Understanding AI
Glossary

AI terms, in plain English

The 33 core terms behind the lessons, what each one means, and where you are likely to meet it at work. Nothing here is required reading. Dip in when a word trips you up.

Palm card (PDF)

All 33 terms and the harness cheat sheet on one A4 page, two cards to a sheet. Print it, cut it in half, keep it by your desk. Optional.

Download PDF

Foundations: How AI Works

Seven core concepts that explain what's happening under the hood

LLMLarge Language Model
AI technology trained on vast amounts of text data to understand and generate human language. Powers tools like ChatGPT, Claude, and Gemini.
Where you will see it: Used whenever you interact with an AI chatbot or generative tool. Your business might use this to automate customer service, draft content, or analyse documents.
Prompt
The question or instruction you give to an AI. The clearer and more specific your prompt, the better the AI's response.
Where you will see it: Every time you write to ChatGPT or ask an AI for help. A well-crafted prompt to Claude could save your team hours on research or analysis.
Token
A small unit of text that an AI processes. Roughly 4 characters or 1 word. Important because many AI tools charge based on token usage.
Where you will see it: Matters when calculating costs for AI API usage. A 1000-word document uses approximately 1300 tokens.
Hallucination
When an AI generates confident-sounding but false or made-up information. It looks plausible but is actually incorrect.
Where you will see it: Critical to understand when using AI for research or fact-based work. Always verify AI outputs, especially for numbers, dates, or regulatory claims.
Context Window
The maximum amount of text an AI can read and consider at once. Newer models have larger context windows, allowing them to process longer documents.
Where you will see it: Affects what you can ask an AI to do. Claude 3.5 Sonnet has a 200k token context window - enough for entire books in a single conversation.
Chain of Thought
Prompting technique where you ask the AI to show its working - explain step-by-step reasoning rather than just giving a final answer.
Where you will see it: Improves accuracy on complex problems. Instead of asking 'Is this contract risky?', ask 'Walk me through the risk factors in this contract, step by step.'
Temperature
A setting that controls how creative or unpredictable an AI's response is. Low temperature (0.0) equals predictable, repeatable answers. High temperature (1.0) equals more varied, creative responses.
Where you will see it: Customer service chatbots use low temperature for consistency. Content creation tools often use higher temperature for variety.

Retrieval & Data: Making AI Smart About Your Business

How to give AI access to your own information so it stays current and accurate

RAGRetrieval-Augmented Generation
A technique that lets AI access current information by pulling it from your own documents or databases, then generating answers based on that fresh data.
Where you will see it: An accountancy firm could use RAG so their AI references current tax law. A law practice could pull from recent cases. Avoids the AI making up outdated info.
Embeddings
A way to convert text into numerical patterns that capture meaning. Allows the AI to understand relationships between concepts and ideas.
Where you will see it: Behind-the-scenes in RAG systems. Enables semantic search - finding documents by meaning, not just keyword matching.
Vector Database
Specialised storage for embeddings (numerical patterns of text). Allows fast retrieval of similar documents based on meaning rather than keyword.
Where you will see it: Powers the retrieval part of RAG. A business could store product manuals or policies, then instantly find relevant sections even if phrased differently.
Fine-tuning
Training an AI on your own specific data to make it behave differently or know domain-specific information. More feasible than full retraining.
Where you will see it: A legal practice might fine-tune an AI on their firm's case history. A retailer might fine-tune on their product catalogue to improve recommendations.
Knowledge Base
A collection of documents, data, or information that an AI system can access and reference when answering questions.
Where you will see it: A customer support chatbot uses your FAQ and product docs as its knowledge base. Sales teams could create a knowledge base of competitive intel or sales scripts.

Business Outcomes & Platforms

Measuring value and choosing the right platforms for your business

CMSContent Management System
The platform used to build and manage a website. Examples include WordPress, Wix, Squarespace, or Webflow. Affects what technical improvements are possible.
Where you will see it: When assessing your website or planning AI integration. Some CMS platforms support AI plugins more readily than others.
UXUser Experience
How easy, intuitive, and satisfying a website or app is to use. Covers navigation, layout, mobile responsiveness, speed, and how clearly visitors are guided toward action.
Where you will see it: When designing customer-facing tools or AI interfaces. Poor UX means users won't adopt your AI solution, no matter how smart it is.
ROIReturn on Investment
The measurable value generated relative to cost. In an AI context: revenue or leads generated, time saved, or costs reduced compared to the cost of the AI implementation.
Where you will see it: When justifying AI investment to leadership. 'This AI tool costs NZD 500 per month but saves 10 hours of admin work, worth NZD 2000 monthly.'
GEOGenerative Engine Optimisation
Structuring content so AI tools like ChatGPT, Claude, and Perplexity cite or surface it as a direct answer. Different from traditional SEO.
Where you will see it: Your business might use this to ensure the AI tells prospects about you. An accountancy firm could optimise for 'AI-recommended accountants in Auckland.'
AEOAnswer Engine Optimisation
Structuring content so AI tools can find and surface it as a direct answer without the user needing to click through. Overlaps with GEO.
Where you will see it: A law firm optimises their 'Employment Law FAQ' so when someone asks an AI 'What are my rights?', the AI cites their content directly.

Tools & Platforms: Building Blocks

The components you use to build AI-powered business solutions

APIApplication Programming Interface
A standardised way for different software systems to communicate and share data. APIs let you plug AI into your existing tools.
Where you will see it: Your accounting software could use an accounting-focused API to automate invoice processing. CRM systems use APIs to integrate AI features.
Model
The actual AI engine - the trained system that powers the responses. Different models have different strengths (speed, accuracy, creativity).
Where you will see it: When choosing a tool, you are often choosing which model to use. Claude 3.5 Sonnet is a model. ChatGPT GPT-4 is a different model.
Chatbot
An AI system designed for conversation. Can answer questions, provide support, or assist with tasks through a chat interface.
Where you will see it: Customer service teams use chatbots to handle routine queries 24/7. Many businesses are adding AI chatbots to their websites.

Agents & Automation: AI That Takes Action

Moving beyond chatbots to systems that plan, decide, and execute

Agent
An AI system that can decide what action to take, break tasks into steps, and use tools to achieve a goal rather than just answering questions.
Where you will see it: An AI agent could review your calendar and email to automatically propose meeting times. A business agent might search the web, compile data, and write a report with minimal human input.
Agentic AI
The broader concept of AI systems that act autonomously - taking initiative, making decisions, and completing workflows with minimal human involvement.
Where you will see it: Next generation of business automation. Instead of tools that answer questions, agentic AI could handle entire processes (hiring, compliance checks, lead qualification).
Orchestration
Coordinating multiple AI tools or agents to work together in sequence. One tool completes a task, passes results to the next, and so on.
Where you will see it: A workflow might orchestrate: AI searches regulations, passes findings to another AI for analysis, then passes to a third to draft a summary.
Orchestrator
The AI system or software that manages orchestration - deciding which agent or tool to use, when to use it, and what to do with the results.
Where you will see it: In a complex AI workflow, the orchestrator is the conductor. It receives a business request and determines which sequence of agents will best complete it.
Tool Use
Giving an AI the ability to call external tools - search engines, calculators, databases, APIs - to complete tasks beyond pure text generation.
Where you will see it: An AI with tool use can run a real Google search instead of relying on training data. Can query your actual database instead of guessing.
HarnessAgent Harness
The software wrapped around an AI model that lets it do real work. It supplies the instructions and context, offers tools, carries out the actions the model asks for, remembers progress, and enforces permissions. The model is the engine; the harness is the rest of the car.
Where you will see it: Claude Code and ChatGPT's agent mode are harnesses. When you compare AI tools you are often comparing harnesses as much as models: what data they can reach, which actions need your approval, and how much they log. See 'How an AI Harness Works' below.
Prompt Engineering
The practice of designing prompts strategically to get the best responses from an AI. Includes structure, examples, and clear instructions.
Where you will see it: Skilled prompt engineering can save significant time. Instead of rephrasing and retrying, you craft one effective prompt that produces exactly what you need.
MVPMinimum Viable Product
The simplest version of an AI solution that delivers core value. Built quickly to test ideas before investing in full development.
Where you will see it: You might build an MVP chatbot to test whether customers like the idea before spending on enterprise integration. MVP helps you learn what actually works.
Worked example

How an AI Harness Works

The software around the model that turns text generation into safe, useful action

A model on its own only produces text. A harness is everything around the model that turns it into a working assistant: it gives the model its instructions and context, offers it tools, carries out the actions the model asks for, remembers what has happened so far, and keeps people in control. Claude Code, ChatGPT's agent mode and the AI features inside many business apps are all harnesses wrapped around a model.

Want the longer version? Agent Harnesses and MCP is a lesson in the Professionals tier.

  1. You set a goal. You type a request, or another system triggers one.
  2. The harness builds the context. It combines the instructions, your request, relevant files or history, and a list of the tools the model may use, and fits it all inside the model's context window.
  3. The model decides the next step. It either gives an answer or asks for a tool to be used, for example 'search the knowledge base' or 'read this invoice'.
  4. The harness checks and acts. The model cannot act on its own. The harness checks the request against its rules and permissions, runs the tool, and captures the result. Risky actions can pause for a person to approve.
  5. The result goes back to the model. The outcome is added to the context and the model chooses what to do next. This loop repeats until the task is finished or a limit is reached.
  6. The harness keeps memory and records. It trims or summarises long histories, keeps notes between sessions, and logs what was done.

Why it matters for your business

  • Same model, different results. Quality, safety and cost depend on the harness as much as on the model, so two tools built on the same model can behave very differently.
  • The guardrails live here. Permissions, approvals, data access and logging are enforced by the harness, not by the model.
  • Questions to ask a vendor: What tools can it use? What data can it see? Which actions need a person's approval? Is everything logged? Where is the data processed (data residency)?
  • Related terms: Agent, Tool Use, MCP, Orchestration, Guardrails, Context Window.

Safety & Integration: Responsible AI in the Real World

How to keep AI secure, compliant, and connected to your business systems

Guardrails
Rules or boundaries built into an AI to prevent unwanted behaviour - refusal to answer harmful questions, preventing data leakage, maintaining consistency.
Where you will see it: A business chatbot has guardrails to never share customer data or financial details. Compliance-critical industries rely on guardrails to stay within regulations.
Privacy
Protecting personal and sensitive data from unauthorised access. Critical when AI processes customer data, employee records, or business information.
Where you will see it: When implementing AI that handles customer data, you must ensure privacy safeguards. GDPR and NZ Privacy Act 2020 set requirements for how data is used.
Data residency
The physical location where your data is stored. Important for compliance - some jurisdictions require data to stay within their borders.
Where you will see it: Government agencies might require data residency in NZ. An international business might need data in multiple regions for compliance.
Compliance
Meeting legal and regulatory requirements applicable to your business and its use of AI. Varies by industry and jurisdiction.
Where you will see it: Financial services must comply with FSCL rules. Healthcare must comply with Privacy Act. Your AI implementation must respect these constraints.
MCPModel Context Protocol
An open standard that lets AI tools safely connect to external systems and services, providing a standardised way to access databases, APIs, and enterprise tools.
Where you will see it: Enables Claude to integrate with your Slack, Google Drive, or Jira without custom development. Emerging standard for business AI integration.

Further Reading: 20 Additional Terms

These terms are useful to know but less critical for introductory AI literacy. You may encounter them in technical discussions or specialised use cases.

Vibe-coding
An AI-dependent programming practice where a programmer describes a problem in a few sentences as a prompt to an LLM tuned for coding.
Open Claw
An open-source, self-hosted AI personal assistant that runs on your own devices and can autonomously execute tasks across apps and platforms.
NLPNatural Language Processing
How AI understands and processes human language to extract meaning and intent.
MLMachine Learning
Teaching computers to learn patterns from data without explicit programming for each scenario.
Transformer
The neural network architecture that powers modern LLMs by processing text in parallel rather than sequentially.
Attention Mechanism
A technique that lets AI focus on the most relevant parts of input data when generating responses.
Inference
The process of running an AI model on new data to generate predictions or outputs.
Backbone
The core foundational component of an AI system that handles the primary processing.
Feature Extraction
Identifying and pulling out the most important characteristics from raw data for AI to learn from.
Supervised Learning
Training an AI on labelled data (inputs paired with correct outputs) so it learns the relationship between them.
Unsupervised Learning
Training an AI on unlabelled data to find patterns and structure without being told what to look for.
Gradient Descent
An optimisation algorithm that adjusts AI model parameters iteratively to reduce errors during training.
Backpropagation
A technique for calculating how to adjust model parameters based on errors, working backwards through the network.
Parameterization
The process of defining the adjustable values (weights) that an AI model uses to make decisions.
Loss Function
A mathematical measure of how far off the AI's predictions are from the correct answers during training.
Training Dataset
The collection of examples used to teach an AI model how to perform a task.
Bias
Systematic errors or unfair patterns in AI outputs, often caused by biased training data or design choices.
Agent swarm
Multiple AI agents working together as a coordinated group to solve complex problems collaboratively.
Back end
The server-side infrastructure and code that powers applications, hidden from users but essential for functionality.
Tech Stack
The complete set of programming languages, frameworks, tools, and technologies used to build a system.