AI Terms Cheat Sheet: 25 Words in Plain English

The 25 AI terms you hear most, one plain line each: how the model works, giving it your information, getting work done, and keeping it safe and worth it. Print it for the office.

The 25 AI terms you hear most, one plain line each, for the office wall. The full explanations, with what each one means for a business your size, are in the Applied AI Insider article "25 AI Terms Everyone Uses and Few Explain."

• Tokens: the small pieces AI reads text in. Limits and bills are counted in tokens.
• Context window: what the AI can see at one time. Working memory, not a filing cabinet.
• Next-word prediction: answers are written one likely piece at a time, not looked up.
• Hallucination: a confident answer that's wrong. Fluent isn't the same as true.
• Reasoning model: a slower "thinking" mode for problems with several steps.
• Model vs app: the same model can sit behind very different plans and data rules.

• Grounding: tying an answer to a source you can check.
• RAG (retrieval): search your documents first, then answer from them. Bad files, bad answers.
• Embeddings / semantic search: search by meaning. Can miss exact part and invoice numbers.
• Memory: notes an assistant keeps between chats. Know what it saved.
• Projects, GPTs, Gems: saved instructions and files for one recurring job.
• Connectors / MCP: how AI reaches your email, calendar and files. Each one is a key.

• System prompt: the hidden instructions an app gives the model.
• Context engineering: choosing what the AI sees. The right file beats a better prompt.
• Tool calling: the AI asks, your software does. A request isn't proof.
• Agent: an AI working in a loop: plan, act, check, repeat. Autonomy is a dial.
• Computer use: AI clicking and typing in a browser for you. Keep it off payment buttons.
• Skills: a saved routine the AI can reuse. Teach it once.

• Prompt injection: outside text trying to give orders. Outside text is data, never orders.
• Guardrails: rules the software enforces, not rules the AI promises.
• Human in the loop: a person approves the risky steps. Customers and money get a person.
• Evals: ten real examples you re-run whenever something changes.
• Training on your data: business plans say no by default. Check personal accounts.
• Fine-tuning: retraining the model. Most small businesses don't need it.
• Credits / metered billing: pay-per-use AI. Set the spending limit first.

Four questions to ask about any AI tool: What can it see? What can it do? Who approves it? What does it cost?

From Applied AI Insider, the free applied-AI library for people who run a small business: appliedaiinsider.com

A template from the template library.