You sit in a vendor call and the words go past: tokens, context window, RAG, agents, embeddings, MCP. You nod, because asking would take the meeting sideways, and you leave not quite sure what you were offered. Then a proposal arrives full of the same words and you cannot tell whether the price is fair because you cannot tell what the parts are.
This AI glossary for business owners fixes that. Forty-five terms, each in two plain sentences, each with a line on why it matters to someone who runs a company rather than someone who writes code. Read it straight through once, then keep it open during the next call.
What this AI glossary for business covers
Most of these words describe one of four things: the model itself (the software that produces text), the material you feed it, the actions you let it take, or the safeguards around it. A useful analogy: the model is a very well-read temp, the material is what you put on their desk, the actions are the keys you hand them, and the safeguards are the manager who checks their work. Every term below sits somewhere in that picture.
The model and how it works
1. AI (artificial intelligence)
The broad label for software that does tasks that used to require a person: reading, writing, sorting, deciding. In business in 2026 it almost always means a large language model with software around it. Why you care: when a vendor says “AI,” ask which of the four parts they mean.
2. LLM (large language model)
Software trained on an enormous amount of text to predict the most likely next words for any input, which produces fluent, mostly accurate writing and reasoning. It composes; it does not look things up. Why you care: everything it does well and badly follows from that (What Is an LLM? A Plain Explanation for Business Owners).
3. Model
A specific trained LLM, such as one of the current Claude models, a GPT-class model from OpenAI, or Gemini from Google. Why you care: the model choice affects cost, quality, and your data terms.
4. Provider
The company that trains and serves the model: OpenAI, Anthropic, Google, and others. You usually pay them per usage through an API. Why you care: the provider’s terms decide what happens to your data.
5. Open-weight model
A model whose trained weights are published so anyone can run it on their own hardware. Why you care: it is the route to keeping data entirely in-house, at the cost of running your own servers.
6. Training
Showing the model huge amounts of text so it learns to predict language. It happens once, at the provider. Why you care: the model learned nothing about your business during training.
7. Training cutoff
The date after which the model saw no new text. Anything newer has to be handed to it at question time. Why you care: a model’s general knowledge is never current or specific to you.
8. Inference
The act of running the model to produce an answer, as opposed to training it. This is what you pay for per use. Why you care: “inference cost” on an invoice means usage cost.
9. Token
The unit the model reads and writes in, roughly three quarters of a word in English. Providers bill per token, usually quoted per million. Why you care: cost depends on how much text goes in and out, not on how many questions are asked.
10. Context window
The maximum amount of text the model can consider at once: instructions, documents, conversation, and answer combined. Large in the current models, but accuracy drops when it is filled with irrelevant material. Why you care: good systems select what the model sees rather than dumping everything in.
11. Prompt
The text sent to the model, including standing instructions about its role, rules, and output format. In a business system it is a specification, not a casual question. Why you care: most of the quality lives here, and it should be versioned like a contract template.
12. System prompt
The part of the prompt that sets the standing rules for every conversation: who the model is, what it may and may not do, how to answer. Why you care: ask to read your vendor’s system prompt; it is where the promises are kept or broken.
13. Temperature
A setting that controls how much variation the model allows in its answers. Low means consistent, high means creative. Why you care: business systems should run low so the same question gets the same answer.
14. Hallucination
A confident, fluent, false statement produced because the model composes likely text rather than retrieving facts. It is ordinary behavior, not a rare bug. Why you care: every customer-facing system must assume it will happen.
15. Reasoning model
A model configured to work through a problem step by step before answering, at the cost of more time and tokens. Why you care: useful for hard judgment tasks, wasteful for sorting emails.
16. Multimodal
A model that accepts images, audio, or video as well as text. The current models can read a photo of an invoice or a scanned form. Why you care: this is what makes document extraction from messy PDFs work.
Your material and how the model reads it
17. Knowledge base
The written material you want the AI to answer from: policies, FAQs, price lists, manuals, past answers. Why you care: the knowledge base is the product; a bot is only as good as what is in it.
18. RAG (retrieval-augmented generation)
Finding the relevant passages from your material first, then having the model write an answer from only those passages. It is how you stop the model guessing. Why you care: when a vendor says “it knows your documents,” this is what they should mean (What Is RAG? Retrieval-Augmented Generation for Business Owners).
19. Retrieval
The “finding the relevant passages” step in RAG. Its quality decides whether the model gets the right three paragraphs or the wrong ones. Why you care: most RAG failures are retrieval failures, not model failures.
20. Chunking
Splitting long documents into pieces small enough to search and hand to the model. Chunk badly and the answer gets cut in half. Why you care: it is a detail that decides whether your bot can answer questions about page 40.
21. Embedding
A numerical fingerprint of a piece of text that captures its meaning, so similar passages have similar fingerprints. Why you care: it is why the bot can match “money back” to your “refund policy.”
22. Vector database
A database built to store embeddings and find the closest matches quickly. Pinecone is a hosted one; pgvector adds the same ability to an ordinary Postgres database. Why you care: pgvector often means you do not need a new vendor.
23. Fine-tuning
Adjusting a model’s weights by training it further on your own examples. Expensive to do well and hard to correct. Why you care: for most businesses RAG is the better tool; be skeptical of a vendor who leads with fine-tuning.
24. Grounding
Tying the model’s answer to specific source material so it can only say what the material supports. Often paired with citations. Why you care: “grounded” is the word to look for in any customer-facing AI.
25. Citation
The model pointing to the exact passage its answer came from. Why you care: a citation lets a human check the answer in seconds, and a system that cannot cite cannot be trusted with customers.
Actions, tools, and agents
26. API
A way for one piece of software to talk to another over the internet. Your automation talks to the model through the provider’s API. Why you care: “API access” is what turns a chat toy into something that can plug into your systems.
27. Tool calling (function calling)
Giving the model a menu of actions it can request, such as look up an order or check the calendar, with real software doing the action. Why you care: this is how AI touches your systems, and the menu is where you set the limits.
28. Structured output
Forcing the model to answer in a fixed shape (named fields, allowed values) rather than free prose, so ordinary software can act on it. Why you care: money should only move on structured, validated output, never on a paragraph.
29. Agent
An AI system that takes several steps on its own, choosing tools and deciding what to do next, rather than answering once. Why you care: more capable, more risk; most businesses should start with fixed workflows (AI Agents vs Automation: Which One Your Business Needs).
30. Workflow (automation)
A fixed sequence of steps triggered by an event: a form arrives, the AI classifies it, a record is created, a person is notified. n8n, Make, and Zapier build these. Why you care: predictable, cheap, and the right first project for most businesses.
31. MCP (Model Context Protocol)
A standard way for an AI to connect to tools and data sources, so a connection built once can be reused across models and products. Why you care: less lock-in, and security questions to ask (MCP (Model Context Protocol) Explained for Business Owners).
32. Human in the loop
A design where a person reviews or approves the AI’s output before it takes effect. Why you care: it is the difference between a system you can trust with customers and one you cannot.
33. Agentic coding
Software that plans and writes software under instruction, such as Claude Code, Cursor, and similar tools. Why you care: it has changed what custom software costs and how fast it arrives.
Building and running it
34. Integration
Connecting the AI system to something you already run: your CRM, accounting, calendar, phone system. Why you care: it is where most of the effort and value in a project sits.
35. Webhook
A way for one system to notify another the moment something happens (a payment lands, a form is submitted). Why you care: it is the trigger for most automations and why they run in seconds rather than overnight.
36. Latency
The delay between asking and getting an answer. Matters most for voice, where anything over about a second feels wrong. Why you care: ask a phone-agent vendor for their latency number.
37. Rate limit
The cap a provider places on how many requests you can make per minute. Why you care: a system built without handling it will fail on your busiest day.
38. Self-hosting
Running software (an automation platform like n8n, or an open-weight model) on a server you control rather than the vendor’s cloud. Why you care: more control and lower cost at volume, in exchange for owning updates and backups.
39. Evals (evaluations)
A set of test questions with known good answers, run against the system before launch and after every change to catch regressions. Why you care: a vendor who has no evals is guessing about quality.
40. Logs (audit log)
The record of every input, output, tool call, and decision the system made. Why you care: without logs nothing can be debugged or reviewed; in the systems we build it is the first thing added.
Risk and safety
41. Guardrails
The rules and checks that constrain what the AI may say and do: allowed topics, forbidden actions, required refusals, output validation. Why you care: this is the word for the safeguards you should be asking about.
42. Prompt injection
An attack in which a document, email, or web page contains hidden instructions that hijack the AI’s behavior. Why you care: the moment your AI can read outside content and take actions, this is a real risk.
43. Least privilege
Giving the AI system only the access it needs for its task: its own account, limited keys, spending caps. Why you care: an AI with your admin login is a liability.
44. Zero data retention
A provider setting or contract term under which your inputs are not stored after the answer is produced. Why you care: it is the setting to ask for when customer data passes through the model (AI Data Privacy for Business, What Happens to Customer Data).
45. PII (personally identifiable information)
Data that identifies a person: names, addresses, phone numbers, account numbers, health details. Why you care: it should be masked or minimized before it reaches any model, and your provider terms should cover it.
Picture a business like this one
The business below is a composite of the kind of company that writes to us, not a client. The numbers describe the shape of the problem, not a case study.
Picture a business like this one: a property management company with nine staff and about six hundred units. They receive a proposal for “an AI tenant assistant with RAG over your knowledge base, agentic maintenance triage, and MCP integrations.” Six months ago the owner would have signed or walked away on instinct.
With the glossary open, the questions write themselves. What is in the knowledge base and who keeps it current? What does the agent do on its own, and which steps have a human in the loop? Which tools does the MCP connection expose, under whose account? Are answers grounded with citations? Where are the logs? What is the zero-retention position on tenant data?
What gets built is smaller and better than what was proposed: a grounded assistant that answers lease and policy questions with citations, a classification step that routes maintenance requests by urgency, and a log the office manager reads on Fridays. What changes is that the owner bought parts they understood, at a price they could judge.
What it costs to run
The terms above map onto a short cost sheet. Model usage (tokens) for a business this size is typically tens of dollars a month. A vector database is $0 to $50 a month with pgvector on an existing Postgres, more for a hosted service. An automation platform is a subscription or a $10 to $30 server if self-hosted. Integrations and evals are one-time build work plus a few hours a quarter of upkeep. Check each vendor’s current pricing page.
The mistakes we see most
- Hearing “AI” and not asking which of the four parts is meant.
- Accepting “it knows your documents” without asking about retrieval, chunking, and citations.
- Buying an agent when a workflow with one AI step would do.
- Signing without knowing the data retention terms.
- Never asking to see the logs, the system prompt, or the evals.
When to bring in help
An owner armed with this AI glossary for business meetings can run a no-code automation on Make or Zapier, subscribe to a chat assistant, and vet a vendor proposal without being talked over.
A developer is needed when the system must be grounded in your material, integrated with your real tools, logged, evaluated, and guarded, because those are built, not configured. Levelbrook builds exactly that for businesses, fixed price from a written scope, everything in accounts you own, and every term on this page shows up in the scope in plain English. The form below is how that conversation starts.