AI glossary

Plain-English definitions of the AI terms you'll keep hearing, without the jargon. Written for Australian small-business owners who don't need to read another vendor white paper.

A

  • Agent , An AI system that can autonomously plan, execute and iterate on a multi-step task using a language model plus tools. Different from a chatbot, agents act, they don't just respond.
  • Agentic AI , An AI system that takes actions in the world (reads files, writes responses, books appointments, sends emails) in an autonomous loop, rather than only chatting in response to prompts.
  • Agentic loop , The repeating cycle where an AI agent decides what to do, takes an action, observes the result, and decides what to do next, until the task is complete.

C

  • Context window , The maximum amount of text (in tokens) a language model can hold in its working memory at once. Modern frontier models support 200k-1M+ tokens of context.

E

  • Embedding , A numerical representation of text (or images, audio) as a vector. Similar meanings produce similar vectors. The foundation of semantic search and most RAG systems.

F

  • Fine-tuning , Continued training of a pre-trained language model on a smaller dataset of domain-specific examples. Adjusts the model's behaviour without training from scratch.
  • Function calling , When a language model emits a structured request to invoke a specific function (with typed arguments), rather than free-form text. The underpinning of tool use.

H

  • Hallucination , When a language model generates information that's false but stated confidently. The defining failure mode of modern AI. Mitigated, not eliminated.

I

  • In-context learning , When a language model learns from examples provided in the prompt itself, without any weight changes. The trick behind few-shot prompting.

J

  • Jailbreak , Coaxing an AI to produce output it's been trained to refuse. Different from prompt injection; both are security concerns for businesses deploying AI.

L

  • Large Language Model , A neural network trained on enormous amounts of text to predict and generate language. The 'brain' behind Claude, GPT, Gemini and other modern AI products.
  • Latency , The time between sending a request to an AI model and getting a response. Matters most for user-facing features; not so much for background batch work.

M

  • Model Context Protocol , An open standard for connecting AI assistants to external tools and data sources. Anthropic's protocol that lets Claude (and others) talk to Xero, Shopify, Slack, etc, in a structured way.
  • Multimodal , A model that accepts more than one type of input (text + images, sometimes audio + video). Modern Claude, GPT, Gemini are all multimodal.

P

  • Prompt , The input you send to a language model. Includes your question or instruction, plus any context, examples, and rules you give the model to work from.
  • Prompt caching , A feature that lets language models reuse previously-processed prompt content at a discounted rate, usually ~90% off the input price. Major cost-saver on multi-turn conversations.
  • Prompt injection , A security risk specific to AI systems: untrusted text (in an email, document, or web page) containing hidden instructions that hijack the AI to do something the user didn't intend.

R

  • Reasoning model , A language model trained or configured to think step-by-step before answering, often visibly. Better at hard problems, slower and more expensive than standard models.
  • Reinforcement Learning from Human Feedback , A training technique where humans rank a model's outputs and the model is fine-tuned to prefer the higher-ranked responses. How Claude, GPT and others were taught to be helpful instead of just plausible.
  • Retrieval-Augmented Generation , A pattern where an AI model retrieves relevant information from a knowledge base before generating its response. Lets AI answer questions about your specific data without retraining the model.

S

  • Structured outputs , Constraining a language model to produce output that conforms to a specific schema (e.g. valid JSON matching a defined shape). Removes the guesswork from parsing AI responses.
  • System prompt , The instructions you give a language model upfront, before any user message. Sets the model's role, voice, rules and constraints for the whole conversation.

T

  • Token , The unit of text a language model reads and writes. Roughly 4 characters of English text per token. Models are priced per million tokens.
  • Tool use , When a language model calls an external function or API as part of generating its response. The foundation of agentic AI, turns the model from a text generator into a system that can act.

V

  • Vector database , A database optimised for storing and searching numerical representations (embeddings) of text, images, or other data. Used to give AI systems memory and search over large document collections.