AI Glossary

An agent is a software entity that can observe information, decide what to do, and take action to achieve a goal.

Agentic AI refers to AI systems designed to act rather than just respond: they can plan steps, use tools, retain context, and complete tasks semi-autonomously within defined limits.

AI adoption is the process by which an organisation integrates AI into its daily workflows, tools, and decision-making, from initial experiments through to structured use across multiple areas.

An AI agent is software that uses a language model to reason, plan, and carry out tasks independently, often by using tools or accessing data.

AI maturity describes the level at which an organisation has progressed in its use of AI, from initial experiments to structured, scalable deployment with governance and measurable results.

AI readiness is a measure of how well an organisation is prepared to deploy AI successfully, assessed across data quality, skills, systems, and governance.

Artificial intelligence is the broad field of technology focused on building systems that can perform tasks that normally require human intelligence, such as understanding language, recognising patterns, or making decisions.

An autonomous agent is an AI agent that can operate with minimal human input while following pre-defined rules, permissions, and safety constraints.

A Chief AI Officer (CAIO) is a leadership role responsible for AI strategy, governance, and the responsible and effective use of AI within an organisation.

ChatGPT is a conversational AI product from OpenAI, built on large language models, that lets users carry out tasks such as writing, summarising, analysing, and brainstorming in plain language.

Context injection is the practice of providing additional background information, instructions, or business data to an AI model before it performs a task, in order to improve the relevance and accuracy of the output.

The maximum amount of information, expressed in tokens, that an AI model can take into account at one time when generating a response.

The executive responsible for an organisation's technology infrastructure and technical decisions, including the selection and implementation of AI platforms and systems.

A branch of machine learning that uses neural networks with multiple layers to recognise complex patterns in data such as text, images, or speech.

A working model in which humans and AI agents collaborate, with AI handling repetitive or data-intensive tasks and humans focusing on judgement, client contact, and exceptions.

A technique in which an AI system constrains its responses to specific, supplied documents or sources, reducing hallucinations and improving trustworthiness.

Numerical representations of text or data in which meaning and context are captured as vectors, allowing AI systems to compare, search, and retrieve information based on semantic similarity.

The European Union's legal framework regulating AI systems by classifying them according to risk level and imposing obligations on transparency, safety, human oversight, and accountability.

An automation approach in which workflows are triggered by specific, predefined events in a system, such as a new email, a form submission, or a status change in an application.

The way AI or automation systems detect, manage, and escalate situations that fall outside normal processing rules or expected behaviour.

The ability to understand and explain how an AI system arrived at a particular output or decision, so that people can assess, challenge, or improve that outcome.

The process of further training a pre-trained AI model on a smaller, targeted dataset so that it performs better on a specific domain, task, or desired output style.

A foundation model is a large AI model trained on broad, diverse datasets that serves as a starting point for a wide range of tasks, from text generation to reasoning and classification.

Function calling is a capability of AI models that allows them to invoke predefined software functions in external systems, so they can take actions rather than only generating text.

Gemini is a family of multimodal AI models developed by Google, designed to reason across text, images, audio, video, and code in combination.

GPT stands for Generative Pre-trained Transformer: a type of AI model architecture developed by OpenAI, trained on large volumes of text to understand and generate language.

A hallucination is output from an AI system that sounds convincing and correct but is factually wrong or unsupported by any source.

Human-AI collaboration is a working model in which people and AI systems work complementarily: AI handles speed, repetition, and data processing, while people provide judgement, context, and final accountability.

Human-in-the-loop (HITL) is a design approach in which people provide oversight, approval, or correction at specific moments in an AI workflow, particularly for high-impact decisions or when the system is uncertain.

Inference is the process by which a trained AI model processes new input data and generates output based on what it has learned, such as an answer, a classification, or a decision.

Intelligent automation is the combination of traditional workflow automation with AI capabilities such as language understanding, reasoning, and decision support, making it possible to automate tasks that are too variable or unstructured for rule-based software alone.

A knowledge base is a structured collection of documents, procedures, and company knowledge that AI systems use to generate accurate, source-grounded answers and make consistent decisions.

LangChain is an open-source development framework for building AI applications that connect language models to tools, memory, external sources, and multi-step workflows.

A large language model (LLM) is an AI model trained on very large amounts of text to understand and generate human language.

Machine learning (ML) is a branch of AI in which systems learn patterns from data and improve over time without being explicitly programmed for every situation.

Memory in AI agents refers to the mechanisms that allow an agent to retain information temporarily within a session (short-term memory) or preserve knowledge across multiple sessions (long-term memory).

An AI model is a trained system that performs a specific task, such as generating text, classifying images, or predicting patterns, based on patterns learned from training data.

Model alignment is the practice of training and configuring AI systems so their behaviour stays in line with human values, business rules, and intended goals.

Model bias refers to systematic errors in an AI model's output caused by imbalances or blind spots in the training data.

A multimodal model is an AI model that can process and combine multiple types of input, such as text, images, audio, or video, and generate output based on that combined context.

n8n is an open-source workflow automation platform that lets users connect applications, APIs, and AI models through visual, customisable workflows.

NotebookLM is an AI tool by Google that lets users upload their own documents and notes to ask questions, create summaries, and discover connections within them.

OpenAI is an American AI research and technology company, founded in 2015, known for the GPT model series and the ChatGPT chat product.

The OpenAI API is a developer interface that allows applications and workflows to integrate AI capabilities from OpenAI, such as language generation, reasoning, and image analysis, via standardised HTTP requests.

Orchestration is the coordination of multiple AI agents, tools, and workflows to execute complex tasks in the correct sequence.

Planning is the ability of an AI system to break a goal into smaller steps and determine the sequence of actions needed to reach it.

A prompt is the input or instruction given to an AI model to direct its response or output.

Prompt chaining is the technique of linking multiple prompts together so that the output of one step becomes the input for the next, allowing complex tasks to be handled in manageable, verifiable stages.

Prompt engineering is the practice of systematically designing and refining prompts to improve the quality, accuracy, and consistency of AI output.

A proof of concept (POC) is a small-scale implementation used to test whether an AI idea or approach is feasible and valuable before wider deployment.

A reflection loop is a mechanism by which an AI system reviews its own output, identifies weaknesses, and generates an improved version, resulting in higher-quality final output than a single pass would produce.

Responsible AI is the practice of developing and deploying AI in ways that are ethical, transparent, safe, and accountable, with attention to fairness, privacy, and human oversight.

Retrieval-Augmented Generation (RAG) is an AI approach in which a language model retrieves relevant information from external sources or documents before generating a response, so that the output is grounded in current and specific data.

Robotic Process Automation (RPA) is technology that automates repetitive, rule-based tasks by mimicking human interactions with software, such as clicking, copying, and filling in forms.

Semantic search is a search technology that interprets the meaning of a query rather than matching exact words, returning relevant information even when the phrasing differs from the source documents.

A system prompt is a high-level instruction passed to an AI model before a user conversation begins, defining the model's role, behaviour, and boundaries for that session.

Temperature is a setting in AI models that controls how predictable or varied the generated responses are: a low value produces consistent, focused output, while a high value produces more creative and diverse responses.

A token is the smallest unit of text that an AI model processes: typically a fragment of a word, a whole word, or a punctuation mark. AI models read, generate, and are billed by tokens.

Tool calling is the ability of an AI model to invoke external tools, APIs, or software systems during a task, such as a database, calendar, or accounting package.

Top-p, also known as nucleus sampling, is a setting that controls how many options an AI model considers when generating text: the model restricts itself to the most probable options until their cumulative probability reaches the set threshold.

Training is the process by which an AI model learns patterns through repeated exposure to large datasets, enabling it to perform tasks such as text generation, classification, or reasoning.

A vector database is a specialised database that stores embeddings, numerical representations of text or data, and retrieves them based on meaning and similarity rather than exact matches.

A vector embedding is a numerical representation of text, an image, or other data as a sequence of numbers, where the position in that space captures meaning and relationships to other items.

Workflow automation is the automation of multi-step processes by connecting systems, tools, and triggers so that work moves from one step to the next without manual intervention.