The AI Glossary for Small Business

Agent

An AI that takes actions, not just describes them. Where a chatbot tells you what to do, an agent goes and does it: drafts the email, books the call, files the report. Agents are early. They get confused in long workflows. They sometimes do exactly what you said instead of what you meant. See our primer on agents and MCPs.

AGI Artificial General Intelligence

A hypothetical AI that's as capable as a human across all tasks, not just specific ones. Frontier labs are aiming for it. Whether and when it'll arrive is one of the loudest arguments in tech. For a small business today, it's mostly a marketing word.

API Application Programming Interface

A way for one piece of software to talk to another. When your invoicing tool pulls data from Stripe, that's an API. Every major AI tool has an API, which is how developers wire AI into other systems without using the chat window.

Bias

When an AI's outputs carry the patterns and prejudices of the data it was trained on. Important in hiring, customer service, lending, and anything else where consistent fair treatment matters. Bias in AI is the same as bias in the people who wrote the text it learned from.

ChatGPT

OpenAI's flagship chatbot, and probably the thing most people mean today when they say "AI." Free version uses an older model; paid version (Plus, around $20 a month) gives you the current frontier model plus image generation, voice mode, and Custom GPTs.

Claude

Anthropic's chatbot. The one we use most internally. Better than the alternatives at long documents, more honest about what it doesn't know, less prone to hallucinating. Made by Anthropic, which spun out of OpenAI in 2021.

Context

The background information an AI needs to be useful to you. Who you are, who you're talking to, what's already been said, what you're trying to do. Without context, you get generic-sounding answers. With it, you get something you'd actually send. See our primer on context.

Context window

How much information an AI can hold in mind at once. Measured in tokens. A large context window (Claude's is over 200,000 tokens, ChatGPT's around 128,000) means you can paste in a long contract or a year of meeting notes and the AI can reason across all of it.

The Core

Our name for one governed source of shared business context: AI can support work across the organization while people keep the consequential decisions. It is built from the Foundation and AI System around the organization using it. See The Core.

Custom GPT

A version of ChatGPT you've configured with specific instructions, files, and context for a particular use. Lets you save a "version" of ChatGPT trained on, say, your customer voice or your internal SOPs. Available on ChatGPT Plus and above.

Embedding

A way of turning words, sentences, or documents into numbers an AI can compare. The reason "search your files with AI" works at all: your files are embedded into vectors that can be searched semantically, not just by exact keyword.

Foundation

Our word for your business written down in structured files an AI can read. Voice, audience, processes, products, common tasks. It's both the engagement where we capture it and the file you walk away with. Yours to keep, portable, no lock-in. Every later AI build grows out of it.

Fine-tuning

Training an existing AI model on extra examples specific to your use case. Used to be the main way to customize AI for a business. Today it's mostly replaced by giving the model good context at runtime, which is faster, cheaper, and easier to update.

Frontier model

The best AI models available at any given moment. As of mid-2026 that includes GPT-5, Claude Opus 4.x, and Gemini Ultra 2. Frontier models change every few months as labs release new versions. The phrase exists mostly so the labs have a flattering way to describe their flagship.

Gemini

Google's AI tool. Lives inside Google Workspace: Gmail, Docs, Drive, Sheets. Comes free with most paid Workspace plans. Strongest for teams already deep in Google's ecosystem; weaker as a standalone chatbot.

Generative AI

AI that produces new content (text, images, audio, video) instead of just analyzing existing content. Generative AI is what most people mean when they say "AI" today, and what every product in this glossary refers to.

GPT Generative Pre-trained Transformer

The underlying architecture behind ChatGPT (the "GPT" in the name) and a lot of other modern AI. Originally developed at OpenAI. The architecture is now used widely across the industry, including in some models that aren't made by OpenAI.

Guardrails

Rules and constraints built around an AI system to keep it from doing things you don't want. Examples: refusing to discuss competitors, requiring human approval before sending external emails, only answering inside a specific domain. The bigger the agent, the more guardrails matter.

Hallucination

When an AI confidently makes something up. The biggest practical problem with AI tools. Models can invent court cases, fabricate citations, or insist a product feature exists when it doesn't. The best defense is simple: verify anything the AI tells you that you can't check independently. The newer reasoning models hallucinate less, but not zero.

Inference

What's happening when you're actively using an AI tool (asking it questions, getting answers). Distinct from training. Inference is what costs the AI companies money every time someone sends a prompt, and it's what your subscription is paying for.

LLM Large Language Model

The general term for the kind of AI that powers ChatGPT, Claude, Gemini, and most modern AI tools. Trained on enormous amounts of text. Predicts the most likely next word, over and over, very fast. The underlying machine you're chatting with.

MCP Model Context Protocol

An open standard that lets AI tools plug into your other tools (Google Drive, your CRM, your invoicing software, Slack) without each pair needing a custom integration. Think of it as USB for AI. See our primer on agents and MCPs.

Memory

When an AI remembers things between conversations. ChatGPT, Claude, and Gemini all have some version of this now. Useful when set up well; can also produce weird outputs if the AI remembers something incorrectly from a months-old chat.

Model

The actual AI doing the thinking. ChatGPT, Claude, and Gemini are products built around models. The model is the brain; the product is the app you're talking to. Most products let you pick between models (GPT-4o, GPT-5, Claude Sonnet, Claude Opus, etc.).

Multimodal

An AI that can handle multiple kinds of input or output (text, images, audio, video). All three major chatbots are multimodal today. You can paste an image into Claude and ask questions about it; you can talk to ChatGPT.

Open source vs. closed source

Open-source models (Llama, Mistral, DeepSeek, Qwen) have weights you can download and run yourself. Closed-source models (GPT-5, Claude, Gemini) only run on the company's servers. Open-source is mostly relevant if you're building product, or have strict privacy requirements, or want to avoid vendor lock-in.

Parameters

The internal numbers a model uses to make predictions. Often used as a rough measure of model size. A frontier model has hundreds of billions of parameters; small open-source models have a few billion. More parameters usually means smarter but slower and more expensive.

Plugin

A small piece of software that extends an AI's abilities. ChatGPT plugins can search the web, run code, generate images, query databases. The plugin model is partly being replaced by MCPs, which do something similar in a more general, cross-vendor way.

Prompt

Whatever you type into the AI. The prompt is the input. "Write me a follow-up email" is a prompt. So is a 12-page document you paste in along with three sentences of instructions.

Prompt engineering

The skill of writing prompts that get useful answers. Real, but overhyped. Mostly comes down to: be specific, give context, say what you want and what you don't want, give an example if you can. Less an art form than the marketing makes it sound.

RAG Retrieval-Augmented Generation

A way of giving an AI access to your specific information without retraining the model. The AI looks up relevant chunks of your data, then uses them to answer. Most "chat with your docs" tools use RAG under the hood.

Reasoning model

A model that thinks step-by-step before answering, often showing its reasoning as it goes. Slower and more expensive than a regular model, but much stronger on math, code, multi-step logic, and ambiguous problems. ChatGPT's "Thinking" mode and Claude's "Extended Thinking" are examples.

System prompt

Hidden instructions that shape how an AI behaves before you ever type anything. Your ChatGPT's tone, what it knows about you, how cautious it is, all set by the system prompt. You can write your own in Custom GPTs or via API.

Temperature

A setting that controls how random an AI's output is. Low temperature gives you predictable, repeatable answers. High temperature gives you more creative but less reliable output. Most chat interfaces hide this setting; developers use it directly via API.

Token

A chunk of text, roughly three-quarters of a word. AI bills, context windows, and rate limits are all measured in tokens. "Hello, how are you?" is about 6 tokens. A 50-page document is around 25,000 tokens.

Training data

The text, images, and other content the AI learned from before it ever met you. For LLMs, this is usually a huge chunk of the public internet plus licensed datasets. What's in the training data shapes what the model knows and how it talks.

Vector database

A specialized database that stores embeddings (vectors). The thing that makes "find me documents semantically similar to this one" fast. Used inside most RAG systems and inside the search underneath a lot of AI features.

Vision

An AI's ability to see and interpret images. All three major chatbots can read photos, screenshots, charts, invoices, and whiteboards now. Useful for "what's in this receipt," "summarize this slide deck," or "tell me what changed in this contract redline."

Voice mode

Talking to an AI like you'd talk to a person. ChatGPT's voice mode is currently the strongest; Claude and Gemini also have it. Useful in cars, kitchens, walks. Less useful at the office, in our experience.

Workflow automation

Stringing AI together with other tools so something useful happens without you manually doing it. This used to be Zapier and n8n territory. Increasingly it's the job of AI agents, which can figure out the steps instead of needing you to write rules.

Zero-shot / few-shot

How an AI handles a task with no examples (zero-shot) versus with a few examples (few-shot) given in the prompt. Most modern AI tools handle a lot of things zero-shot. Few-shot prompting still helps when you need the output in a specific format the AI hasn't seen before.