AI terms in plain English, for a business of 5 to 50 people.

Agent

Software that takes actions, not just answers questions. It can look up a customer, book an appointment, draft an invoice. The question to ask is which actions it can take on its own and which wait for a person.

API

The door one piece of software opens so another can use it. Your phone system, QuickBooks and your scheduler each have one, or do not. Whether they do decides how hard the wiring is, and therefore what it costs.

Approval step (human in the loop)

A person checks before anything goes out. For invoices, quotes and replies to customers this is the default, not an add-on. A system that skips it is asking you to trust a machine with your name on the letter.

Audit trail

A record of what the system did, when, and why, kept in the same place your staff's actions are kept. If a vendor cannot show you one, the system is not ready to touch your books.

Automation

Rules that move work along without a person: when a job is marked done, draft the invoice. Most of the value in "AI" for a small business is plain automation, with AI handling the messy parts like reading an email that a rule cannot.

Chatbot

A chat window that answers questions. The least useful shape of AI for most businesses, because the work is not in answering, it is in doing. Useful on a website for the questions your front desk is tired of, and not much else.

Fine-tuning

Training a model further on your own examples. You almost never need it. Grounding, below, gives the model your information at a fraction of the cost and without the model forgetting it next quarter.

Grounding (retrieval, RAG)

Giving the model your documents to answer from, so it uses your price list and your procedures rather than the internet's. Vendors call this RAG. It is how an assistant "knows your business", and it is the part to test with a question only your business could answer.

Hallucination

When the model states something untrue, fluently and with confidence. It is why the approval step exists, and why nothing important leaves the building unread.

Integration

Connecting two systems so information moves between them without retyping. The word covers most of what an implementer actually does all day, and most of what you are paying for.

Intake

Whatever arrives: a call, a web form, an email, a text. Intake automation reads it, pulls out what matters, and puts it where it belongs, with the urgent ones on top.

Knowledge assistant

An assistant grounded in your own files and procedures, so a new hire can ask how something is done here and get the answer the twenty year employee would give. The most common first project in small companies, for good reason.

Local model

A model that runs on your own machine or server, so data never leaves the building. Slower and less capable than the big hosted ones. Sometimes required by a contract or a regulator, and worth it then.

MCP (Model Context Protocol)

A standard way for an AI to use tools and systems, so one integration works across models instead of being rebuilt for each. Ask a vendor whether they build to it. The answer tells you how locked in you will be.

Model (LLM)

The thing that reads and writes text. Claude, GPT and Gemini are models. Which one you use matters less than what it is wired to and who checks its work.

Prompt

The instructions given to a model. In a working system your staff never write prompts. The system does, the same way every time, and that is the difference between a tool and a toy.

Prototype

A version that works on your data for one case, built to find out whether the real thing is worth building. It should exist before you commit real money, which is why the Opportunity Map ends with one.

System of record

The place the truth lives: QuickBooks for money, your CRM for customers, your scheduler for jobs. AI should work inside it, through the same rules your staff follow, not beside it in a spreadsheet nobody reconciles.

Token

The unit models are metered in, roughly three quarters of a word. It is why AI usage for a small business costs $20 to $150 a month rather than thousands, and why the engineer's time is the expensive part.

Training on your data

A vendor using your information to improve their model for everyone. The answer you want is no, in writing, before anything goes live. Grounding does not require it.

Trigger

The event that starts an automation: a call ends, a form is submitted, a job is closed. Every automation has exactly one, and naming it is the first design decision.

Voice AI (AI receptionist)

A system that answers the phone, talks in real time, and can take action such as booking or looking up an account. Judge it on what it does when it does not know the answer, because that is the call that matters.

Workflow

The path a piece of work follows through your business, from arrival to done, including the retyping and the waiting. Mapping it is the first step. Automating it is the second. Doing them in the other order is how money gets wasted.

Missing a term a vendor used on you? Send it and I will add it. For what these things cost, see what AI automation costs a small business. For who to hire, see how to choose an AI consultant in Portland.