AI glossary.

A model trained on very large amounts of text to predict the next token, which lets it write, summarise, translate, reason over documents and follow instructions. ChatGPT, Claude and Gemini are built on LLMs. Go deeper

The unit a language model reads and writes, usually a word fragment. In English a token averages about three quarters of a word. API prices are quoted per million input and output tokens. Go deeper

The maximum amount of text, measured in tokens, a model can consider in one request, including your instructions, documents and its own answer. Go deeper

The instructions and material you give a model. Good business prompts state the task, the audience, the source material, the format and what to do when information is missing. Go deeper

Standing instructions set by the builder of an AI system that apply to every conversation, such as tone, scope and rules the model must follow.

A confident output that is false or unsupported by the source material. It is a property of how language models generate text, so important outputs need checking against sources. Go deeper

A model that spends extra computation working through a problem before answering. Usually more accurate on multi-step tasks, and slower and more expensive per request.

A model that can take more than one kind of input, such as text, images, audio or video, and sometimes produce more than one kind of output. Go deeper

A model whose trained weights are published so it can be run on your own hardware or a cloud of your choice, rather than only through the developer’s API. Go deeper

Further training a model on your own examples to change its behaviour or style. Often less useful for business knowledge than retrieval, which keeps facts in documents you control. Go deeper

A list of numbers that represents the meaning of a piece of text, so similar meanings sit close together. Embeddings power semantic search and retrieval.

Running a trained model to get an output. Inference cost and speed, not training, are what most businesses pay for.

A system in which a model decides which steps to take and which tools to use to reach a goal, observing results and adjusting, rather than following a fixed script. Go deeper

AI that acts, not just answers: it plans, calls tools, reads results and takes further steps, usually within limits set by people. Go deeper

A model’s ability to request that software run a defined function, such as searching a database or creating a draft invoice, with structured arguments. Also called function calling. Go deeper

An open standard for connecting AI applications to tools and data sources through MCP servers. It is maintained under the Linux Foundation’s Agentic AI Foundation. Go deeper

Several AI agents with different roles working together, such as a planner, workers and a reviewer. Useful for some complex tasks, but harder to test than one well-designed agent. Go deeper

A design where a person reviews, approves or corrects AI output before it takes effect. The standard starting point for anything that reaches customers or changes records. Go deeper

Checks around an AI system that block or route unsafe, off-topic or malformed inputs and outputs, such as validation, permission limits and escalation rules.

An AI agent that reads and edits code, runs commands and tests, and proposes changes, such as Claude Code or Codex. Go deeper

An engineer who works inside a client’s operating team, on its systems and data, building until a system creates value in production. Go deeper

Software that moves work through fixed steps automatically. AI adds judgement steps, such as reading a document, inside an otherwise predictable process. Go deeper

A pattern where the system first retrieves relevant passages from your documents and then has the model answer from them, ideally with citations. It keeps answers grounded in sources you control. Go deeper

A database built to store embeddings and find the most similar ones quickly. Often part of a retrieval system. Go deeper

Splitting documents into passages for retrieval. Chunk size and boundaries strongly affect whether the right passage is found.

Combining keyword search with semantic search, which usually finds the right passage more reliably than either alone.

A second step that reorders retrieved passages by relevance before the model sees them.

Tying a model’s answer to specific source material, so every claim can be traced and checked.

Model output in a defined machine-readable format, such as JSON matching a schema, so software can use it reliably. Go deeper

Optical character recognition: turning images of text into text. Vision models now handle many documents OCR struggles with, at a different cost and error profile. Go deeper

Where data is stored and processed. As of October 2026 the major AI model providers do not process AI requests inside New Zealand. Go deeper

A repeatable test of an AI system on representative cases with a defined measure of success, run whenever the prompt, model or data changes. Go deeper

A set of real, representative cases with known correct answers, used to evaluate a system and catch regressions.

Using a language model to grade outputs against a rubric. Useful at scale, but the judge itself must be checked against human grades.

A standard public test for comparing models or systems. Useful for shortlisting, but no substitute for testing on your own work. Go deeper

How well a model’s stated confidence matches how often it is right. A well-calibrated 70 percent prediction is right about 70 percent of the time. Go deeper

Precision is the share of what a system returned that was correct; recall is the share of everything correct that it found. Most AI workflows trade one against the other.

Logging and measuring what an AI system does in production, including inputs, outputs, tool calls, cost, latency and quality, so problems are found quickly. Go deeper

Reusing the processed form of a repeated prompt prefix across requests, which providers bill at a fraction of normal input prices. Go deeper

A way to submit many requests for processing within hours rather than seconds, typically at half the normal price.

The policies, roles and controls an organisation uses to decide where AI is used, who is accountable and how risks are managed. Go deeper

A short, practical document telling staff which AI tools they may use, for what, with which information, and how outputs are checked. Go deeper

New Zealand’s main privacy law. It applies to personal information you put into AI tools and sets out information privacy principles on collection, use, accuracy, security and disclosure. Go deeper

The rules at the heart of the Privacy Act 2020. Since 1 May 2026 there are 14, numbered IPP 1 to 13 with IPP 3A inserted.

The principle in force since 1 May 2026 requiring reasonable steps to tell people when their personal information is collected from someone other than them, such as a third-party data source.

The principle limiting disclosure of personal information overseas. A provider that only processes data on your behalf is generally not a disclosure under section 11 of the Act. Go deeper

An attack where instructions hidden in content the AI reads, such as an email or web page, try to make it ignore its rules or leak data. Go deeper

A widely used list of the most significant security risks for applications built on language models, published by the OWASP Foundation. Go deeper

A voluntary framework from the United States National Institute of Standards and Technology for identifying and managing AI risks.

An international standard for AI management systems, setting requirements for governing the development and use of AI in an organisation.

A commitment released by Stats NZ in July 2020 that signatory government agencies use algorithms transparently and accountably.

The principle that Māori data should be subject to Māori governance, set out in frameworks such as Te Mana Raraunga’s principles. Relevant whenever AI systems use data about Māori people, land or taonga.

A structured look at whether a workflow, its data, systems and people are ready for AI, ending in a ranked, costed recommendation. Go deeper

A small, fast build to test whether an idea is technically feasible. It should answer one risky question, not prove the whole business case. Go deeper

A limited production trial with real users and a success measure agreed in advance. Go deeper

A voice agent that answers phone calls, handles common questions, books appointments and passes complex calls to people. Go deeper

For AI work: the value of time saved, errors avoided or revenue gained, minus build and running costs, including the time people spend reviewing output. Go deeper

Everything a system costs over its life: build, model and hosting fees, monitoring, support, review time and change requests. Go deeper

The decision between an off-the-shelf AI product and a custom system. Buy when the workflow is generic; build when it depends on your own data, rules or systems. Go deeper