AI glossary for small business: plain-English definitions of the terms you keep hearing

Most AI jargon describes a few simple ideas. An AI tool like ChatGPT, Claude, Copilot, or Gemini runs on a large language model, which writes text by predicting what should come next. You give it instructions, called a prompt, and it can make things up, called a hallucination, so a person checks the output before it goes anywhere. A business plan keeps your data under company control, and connectors let the tool read your email and files. Agents and automations let it take steps on its own. The definitions below cover 58 terms in plain English, grouped by topic, with a small-business example where one helps.

Jump to a term

The basics

Artificial intelligence (AI)

Software that does tasks we used to assume needed a person, like reading a document, answering a question, or writing a reply. When people say "AI" today, they usually mean generative AI tools like ChatGPT.

Machine learning

The way most AI is built: instead of programming every rule by hand, developers show the software a large number of examples and it learns the patterns. Spam filters and fraud alerts are older, familiar uses.

Generative AI

AI that creates new content, such as text, images, audio, or code, instead of only sorting or scoring what already exists.

Example: drafting a reply to a customer's question about a delayed order.

Large language model (LLM)

The engine behind tools like ChatGPT and Claude. It is trained on huge amounts of text and writes by predicting, one piece at a time, what text should come next. That is why it can sound fluent and still be wrong (see hallucination).

Model

A specific version of an AI engine, usually with a name and number, such as a new GPT, Claude, or Gemini release. Vendors release new models often, and the tool you pay for usually switches to them for you. Good setups don't depend on any one model.

Chatbot

A program you talk to by typing or speaking. Older chatbots followed scripts. AI chatbots, like ChatGPT, write their answers on the spot. A customer-facing chatbot on your website is a bigger project than an internal one, because it speaks for your business without anyone checking each reply.

AI assistant

A general-purpose AI tool for work, such as ChatGPT, Claude, Copilot, or Gemini. People use "assistant" and "chatbot" almost interchangeably. "Assistant" usually suggests it can also work with your files, email, and calendar.

Prompt

The instructions you give an AI tool. A good prompt says what you want, who it is for, what to use as source material, and what the result should look like.

Example: "Summarize this vendor contract in five bullet points for our office manager, and list every date and dollar amount."

Prompt engineering

The skill of writing prompts that get reliable results. For most office work, it comes down to being specific and giving the tool the right material. A saved set of prompts that work is sometimes called a prompt library.

Context window

How much text an AI tool can consider at once, including your prompt, any files you attach, and the conversation so far. It is measured in tokens. Very long documents or very long chats can push early details out of view, which is one reason answers drift in long conversations.

Token

The unit AI models use to count text. A token is a word or a piece of a word; in English, a token averages a bit less than one word. Usage limits and developer pricing are usually counted in tokens.

Hallucination

When an AI tool states something false with confidence, such as a made-up policy number, a wrong date, or a quote nobody said. It happens because the model predicts likely text and does not check facts. The fix is to give the tool the source material (grounding) and have a person check anything that matters (human in the loop).

Example: a draft bid that cites a warranty term your company doesn't offer.

Multimodal

An AI tool that handles more than text, such as photos, PDFs, scanned forms, audio, or screenshots.

Example: photographing a handwritten job sheet and asking the tool to type it up.

Reasoning model

A model that works through a problem in steps before answering. It is slower, and usually better at math, analysis, and multi-step questions. Many tools let you choose between a fast mode and a "thinking" mode.

Tools and plans

ChatGPT

OpenAI's AI assistant and the best known one. Its plan for companies is ChatGPT Business, which was called ChatGPT Team until August 2025. We compare it with the other main options in ChatGPT Business vs Microsoft Copilot vs Claude Team.

Claude

Anthropic's AI assistant. Many teams use it for long documents, careful writing, and analysis. Its plan for small companies is called Claude Team.

Microsoft Copilot

Microsoft's AI assistant, built into Word, Excel, PowerPoint, Outlook, and Teams. It was called Microsoft 365 Copilot until 2026, so you will see both names. It shows each person only what they already have permission to open, which makes a permissions review worth doing before you turn it on.

Copilot Chat

A version of Microsoft's assistant that comes with many Microsoft 365 business plans at no extra cost. It is a reasonable way to get started. The paid Copilot license is what puts it inside the apps and connects it to your work content.

Gemini

Google's AI assistant. Google Workspace business plans include the Gemini app, and features vary by plan. In Workspace, it can also help inside Gmail, Docs, Sheets, and Meet.

Business account vs personal account

A personal account belongs to the employee. A business account belongs to the company, which controls who has access, what data settings apply, and what happens when someone leaves. Business plans from the major vendors also say they don't train their models on your business content by default (see training data).

Example: a bookkeeper pasting client statements into a free personal account means you lose track of that data when they leave.

Seat (license)

One paid user on a business plan. Most AI business plans are priced per seat per month. You can usually add seats as more people need them.

Admin console

The settings area of a business plan where an administrator adds and removes users, manages billing, sets data and sharing rules, and sees usage. Someone at your company should own it, even if your IT provider helps.

Projects, custom GPTs, and Gems

Ways to save instructions and reference files so you don't have to repeat them every time. ChatGPT and Claude call them projects, ChatGPT also has custom GPTs, and Gemini has Gems. Shared versions let a whole team start from the same setup.

Example: a project with your price sheet and standard terms that anyone in the office can use to draft quotes.

Deep research

A mode in several AI assistants where the tool spends several minutes searching the web or your files and comes back with a longer, sourced report. Names vary by tool. It still needs checking, but it can save hours of background research.

AI search and AI answers

Search results written by AI instead of a list of links, as in ChatGPT search, Google's AI answers, or Perplexity. For a small business, being cited in these answers is a new way to be found, alongside ordinary search rankings.

How AI works with your data

Training data

The text, images, and other material a model learned from before you ever used it. When vendors say a business plan "doesn't train on your data," they mean what you type in is not used to teach future models. Your data can still be stored for a time (see data retention).

Grounding

Giving the AI the actual source material and telling it to answer only from that, instead of from memory. Grounded answers are far less likely to be made up, and they can point to where each answer came from.

Example: attaching your employee handbook and asking how many PTO days a new hire gets.

Retrieval-augmented generation (RAG)

The technical name for one common way to ground answers. When someone asks a question, the system first searches your documents for the relevant passages, then hands those passages to the model to write the answer. Most "chat with your documents" features work this way.

Connector

A built-in link between an AI tool and another app, such as Gmail, Outlook, Google Drive, SharePoint, or Slack, so the tool can read from it (and sometimes act in it) without copy and paste. An admin usually decides which connectors are allowed.

Model Context Protocol (MCP)

An open standard, introduced by Anthropic in 2024 and now supported by many AI tools, that lets AI assistants connect to other software in a common way. You will mostly see it as "MCP server" in a list of available connections. Treat each one like any other app you give access to your data.

Knowledge base

An organized collection of your company's documents and answers, such as policies, procedures, price sheets, and FAQs, that people and AI tools can look things up in. An AI tool is only as good as the knowledge base it reads from, so outdated or conflicting files lead to wrong answers.

Data retention

How long a vendor keeps your chats, files, and outputs, and whether you can delete them. Business plans usually give admins more control here than personal plans. Check the vendor's settings and terms if you have legal or client requirements about how long data is kept.

Sensitive data

Information that would cause harm if it leaked, like Social Security numbers, bank details, health information, passwords, and client confidential files. Every AI policy should say plainly what never goes into an AI tool, even a business one.

Single sign-on (SSO)

Logging in to the AI tool with your existing work account, such as Microsoft or Google, instead of a separate password. It means one place to turn off access when someone leaves. Your IT provider usually sets it up.

Permissions

The rules for who can see which files, folders, and systems. AI tools that read your company data generally respect the permissions you already have, so if a salary spreadsheet is shared with "everyone," an AI tool may surface it to anyone who asks.

Shadow AI

Staff using AI tools the company hasn't approved, usually on personal accounts. It is common, and it usually means people want the tools. The fix is to give them an approved business account and a few clear rules, not a ban.

Automation and agents

Workflow automation

Software that moves information between apps and does routine steps without a person clicking through each one. Adding AI lets an automation handle messy inputs, like reading an email and deciding what kind of request it is.

Example: a new web form submission creates a contact in your CRM and posts a note in Teams.

Trigger and action

The two halves of most automations. The trigger is the event that starts it, like a new email or a new row in a spreadsheet. The action is what happens next, like creating a record or sending a notification.

Automation platforms (Zapier, Make, Power Automate)

Tools that let non-programmers build workflow automations by connecting apps with triggers and actions. Zapier and Make work across thousands of apps. Power Automate is Microsoft's and fits naturally in a Microsoft 365 company.

API

Short for application programming interface: the way one piece of software talks to another. Developers use AI vendors' APIs to build AI into their own systems. If a tool "has an API," it can usually be connected to other software.

Integration

Any connection between two systems so data flows between them, whether through a connector, an automation platform, or a custom build on an API.

AI agent

An AI tool that can take a series of steps on its own toward a goal, such as searching, filling in a form, updating a spreadsheet, or sending a draft, instead of only answering one question. Agents are improving fast. They also make mistakes faster, so start them on low-risk work with a review step.

Example: an agent that reads incoming invoices, matches them to purchase orders, and lists the mismatches for a person to review.

Agentic AI

The general term for AI that acts, not just answers. In practice, it means the same thing as using AI agents.

Human in the loop

A design where a person reviews or approves the AI's work before it counts, for example before an email is sent, a record is changed, or a payment goes out. For most small businesses, this should be the default.

Document extraction

Pulling specific details out of documents, like invoice totals, dates, policy numbers, or line items, and putting them into a spreadsheet or system. Older tools did this with OCR (optical character recognition); AI tools also understand layout and meaning. It is a back-office job AI handles well today.

Meeting transcription and notes

AI that records a meeting or call, writes a transcript, and produces a summary with action items. Tell people when a meeting is being recorded, and check your state's consent rules for recording calls.

Adoption and results

AI adoption

Getting people in a company to use AI on real work, regularly, in a way that actually saves time. Buying the software is a small part of it. We cover how to begin in How to start using AI in a small business without an IT team.

Use case

A specific task AI is used for. "Draft replies to certificate-of-insurance requests" is a use case. "Use AI for marketing" is a wish. For examples that fit an office, see What can AI actually do for a small business back office?

Pilot

A small, time-boxed test of an AI tool with a few people and a few tasks before buying it for everyone. A good pilot has a baseline and a decision date.

Baseline

How long a task takes, or how often it goes wrong, before AI is involved. Without one, you can't tell later whether anything improved.

Time saved

The difference between the baseline and the time the task takes with AI, including the time spent checking the AI's work. Leaving out checking time is the most common way savings get overstated.

ROI (return on investment)

What you got back compared with what you spent. For AI, the spend is seats, setup, and staff time; the return is usually hours saved, fewer errors, or work you can take on without hiring. It only means something when it is measured against a baseline.

AI policy

Your company's rules for using AI: which tools and accounts are approved, what must never be entered, and what a person must check before AI output is used. For a small business, it can start as a single page. Some clients and insurers may ask whether you have one.

Acceptable use

The part of an AI policy that says what staff may and may not do with AI tools. It is often the first section people read, so keep it short and concrete.

AI governance

The ongoing oversight of AI in a company: who approves new tools, who owns the policy, and how use and risk are reviewed. In a 30-person company, this is usually one owner and a regular check, not a committee.

Change management

The people side of any new tool: explaining why, involving staff, training them, and following up so the change sticks. When a rollout stalls, it is more often here than on the technology.

Prompt library

A shared, saved collection of prompts that work well for your company's common tasks, so people don't start from a blank box every time.

AI operations audit

A structured review of where your team's time actually goes, which tasks AI can reliably help with, and what each change is likely to be worth, ending in a ranked plan. We explain what one involves in What is an AI operations audit?

Questions owners ask

Do I need to understand all of these terms to start using AI?

No. To get started, you need about five: prompt, hallucination, business account, permissions, and human in the loop. The rest become useful as you go further.

What is the difference between AI, generative AI, and an LLM?

AI is the broad field. Generative AI is the kind that creates content. An LLM is the type of generative AI behind text tools like ChatGPT, Claude, Copilot, and Gemini.

Is an AI agent the same as a chatbot?

No. A chatbot answers when you ask. An agent can take a series of actions on its own, such as looking things up, updating records, or preparing drafts, which is why agents need tighter limits and a review step.

Is it safe to put company information into AI tools?

On a business plan with sensible settings, routine work information is generally fine; that is what those plans are built for. Keep sensitive data out unless your plan and your industry's rules clearly allow it, and put that rule in writing.

Still not sure what applies to you?

Armature helps small and midsize businesses in Washington figure out where AI fits in their work and put it to use without the jargon. Email [email protected] for a free 30-minute call.