AI glossary of terms

Agent (AI Agent)

A system that uses an AI model to plan steps, use tools and carry out a task with some independence, rather than just answering a single question. Learn more

Agentic AI

AI that can act on a goal instead of only responding to a prompt. It plans steps, uses tools and adjusts as it goes, often with little human input. Learn more

Agentic Workflow

A business process carried out by an AI agent that decides its own steps toward a goal, instead of following a fixed script every time. Learn more

AI Alignment

The practice of making sure an AI system behaves the way its designers and users actually intend, including in situations it was not directly trained for. Learn more

AI Governance

The policies and oversight a business puts in place to manage how it develops or uses AI responsibly, including who approves new AI use cases and how outputs get checked. Learn more

AI Model

The trained system behind an AI application. It has learned patterns from data and can be given new inputs to produce predictions, text or decisions. Learn more

AI Safety

The field concerned with preventing AI systems from causing unintended harm, from a single system misbehaving through to longer term risks from more capable systems. Learn more

AI Washing

When a company overstates its use of AI in marketing or investor communications to look more innovative than it really is. Learn more

Algorithm

A defined set of steps a computer follows to solve a problem or complete a task. Not every algorithm involves AI, but every AI system runs on algorithms. Learn more

API (Application Programming Interface)

A defined way for one piece of software to talk to another. Most businesses use AI by calling a provider’s API rather than building a model themselves. Learn more

Artificial General Intelligence (AGI)

A hypothetical AI system able to understand, learn and apply knowledge across almost any task at a human level. AGI does not exist yet. Learn more

Artificial Intelligence (AI)

Technology that allows machines to perform tasks that would normally need human intelligence, such as understanding language, recognising images or making decisions. Learn more

Artificial Narrow Intelligence (ANI)

AI built to do one thing well, such as recommend products or detect fraud, rather than reason generally like a human. Every AI system in use today is narrow AI. Learn more

Artificial Superintelligence (ASI)

A theoretical AI that would exceed human ability across every domain, including creativity and judgement. It remains speculative. Learn more

Automation

Using technology to carry out a task without manual effort. Not all automation involves AI, but AI has made it possible to automate more judgement based tasks. Learn more

Autonomous Agent

An AI agent that plans and carries out a task without asking for approval at each step, as opposed to one that pauses for a person to check in along the way. Learn more

Bias (Algorithmic Bias)

When an AI system produces skewed results for certain groups, usually because of patterns or gaps in the data it was trained on. Learn more

Big Data

Datasets too large or complex for traditional tools to process. It is often the raw material that AI models are trained or fine tuned on. Learn more

Chain-of-Thought (CoT) Reasoning

A technique where an AI model works through a problem in visible steps before giving its final answer, which tends to improve accuracy on harder tasks. Learn more

Chatbot

Software that holds a text or voice conversation with a user, usually powered today by a large language model rather than a fixed script. Learn more

Compute

Shorthand for the processing power, chips, servers and energy needed to train or run an AI model. Compute cost is one of the biggest factors in any serious AI deployment. Learn more

Constitutional AI

A training method, developed by Anthropic, where a model is guided by a written set of principles that it uses to review and revise its own responses. Learn more

Context Engineering

The practice of deliberately gathering the documents, examples and rules an AI system needs to do a task well, rather than relying on a single clever prompt. Learn more

Context Window

The amount of text an AI model can take into account at once, measured in tokens. Once a conversation grows past this limit, earlier content gets dropped. Learn more

Conversational AI

AI built specifically to understand and hold natural, back and forth dialogue. It is the technology behind chatbots, voice assistants and AI customer support. Learn more

Copilot

An AI assistant that works alongside a person, suggesting a draft, a next step or a line of code, while leaving the person to review and approve it. Learn more

Data Privacy

Protecting personal or sensitive information from being collected, used or exposed inappropriately, a central concern whenever an AI system is trained on or given access to business data. Learn more

Deep Learning

A branch of machine learning that uses neural networks with many layers to find patterns in large amounts of data. It is behind most recent advances in image and language AI. Learn more

Deepfake

Synthetic image, audio or video created or altered by AI to look or sound real, often showing someone doing or saying something they never did. Learn more

Diffusion Model

A type of generative AI model, commonly used for images, that builds its output by starting from random noise and gradually refining it into a clear result. Learn more

Edge AI

Running an AI model directly on a local device, such as a phone or a camera, instead of sending data to the cloud to be processed. Learn more

Embeddings

A way of turning words, images or other data into lists of numbers that capture their meaning, so a computer can measure how similar two things are. Learn more

EU AI Act

The European Union’s AI law, in force since 2024 and applying in stages through 2027. It sorts AI systems by risk and applies to any business whose AI touches the EU market. Learn more

Explainable AI (XAI)

Techniques that make an AI model’s decisions understandable to a person, rather than leaving them as an unexplainable black box. Learn more

Few-shot Learning

Giving an AI model a handful of examples inside the prompt itself, so it can pick up a pattern without any retraining. Learn more

Fine-tuning

Taking an already trained AI model and training it further on a smaller, specific dataset so it performs better on a particular task. Learn more

Foundation Model

A large AI model trained on broad, general data and built to be adapted to many different tasks, rather than one narrow purpose. Learn more

Frontier Model

The most advanced AI models available at any given time. Because the frontier keeps moving, the term describes a moving target rather than one fixed model. Learn more

Function Calling

The mechanism that lets an AI model trigger a specific action in outside software, such as looking up a record or sending an email, by producing a structured request. Learn more

GAN (Generative Adversarial Network)

An earlier generative AI approach where two neural networks compete, one generating fake content and one trying to spot it, with both improving as a result. Learn more

General-Purpose AI (GPAI)

The EU AI Act’s term for AI models capable of a wide range of tasks rather than one narrow function, broadly the large models built by major AI labs. Learn more

Generative AI (GenAI)

AI that creates new content such as text, images, audio or code, rather than only classifying or analysing existing data. Learn more

Guardrails

Rules and checks put around an AI system to keep its behaviour within acceptable limits, such as stopping a chatbot from giving financial advice it is not authorised to give. Learn more

Hallucination

When an AI model generates something that sounds confident and plausible but is factually wrong or made up. It is a normal feature of how these models work, not a rare fault. Learn more

Human-in-the-Loop (HITL)

A system design where a person reviews or approves an AI’s output at one or more points, rather than letting it act fully on its own. Learn more

Inference

The process of an already trained AI model actually being used, taking a new input and producing an output. This is what happens every time someone sends a prompt. Learn more

Inference-time Compute

Extra computing power a model uses while it is answering a question, rather than only during training, in effect letting it think longer before it responds. Learn more

Large Language Model (LLM)

An AI model trained on huge amounts of text to predict and generate language, now used as the basis for summarising, coding, analysis and conversation. Learn more

Latency

The delay between sending a request to an AI system and getting a response back. Even an accurate response can feel broken if it takes too long. Learn more

Machine Learning (ML)

The branch of AI where systems improve at a task by learning patterns from data, rather than following fixed rules written by a person. Learn more

Mixture of Experts (MoE)

A model design that splits its capacity into many smaller specialist sub-networks and only uses the relevant few for any given input, making large models cheaper to run. Learn more

Model Context Protocol (MCP)

An open standard, introduced by Anthropic in late 2024, that gives AI systems a common way to connect to a business’s tools and data instead of needing a custom integration for each one. Learn more

Multi-agent System

A setup where several AI agents, often specialised for different tasks, work together to complete something more complex than any one of them could handle alone. Learn more

Multimodal AI

AI that can work with more than one type of input or output, such as text, images, audio and video, in a single system. Learn more

Natural Language Processing (NLP)

The field of AI focused on helping computers understand, interpret and generate human language. It underpins everything from spam filters to large language models. Learn more

Neural Network

A model design loosely based on the brain’s structure, made of connected layers that adjust as they learn from data. It is the basis of deep learning. Learn more

Orchestration

Coordinating several AI models, agents or tools so they work together on a task, deciding what runs when and passing information between steps. Learn more

Overfitting

When a model learns its training data too closely, including its noise and quirks, so it performs well on that data but poorly on new, real world examples. Learn more

Parameters

The internal values a model adjusts during training to capture patterns in its data. Parameter count is often used as a rough measure of a model’s scale. Learn more

Pre-training

The first, resource heavy stage where a model learns general patterns from a huge, broad dataset, before any fine tuning for a specific task. Learn more

Prompt

The instruction, question or piece of text given to an AI model to get a response. Everything a person types into an AI tool is a prompt. Learn more

Prompt Engineering

Writing and refining prompts to reliably get better, more accurate output from an AI model, by being specific about format, context and examples. Learn more

Prompt Injection

A security risk where hidden instructions, planted in a document or web page an AI system reads, try to override its original instructions. Learn more

Quantization

A technique that shrinks a model’s size and computing needs by reducing the precision of its internal numbers, with a small trade off in accuracy. Learn more

RAG (Retrieval-Augmented Generation)

A technique where an AI model looks up relevant information from a set of documents before generating its answer, rather than relying only on what it learned during training. Learn more

Reasoning Model

An AI model built to work through a problem in explicit steps before answering, trading extra time for better performance on complex maths, coding and logic tasks. Learn more

Reinforcement Learning

A training approach where a model learns by trial and error, getting a reward for good outcomes and a penalty for bad ones. Learn more

Reinforcement Learning from Human Feedback (RLHF)

A version of reinforcement learning where human raters’ preferences between different AI outputs are used to guide a model’s later training. Learn more

Responsible AI

A general term for developing and using AI in ways that are fair, transparent, safe and accountable. Learn more

Small Language Model (SLM)

A language model built with far fewer parameters than a frontier model, trading some general capability for lower cost, faster responses and the ability to run on smaller hardware. Learn more

Sovereign AI

A country or organisation’s push to build and run AI models and infrastructure on its own soil and under its own control, rather than depending entirely on foreign providers. Learn more

Supervised Learning

Training a model on labelled examples, meaning inputs paired with the correct output, so it learns to predict outputs for new, unlabelled inputs. Learn more

Synthetic Data

Artificially generated data used to train or test an AI model, made to resemble real data without exposing actual personal or sensitive records. Learn more

Temperature

A setting that controls how predictable or varied an AI model’s output is. A low temperature gives more focused answers, a high one gives more varied, creative results. Learn more

Token / Tokenization

The process of breaking text into the small chunks, often pieces of words, that an AI model actually reads and generates. Pricing and limits are usually measured in tokens. Learn more

Tool Use

An AI model’s ability to call outside software, such as a search engine or a database, mid task rather than relying only on what it already knows. Learn more

Training Data

The information a model learns from during training. The scale, quality and diversity of this data has a direct impact on how capable and reliable the model is. Learn more

Transformer

The neural network design, introduced in 2017, that underpins almost all modern large language models, thanks to a mechanism called attention that lets it weigh every word against every other word. Learn more

Unsupervised Learning

Training a model to find structure in data that has no labelled correct answers, such as grouping similar customers together without being told what the groups should be. Learn more

Vector Database

A database built to store and quickly search embeddings, so a system can find the information most relevant to a query by meaning rather than exact keywords. Learn more

Vibe Coding

Writing software by describing what you want in natural language and letting an AI generate the code, with the person guiding by feel and testing the result rather than reading every line. Learn more

Weights (Model Weights)

The specific numerical values inside a neural network, set during training, that determine how it turns an input into an output. Often used as shorthand for the trained model itself. Learn more

Zero-shot Learning

An AI model completing a task correctly despite never having seen a labelled example of that exact task, by generalising from everything else it learned in training. Learn more