Resources / Glossary

AI in plain English: a glossary for business owners

Every AI term a business owner is likely to hear in 2026, defined in 1–2 sentences. We update this quarterly. If a term you need is missing, email us and we'll add it.

Reviewed May 5, 2026 by Level Up Automate.

A

Acceptable Use Policy: A document telling staff how they can and cannot use AI tools at work. Usually one page; lives next to your employee handbook.
Agentic AIalso: AI agents: AI systems that take multi-step actions on a user's behalf — for example, booking travel or processing a refund — rather than just answering a question.
AI Act: Usually refers to the European Union's AI Act, the first comprehensive AI law in a major economy. Sorts AI uses into four risk tiers and applies to anyone whose AI affects EU residents.
AI Governance: The combination of policies, processes, and oversight that lets a company use AI safely. The smaller the company, the simpler this can be.
AI Incident: Any time AI causes harm — wrong information sent to a customer, data leaked, an embarrassing output — or had a near-miss. Worth documenting even if no real damage was done.
AI Policy: A short document setting the rules for AI use at your company. Best ones are one page, plain-English, and reviewed quarterly.
AI Readiness: How prepared your business is to use AI safely and effectively today. Most small businesses overestimate this.
Algorithmic Discrimination: When an AI system's decisions consistently disadvantage a protected group. A legal risk in hiring, lending, housing, and similar 'consequential decision' uses.
Anthropic: The AI safety company behind the Claude family of AI assistants and the Claude Code coding tool. Headquartered in San Francisco. Considered one of the two leading consumer-AI vendors alongside OpenAI.

B

Bias (in AI): Patterns in AI output that reflect skewed training data or flawed model design — often along lines of gender, race, geography, or age. Real but manageable in most small-business uses.

C

Chatbot: A software application that converses in text. Modern chatbots are usually built on large language models (LLMs).
ChatGPTalso: GPT chat: The most widely-used consumer AI assistant, made by OpenAI. Plans: Free, Plus (individual), Team and Enterprise (business with data protection). The closest direct competitor is Anthropic's Claude.
Claudealso: Anthropic Claude, Claude AI: Anthropic's family of AI assistants. Plans: Free, Pro (individual), Team, and Enterprise. Widely considered the strongest assistant for long-form writing, analysis, and conversational tasks. Claude Free and Pro do not train on user conversations by default — a more conservative posture than the equivalent ChatGPT consumer tiers.
Claude Codealso: Anthropic Claude Code: Anthropic's coding assistant — a terminal- and IDE-integrated tool that lets developers delegate code-writing, refactoring, and review tasks to Claude. Peer to GitHub Copilot and Cursor. Used by engineering teams as part of the same approved-tools list as their other AI assistants.
Colorado AI Act: A 2026 Colorado state law requiring risk management and consumer notice for 'high-risk' AI systems making consequential decisions. The first broad AI law in the United States.
Confidence Threshold: A score above which AI output is allowed to act automatically; below the score, a human reviews. Setting these well is most of the work in building safe AI workflows.
Copilot: Microsoft's AI assistant, integrated into Office, Windows, and Edge. Many versions exist; the one in your business depends on your Microsoft 365 plan. Note: GitHub Copilot is a different (and older) Microsoft product aimed at developers — peer to Anthropic's Claude Code and Cursor.

D

Data Leak: Confidential information ending up somewhere it shouldn't. AI tools create new leak paths because employees paste data into them, often without thinking.
Deepfake: Synthetic audio, video, or images that convincingly impersonate a real person. Increasingly used in scams targeting finance and HR teams.

E

Embedding: A way of representing text, images, or other data as numbers so AI can compare and search them. Powers most AI search and recommendation features.
EU AI Act: The European Union's comprehensive AI law. Applies to any AI system whose output is used in the EU, even if the company is outside Europe.

F

Fine-tuning: Adjusting an AI model with additional training data so it performs better at a specific task. Usually done by vendors; rarely something a small business does directly.
Foundation Modelalso: Base model, General-purpose AI: A large, general-purpose AI model that other AI products are built on. Examples: GPT-4, Claude, Gemini.

G

Generative AIalso: GenAI: AI that produces new content — text, images, audio, video, code — rather than only classifying or scoring existing content.
GPT: Generative Pre-trained Transformer — the family of language models from OpenAI. Numbers (GPT-3.5, GPT-4, GPT-5) refer to successive generations. Anthropic's equivalent is the Claude family of models; Google's is Gemini.
Guardrails: Rules and constraints built around an AI system so it stays within safe behavior. Often a combination of model rules, software checks, and human review.

H

Hallucination: When AI confidently produces information that is wrong or invented. The single biggest reason every AI output for a customer needs human review.
High-Risk AI System: Term used in the Colorado AI Act and EU AI Act for AI that makes consequential decisions about people in regulated areas like employment, lending, healthcare, and housing.
Human-in-the-Loopalso: HITL: Workflow design where a human reviews or approves AI output before it acts. Standard practice for any high-stakes AI use.

I

Impact Assessment: A written analysis of how an AI system will affect users — covering risks, mitigations, and benefits. Required by several state laws for high-risk AI.
ISO/IEC 42001: An international standard for AI management systems. You can be formally certified against it. Useful mainly for companies with enterprise customers asking about it.

J

Jailbreak: A prompt or trick that gets an AI tool to ignore its safety rules. A risk to be aware of, especially for tools used by the public.

L

LLM (Large Language Model): An AI trained on huge amounts of text that can read, write, summarize, and answer questions. ChatGPT, Claude, Gemini, and Copilot are all built on LLMs.

M

Machine Learningalso: ML: A broader category of AI in which systems learn patterns from data. AI tools you use every day (spam filters, recommendations) are built with machine learning.
Model Drift: When an AI model's performance changes over time as inputs or the world shifts. A reason to monitor AI use even after it's working well.

N

NIST AI RMF: The U.S. National Institute of Standards and Technology's voluntary framework for managing AI risk. Increasingly the de facto baseline for U.S. business AI governance.

O

Open-source AI: AI models whose weights are published openly so anyone can run them. Different from proprietary models like GPT or Claude. Important for businesses needing on-premise control.

P

Prompt: The instruction you give to an AI tool. Better prompts produce better output.
Prompt Engineering: The practice of writing prompts that consistently get good results. Less of a specialized job than it sounds; most staff can learn the basics in an hour.
Prompt Injection: A security attack where hidden instructions in a document, image, or webpage trick an AI into doing something unintended. Affects AI tools that read external content.

R

RAG (Retrieval-Augmented Generation): An AI design pattern where the model is given relevant documents to read before answering a question. Reduces hallucinations and lets AI 'know' your business's documents.
Responsible AI: An umbrella term for the practices that keep AI use ethical, safe, and aligned with business values. Includes governance, fairness, transparency, and accountability.
Risk Assessment: A written analysis of the risks of using a specific AI tool or workflow. Covers data, accuracy, dependence, and mitigation.

S

Shadow AI: AI use inside a company that wasn't authorized. Almost always present at companies that haven't communicated a policy yet.
SOC 2: A third-party audit of a vendor's security practices. Common ask when evaluating AI vendors. Type II is more rigorous than Type I.
Synthetic Media: Audio, video, or images generated by AI rather than recorded from reality. Includes deepfakes but also legitimate uses like AI voiceovers and product imagery.

T

Tokens: The chunks of text AI models read and produce. Roughly four characters or three-quarters of a word per token. Most pricing is per token.
Training Data: The data an AI model learned from. Whether your business's data becomes training data for a vendor's model is a critical contract question.
Transparency: Telling people when they're interacting with AI, what data it uses, and how decisions are made. Required in some jurisdictions for certain AI uses.

V

Vendor Due Diligence: The process of evaluating an AI vendor's security, data handling, and reliability before signing a contract. Worth one to four hours per vendor.

W

Watermarking: Embedding a subtle, machine-detectable signal into AI-generated content so it can be identified later. Increasingly a regulatory expectation for synthetic media.

Z

Zero-shot: When an AI handles a task without being given examples first. Most everyday AI use is zero-shot. Adding even one example often improves results dramatically.