AI glossary

AI that creates new content: text, images, audio and code. Earlier systems could only sort or analyse data that already existed. A generative model produces something that wasn't there before. It's the technology behind today's AI assistants and most of the business AI tools in use now.

A model trained on huge amounts of text to understand and generate language. It works by predicting the next word over and over, from patterns it picked up in training. Chatbots, copilots and most generative AI tools run on one. The "large" refers to the number of internal parameters, now often in the hundreds of billions.

A software system that plans and carries out multi-step tasks on its own. It reads data, calls tools and adjusts course as it goes, rather than answering a single prompt. Humans set the goals and review the outcomes. This ability to act independently is what "agentic AI" refers to.

The craft of writing instructions that reliably get the output you want from an AI model. Small changes in wording can shift the result a lot, so in production teams version and test prompts like code. It's less about clever tricks than about being specific about the task, the format and the constraints.

The unit a model reads and writes. In English one token is about three-quarters of a word (roughly four characters), though this varies by language. Pricing, speed and context limits are all counted in tokens, so token count is what you actually pay for.

When a model states something false and presents it as fact, with no signal that it's unsure. You reduce it by grounding the model in real data (for example RAG), writing clearer prompts and keeping a human in the loop. It never disappears completely on its own, so it's a risk you manage, not a bug you fix.

The EU regulation that sorts AI systems into risk tiers, each with its own obligations: a banned tier, a high-risk tier with strict duties, a transparency tier, and minimal risk for the rest. Most business uses fall in minimal risk. The duty you cannot skip is knowing where yours lands. In force since 2024, phased through 2027.

The structures that keep AI use controlled and accountable: who decides, what is allowed, how models get monitored across their life. It is what regulators and boards ask about first. Frameworks like NIST's AI RMF and ISO/IEC 42001 give it shape, but governance is people and process before it is any tool.

AI tools employees use for work without the organisation's knowledge or approval, like pasting company data into a public chatbot. It is common, and it quietly exposes the business to data leaks and compliance gaps. IBM's 2025 breach research put the added cost of a shadow-AI incident at around $670,000. Get visibility first, then set the rules.

Technical controls that keep an AI system's behaviour inside agreed limits, checking inputs and outputs and blocking anything unsafe before it reaches a user: a leaked secret, a harmful reply, an off-topic tangent. Dedicated frameworks let teams define these rules in code. Guardrails enforce policy at runtime. An AI policy writes it down.

The organisation's written rules for using AI: which tools are approved, what data must never go into them, when a human has to review the output. It is the first governance document most companies need, and the practical answer to shadow AI. Rules people can actually follow beat an exhaustive policy nobody reads.

The practical discipline of building and running AI that is fair, transparent, robust and accountable. Where AI ethics sets the principles, responsible AI turns them into everyday work: testing for bias, documenting models, keeping a human in oversight. Major AI vendors each frame it around a short set of principles backed by concrete tooling and review.

A way of turning text, images or other data into a list of numbers that captures meaning. Content about similar things ends up with similar numbers, so a machine can measure how close two items are. Embeddings are the maths behind semantic search, recommendations and retrieval-augmented generation.

A database built to store embeddings and find the closest matches fast, instead of looking for exact keyword hits. Ask it for the items most similar in meaning to your query and it returns them from millions in milliseconds. It is the retrieval layer under most semantic search and RAG systems.

The practice of designing everything a model sees at the moment it runs: instructions, retrieved documents, tool definitions, conversation history. As AI moved from single prompts to agents working over many steps, wording one prompt well stopped being enough. It is the discipline prompt engineering grew into. The focus shifted from wording to wiring.

Search that matches on meaning rather than exact words. It turns your query into an embedding and finds the content whose meaning sits closest, so a search for "cut cloud costs" also surfaces a page titled "lowering the monthly server bill". It is why modern search and chatbots find the right answer even when the wording differs.

Tying a model's answers to verified sources such as your own documents, a database or live search, so it draws on real information instead of guessing. Grounded systems can cite where an answer came from, which makes them checkable. It is the main defence against hallucination, and retrieval-augmented generation is the common way to do it.