AI Glossary

An AI system that can autonomously perform tasks, make decisions, and take actions to achieve specific goals.

AI systems capable of independent reasoning, planning, and executing multi-step tasks without constant human guidance.

A person who uses AI tools extensively and skilfully in their daily work, staying current with new models and techniques to consistently get strong results.

Overhyped AI systems that claim to predict inherently unpredictable outcomes like crime, job performance, or market movements.

Automatic Speech Recognition: technology that converts spoken audio into text, powering voice assistants and transcription services.

Application Programming Interface: a set of rules that allows different software applications to communicate with each other.

The use of technology to perform tasks with minimal human intervention, improving efficiency and reducing errors.

Away From Keyboard: letting an AI agent work through a task on its own, without someone watching each step. AFK work suits well-defined jobs with strong automated checks, and is riskier where mistakes are costly or hard to spot.

The rules that decide what an agent may do on its own and what needs human approval first, such as editing files, spending money, or sending messages. Setting them well is central to using agents safely.

Systematic errors in AI outputs caused by imbalanced or unrepresentative training data, leading to unfair or inaccurate results.

The decision between building custom software for a specific need and buying a ready-made product. The practical rule we use: buy the commodity capabilities that do not differentiate you, and build the specific workflow that is your edge.

Content Delivery Network: a network of servers that delivers web content to users based on their geographic location.

An AI-powered conversational interface that can understand and respond to user queries in natural language.

AI technology that enables computers to interpret and understand visual information from images and videos.

The practice of strategically managing an LLM's context window to optimise AI responses. This includes selecting, organising, and formatting the information provided to the model for best results.

The maximum amount of text an AI model can process at once, measured in tokens. Larger windows allow for more context-aware responses.

Customer Relationship Management: software that helps businesses manage interactions with customers and prospects.

The information a model has in front of it for a task right now: your prompt, the conversation so far, and any documents or data fed in. A model reasons from its context plus what it learned in training, so getting the right information in front of it is most of the work. See also Context Window and Context Engineering.

Domain Name System: the internet's phonebook that translates domain names into IP addresses.

A class of digital systems, including blockchain, where records are stored and synchronised across many independent computers rather than a single central authority, making the shared ledger transparent and tamper-resistant.

A numerical representation of text that captures its meaning, used for semantic search and similarity matching.

Teaching an AI to perform a task by providing only a few examples in the prompt, without additional training.

The process of training a pre-existing AI model on specific data to improve its performance for particular tasks.

Generative Engine Optimisation: the practice of optimising content to appear in AI-generated search responses.

When an AI generates confident but factually incorrect or fabricated information that sounds plausible.

Human-in-the-Loop: a design pattern where humans are involved in AI decision-making to provide oversight and handle edge cases.

Passing the context and state of a piece of work from one session, agent, or person to the next so it can continue without starting over. A clear written summary is the usual way to carry the important detail across.

The process of using a trained AI model to generate predictions or outputs from new input data.

The tokens you send to a model on a request, including your prompt, instructions, and any supplied context. Providers bill for input tokens, so longer prompts and larger context cost more.

The date beyond which a model has no built-in knowledge, set when its training data was collected. Ask about events after the cutoff and the model will not know, unless you supply the information or connect it to live tools.

Large Language Model: AI systems trained on vast amounts of text data that can understand and generate human-like text.

Model Context Protocol: an open standard for securely connecting AI assistants to business data and tools.

A trained AI system that processes inputs and generates outputs based on patterns learned from training data.

AI systems that can process and generate multiple types of content including text, images, audio, and video.

Minimum Viable Product: the smallest version of a product that delivers real value and tests your riskiest assumption with actual users. It is deliberately incomplete and, above all, shipped, because a prototype that never reaches a customer teaches you nothing.

A system that lets an AI keep useful information across separate sessions, such as your preferences or project details, and reload it later. It works around the fact that models are stateless by saving notes outside the conversation.

The company that builds and runs a model and sells access to it, such as OpenAI, Anthropic, or Google. Most AI products are built on top of one or more providers rather than a model the business trains itself.

A computing system inspired by biological brains, consisting of interconnected nodes that process information in layers.

Natural Language Processing: AI technology that enables computers to understand, interpret, and generate human language.

The core mechanism behind most AI text generation. The model reads the text so far, predicts the most likely next token, adds it, then repeats, producing fluent output one small piece at a time.

The property that the same prompt can produce different answers on different runs. It explains why AI output is hard to reproduce exactly, and why testing and review matter for anything important.

The tokens a model generates in response to a request. Providers usually charge more for output than input, which is why concise answers can be cheaper to produce at scale.

Product Requirements Document: a written specification of what a piece of software should do, for whom, and why, including what is explicitly out of scope. It turns a vague idea into a buildable plan and is the main tool for directing an AI build.

The practice of writing effective LLM prompts to achieve desired outputs. Includes techniques for structuring instructions, providing examples, and guiding AI behaviour.

A security vulnerability where malicious inputs trick an AI into ignoring its instructions or performing unintended actions.

The internal values a model adjusts during training, often numbering in the billions. They store the patterns the model has learned, and a higher count is a rough, imperfect proxy for capability.

Building a rough, throwaway version of an idea to test it quickly before committing to a proper build. AI makes prototyping fast, which is useful for learning, as long as the result is treated as disposable rather than production ready.

Retrieval-Augmented Generation: an AI technique that combines information retrieval with text generation for more accurate responses.

Return on Investment: a measure of whether the value a project produces exceeds what it cost. For an AI project, value is best counted as outcomes like hours saved, cost avoided, and revenue influenced, measured against a baseline taken before the work began.

Software as a Service: cloud-based software accessed via subscription rather than traditional installation.

Structured data code that helps search engines understand your website content better.

Search Engine Optimisation: the practice of improving website visibility in traditional search engine results.

Security protocols that encrypt data transmitted between websites and users, indicated by HTTPS in URLs.

An isolated environment where an AI agent can run code or commands without touching real systems or data. It contains mistakes and limits damage, which makes giving an agent more freedom safer.

One continuous run of work with an AI assistant, from the first message until you clear it or start fresh. Within a session the assistant can see earlier messages; across separate sessions it cannot, unless a memory system carries information over.

A model holds no memory of its own between requests. Each call is answered only from what is in front of it, so any continuity (the earlier conversation or saved notes) has to be passed back every time.

A secondary agent that a main agent starts to handle a specific sub-task and report back. Splitting work this way keeps each agent focused and stops one long task from overwhelming a single context window.

An AI's tendency to agree with the user and give flattering or expected answers rather than accurate ones. It is a side effect of training models to produce responses people rate highly, and a reason to ask for evidence rather than reassurance.

A set of standing instructions placed before a conversation that defines an assistant's role, rules, and tone. It is how you steer behaviour consistently without repeating yourself in every message.

The future cost of choosing a quick or easy solution now instead of a sounder one. Like financial debt it accrues interest: shortcuts taken early are repaid later as slower changes and more bugs. A little is a sensible trade to ship, too much grinds a project to a halt.

A parameter that controls AI output randomness. Lower values produce more predictable responses, higher values more creative ones.

A unit of text that AI models process, typically representing a word or part of a word. Pricing is often based on token usage.

The dataset used to teach an AI model patterns and relationships, fundamentally shaping its capabilities and limitations.

The neural network architecture behind modern AI models like GPT and Claude, enabling parallel processing of text sequences.

Text-to-Speech: technology that converts written text into natural-sounding spoken audio, enabling voice interfaces and accessibility features.

When a model asks to run an external tool, such as a search, a calculation, or an action in another system, and then uses the result. Also called function calling, it is how assistants do things in the world rather than only producing text.

One exchange in a conversation with an assistant: your message plus everything the assistant does in response, including any tool calls, before it hands control back to you.

A database optimised for storing and querying embeddings, enabling fast semantic search across large datasets.

A search technique that finds semantically similar content by comparing numerical representations (vectors) of meaning, rather than matching keywords.

Building software by describing what you want in natural language and letting an AI model generate the code, while the human steers and reviews rather than writing most lines by hand. Powerful for prototypes, and risky for production work without careful review.

An automated message sent from one system to another when a specific event occurs, enabling real-time data transfer.

An AI's ability to perform tasks it wasn't explicitly trained for, using only its general knowledge and the task description.