The AI vocabulary cheat sheet: 15 terms every operator should be able to use in a sentence

You shouldn't have to fake it during AI conversations. Here are 15 terms that come up in every "AI for business" conversation in 2026, defined in plain English, with one example sentence each.

Print this. Use it.

The basics

1. Large Language Model (LLM) A type of AI trained on huge amounts of text that can read and write in natural language. Claude, GPT-4, Gemini, and Llama are all LLMs. "The chatbot is built on an LLM, so it can answer free-form questions instead of just FAQ-style ones."

2. Prompt The instructions you give an AI to get the output you want. Can be one sentence or three pages. "My first prompt was vague, so I rewrote it with more specifics and the output got way better."

3. Token The unit AI uses to count text. Roughly 1 token = 0.75 of a word. Pricing for AI APIs is usually "per million tokens." "That document was 50,000 tokens, which is why processing it cost about $0.30."

4. Context window The maximum amount of text an AI can hold in working memory during a single conversation. Bigger is better — modern models hold 100K–1M tokens. "The 200K context window means I can paste in our whole sales playbook before asking it to draft a proposal."

How it works under the hood

5. Training data The text and images the AI learned from. Older models trained on data through 2023; newer ones go through 2025. "The model has a knowledge cutoff in late 2024, so it doesn't know about anything that happened this year."

6. Fine-tuning Taking a general AI and training it further on your specific data so it gets better at your specific task. "We fine-tuned Claude on our last 200 proposals so it drafts in our voice."

7. Embedding A way of representing text or images as numbers so AI can compare them by meaning, not just by exact words. Powers most "search" features in modern AI tools. "The agent uses embeddings to find similar past proposals when drafting a new one."

8. Retrieval-Augmented Generation (RAG) A pattern where AI looks up relevant info from your documents before generating a response, instead of relying only on its training. Most "AI that knows your company" tools work this way. "RAG is why our agent can answer questions about our internal docs even though they weren't in the original training set."

The agent stack

9. Agent AI that can take multi-step actions across different tools, not just answer questions. The category that's actually changing how growing teams run. "The agent reads my email, decides which ones need replies, drafts them, and pushes them to my outbox for review."

10. Tool use (or function calling) The mechanism that lets an AI call other software (your CRM, your email, your calendar). What turns a chatbot into an agent. "With tool use enabled, the AI can actually create the HubSpot contact instead of just describing what it would do."

11. Workflow / Chain A sequence of AI steps that produce a finished output. Read email → classify → draft → log → notify is a workflow. "The proposal-drafter is a 6-step workflow that pulls from past wins, formats, and routes for review."

The risk side

12. Hallucination When AI makes up something that sounds plausible but isn't true. The most common failure mode in 2026. "The first draft hallucinated a citation that doesn't exist, so we added a verification step before anything goes out."

13. Guardrails Rules that prevent AI from doing certain things. Can be technical (the model refuses) or operational (a human review step). "We put guardrails on the agent so it can't send anything to a client without a human approving."

14. Inference cost What it costs to run AI to answer one query. Usually small — fractions of a cent, but adds up at scale. "At 5,000 daily queries, our inference cost is about $40/month."

15. Model API The connection that lets your software talk to the AI provider. OpenAI API, Anthropic API, Google API. The plumbing under most modern AI tools. "We're using the Anthropic API directly, which gives us better control than buying a SaaS wrapper."

How to use this list

You don't need to memorize this. Bookmark it. The next time a vendor uses one of these terms in a demo, you'll know whether they're using it correctly or just dressing up basic software with AI vocabulary.

Three terms that should set off alarm bells when used wrong:

• "AI-powered" with no model named: marketing, not engineering. Ask which model.
• "Trained on your data" without explaining if it's fine-tuning or RAG: probably RAG, which is fine, but they should say so.
• "Eliminates hallucinations": no AI tool eliminates hallucinations entirely in 2026. Reduces them, sure. Eliminates, no.

The next post in this series covers prompts in detail, how to write the instructions that actually get you what you want.