns the patterns itself, which is why modern AI can handle messy, real-world language.
Large Language Model LLM
The engine behind tools like ChatGPT and Claude — software trained on enormous amounts of text to produce human-like writing. It’s what lets you ask a question in plain English and get a plain-English answer.
Generative AI Gen AI
AI that creates new content — text, images, audio, or code — rather than just sorting or labeling existing data. This is the wave that put AI on every business owner’s desk.
Model the engine
The actual trained “brain” you run, such as GPT-4 or Llama. Models differ in cost, speed, capability, and whether you can run them on hardware you own.
Open-weight model downloadable
A model whose internal files are published, so anyone can download and run it on their own computers. This is what makes private, on-premise AI possible without paying a cloud provider for every use.
Parameters model size
The internal settings a model adjusts during training — often billions of them — that hold everything it has learned. More parameters can mean a smarter model, but also a bigger, costlier one to run.
Training how a model learns
The one-time, expensive process of teaching a model by feeding it huge amounts of data. Almost no business trains a model from scratch; they take one that’s already trained and adapt it with fine-tuning or RAG.
Using AI day-to-day 8 terms
Prompt your input
The instruction or question you give an AI, like “summarize this contract in five bullet points.” What you get back depends heavily on how clearly you ask.
Prompt engineering asking well
The practice of wording and structuring prompts to get reliable, useful results. It’s a real skill, but for most teams a few good reusable templates cover the bulk of daily work.
Token the billing unit
The small chunk of text — roughly three-quarters of a word — that AI reads and charges by. Usage and cost are measured in tokens, so a long document simply costs more to process than a short one.
Context window working memory
How much text a model can keep in mind at once, measured in tokens. A bigger window lets the AI work through longer documents or conversations before it starts “forgetting” the earlier parts.
Inference running the model
The act of actually using a trained model to get an answer — every question you ask is one inference. In the cloud you pay per inference; run it locally and inference is essentially free once you own the hardware.
System prompt the standing brief
A behind-the-scenes instruction that tells the AI how to behave — its role, tone, and rules — before any user types a word. It’s how a business gives a chatbot a consistent personality and guardrails.
Multimodal beyond text
A model that handles more than typed text — it can also read images, documents, audio, or charts. Useful when your work involves scanned paperwork, photos, or screenshots, not just words.
Temperature creativity dial
A setting that controls how predictable or inventive the AI’s output is. Low temperature gives consistent, by-the-book answers; high temperature gives more varied, creative ones.
Making AI yours 8 terms
Fine-tuning teaching your style
Taking an existing model and training it a little further on your own examples so it adopts your style, terms, or task. It’s powerful but often overkill — many businesses get what they need from RAG instead.
RAG Retrieval-Augmented Generation
A setup that lets the AI look up answers in your own documents before responding, instead of relying only on what it memorized in training. This is how you get a chatbot that actually knows your products, policies, and prices — and it sharply reduces hallucination.
Embedding meaning as numbers
A way of turning text into numbers that capture its meaning, so software can find related content even when the wording differs. It’s the quiet machinery that makes document search and RAG work.
Vector database the smart filing cabinet
A specialized store for embeddings that can instantly find the most relevant passages across thousands of your documents. It’s what a RAG system searches before it answers.
Agent AI that acts
An AI that doesn’t just answer but takes steps — looking things up, using tools, and chaining actions to finish a task. Think of it as moving from “advisor” to “assistant that actually does the work.”
Tool use function calling
The ability for an AI to trigger real software actions — check inventory, book a meeting, send an email — instead of only producing text. It’s what connects an agent to the systems your business already runs.
Knowledge base your source of truth
The curated set of your own documents — policies, manuals, FAQs — that you feed an AI so its answers reflect your business, not the open internet. Keeping it clean and current is most of the work in a useful RAG deployment.
MCP Model Context Protocol
An emerging standard that lets AI tools plug into your data and software in a consistent way. It matters because it cuts the custom wiring needed to connect AI to the systems you already use.
Where AI runs 7 terms
Cloud AI you rent it
AI that runs on a provider’s servers and reaches you over the internet, paid monthly or per use. It’s the fastest way to start, but your data is processed on someone else’s systems. See the full local-vs-cloud guide.
Local AI on-premise / you own it
The same kind of AI running on a computer you own, so nothing you type leaves your building. No per-use fees and it works offline, in exchange for buying and maintaining the hardware. Made possible by open-weight models.
Hybrid AI split the work
A split where sensitive work runs on hardware you own and everything else uses the cloud. Most businesses that get serious about AI end up here.
API the software connection
A connection that lets your own software talk to an AI service automatically, instead of a person typing into a chat box. It’s how AI gets built into your website, tools, and workflows behind the scenes.
GPU the AI chip
The specialized chip that does the heavy math AI relies on, whether in the cloud or a local workstation. For on-premise AI, the GPU is usually the biggest hardware decision and cost.
Quantization shrinking the model
A technique that compresses a model so it runs on smaller, cheaper hardware, trading a little accuracy for a lot of savings. It’s often what makes capable local AI affordable for a small business.
Self-hosting infrastructure you control
Running AI software on infrastructure you control rather than a vendor’s platform. It gives you privacy and predictable costs, but the upkeep is yours.
Risks & limits 5 terms
Hallucination confident but wrong
When an AI states something false with total confidence, because it predicts plausible words rather than checking facts. It’s the single biggest reason to keep a human reviewing anything that matters — and the main problem RAG is built to reduce.
Bias skewed output
Unfair or skewed output that reflects patterns in the data a model learned from. In business it surfaces in things like hiring, lending, or customer messaging, where it carries real legal and reputational risk.
Guardrails the limits you set
The rules and limits placed around an AI to keep it on-topic, safe, and on-brand. For a customer-facing chatbot, good guardrails are what stop it from making promises you can’t keep.
Prompt injection the security risk
A trick where hidden instructions buried in a document or web page hijack an AI into doing something it shouldn’t. It becomes a genuine concern once an agent can read outside content or take actions.
Still have a question?
A glossary can define the words, but it can’t tell you which of these apply to your business. If you want a human read on where AI fits — or where it doesn’t — get in touch, or ask GPT_Lenny (the chat bubble in the corner) to talk it through.