AI Terms Explained

The percentage of AI outputs that are correct or acceptable according to a defined standard — measured against a validation dataset.

The percentage of intended users who actively and consistently use an AI system — the ultimate measure of whether an AI program succeeded.

AI systems capable of taking sequences of autonomous actions toward a goal — including planning, tool use, and self-correction.

An AI system that autonomously plans and executes a sequence of actions to achieve a goal — using tools, memory, and reasoning.

Standalone AI tools that help users complete tasks through conversation — ChatGPT, Claude, Gemini, and similar products.

AI tools embedded in existing applications that assist users with tasks — providing suggestions, completions, and automation within familiar interfaces.

The policies, processes, and controls that define how AI is developed, deployed, monitored, and managed within an organisation.

A specific, defined application of AI to a business process — with identified inputs, expected outputs, success criteria, and business value.

Software that performs tasks typically requiring human intelligence — recognising patterns, generating text, making decisions.

The ability to trace, explain, and verify AI decisions — including what inputs were used, what logic was applied, and what output was produced.

Using technology — including AI — to perform tasks with reduced or no human involvement.

AI systems that operate independently to complete complex goals — planning, executing, and adapting without ongoing human input.

Systematic unfairness in AI outputs caused by patterns in training data that reflect historical inequalities or human prejudice.

The structured approach to transitioning people from current ways of working to AI-augmented processes — addressing awareness, skills, and motivation.

An AI task that assigns inputs to one of a set of predefined categories.

Adherence to laws, regulations, and standards governing how AI systems can collect, use, and process data.

All the information available to the AI when generating a response — including instructions, documents, history, and retrieved data.

The maximum amount of text (measured in tokens) an AI model can process in a single request — its "working memory" for that session.

The price charged per token by an AI provider — typically quoted per million tokens, varying by model.

The trade-off between AI model capability and cost — the goal is the cheapest model that meets the required performance threshold.

The process of identifying and correcting errors, inconsistencies, and gaps in data to make it suitable for AI use.

The unintended exposure of sensitive data — through AI model outputs, training data, or system integration vulnerabilities.

The process of defining how data elements from source systems correspond to fields in the AI pipeline or target system.

The automated flow of data from source systems through transformation steps to the AI model and back to destination systems.

The degree to which data is accurate, complete, consistent, and current enough to be useful for AI applications.

The plan-act-observe-adjust cycle that AI agents use to complete tasks iteratively, revising their approach based on results.

An AI task that evaluates inputs against defined criteria and produces a recommended or automated decision.

A type of machine learning using neural networks with many layers that can learn complex patterns from large amounts of data.

The spectrum from AI outputs that are always the same (deterministic) to outputs that vary (variable) — controlled by the temperature parameter.

Mathematical representations of text (or other content) as numerical vectors that capture semantic meaning — enabling similarity search.

An AI task that pulls specific pieces of information from unstructured text — names, dates, amounts, clauses, entities.

Additional training of an existing pre-trained model on a specific dataset to improve its performance on a particular task or domain.

An AI task that creates new content — text, code, images, or other outputs — based on instructions and context.

AI that creates new content — text, images, audio, code, video — rather than just classifying or analysing existing content.

Constraints and controls built into AI systems to prevent harmful, inappropriate, or unintended outputs.

When an AI model generates confidently stated but factually incorrect information — making things up.

A design pattern where humans review, approve, or override AI decisions — especially for high-stakes or low-confidence outputs.

Tokens in the content you send to the model — your prompt, context, documents, and conversation history.

A curated collection of documents, policies, or information used as the source for RAG-based AI systems.

An AI model trained on vast amounts of text that can generate, summarise, classify, and reason about language at scale.

The time delay between sending a request to an AI model and receiving its response.

The multi-dimensional mathematical space in which a model represents and organises concepts internally.

A subset of AI where systems learn from data to improve their performance on a task without being explicitly programmed for every scenario.

Information stored externally and loaded into context at the start of each session — enabling AI to "remember" across conversations.

Information that persists only within a single conversation or session — lost when the session ends.

The trained AI system that processes inputs and generates outputs — the engine behind every AI product or feature.

The gradual decline in AI model performance over time as real-world data patterns shift away from what the model was trained on.

The process of exposing a model to large amounts of data so it learns patterns, relationships, and capabilities.

An architecture where multiple specialised AI agents work together — each handling a specific task, coordinated by an orchestration layer.

AI systems that can process and generate multiple types of input/output — text, images, audio, video — in a single model.

A computational system loosely inspired by the human brain — layers of interconnected nodes that process information and learn patterns.

The coordination layer that manages multiple AI models, tools, and steps in a complex workflow — routing tasks and combining results.

Tokens in the content generated by the model — its response, analysis, or output.

When a model performs extremely well on training data but poorly on new, real-world data — it has memorised rather than learned.

The numerical values within a model that are adjusted during training — the "settings" that encode everything the model has learned.

The instruction you give an AI model — including role, context, task, format, and constraints.

The practice of designing, testing, and refining prompts to consistently produce high-quality outputs from AI models.

The practice of developing and deploying AI systems that are fair, transparent, accountable, safe, and compliant with applicable regulations.

An architecture that combines document retrieval with language model generation — the AI retrieves relevant information before generating a response.

Return on Investment for AI projects — the financial value delivered relative to the cost of implementation and operation.

The ability of an AI system to maintain performance and cost-efficiency as volume increases.

An AI task that identifies the emotional tone of text — typically positive, negative, neutral, or more nuanced emotional states.

Data organised in a defined format with consistent fields — typically in tables, spreadsheets, or databases.

AI output in a specific, machine-readable format — such as JSON, XML, or a defined template — rather than free-form prose.

An AI task that condenses longer content into a shorter, coherent representation while preserving key information.

Standing instructions given to an AI at the start of every session — defining its role, behaviour, constraints, and tone.

The number of tokens or requests an AI system can process per unit of time — a measure of system capacity and speed.

The basic unit of text that AI models read and generate — approximately 4 characters or ¾ of a word. The basis of AI pricing.

Maximising the quality of AI output while minimising the number of tokens used — through prompt compression, caching, and smart retrieval.

The process of breaking text into tokens — the units the model reads and generates, which are the basis for pricing.

The ability of an AI model to call external functions, APIs, or services during its reasoning process to retrieve information or take action.

The set of tools available to an AI agent — defining what actions it can take in the world.

Ready-to-use AI products that require minimal technical setup — subscription-based tools for immediate business use.

Data with no predefined format — text, emails, PDFs, audio, images, videos.

The specific input from a user in a given session — the question, document, or task provided to the AI at runtime.

A database designed to store and search embeddings — enabling fast similarity search across large collections of documents or data.

The sequence of steps — human and AI — that together complete a business process, with defined inputs, outputs, and decision points.