AI Basics and Core Concepts: A Glossary

Part of the Five Stones knowledge base - 51 terms on AI basics and core concepts, in plain English.

Part of the Five Stones knowledge base - 51 terms on ai basics and core concepts, in plain English.

Artificial Intelligence (AI) - Software that performs tasks which normally need human judgement: understanding language, recognising patterns, making a recommendation.
Generative AI - AI that creates new content (text, images, audio, code) rather than only analysing existing content.
Large Language Model (LLM) - The type of AI model behind tools like ChatGPT and Claude, trained on huge amounts of text to understand and generate human-like language.
Foundation model - A large, general-purpose AI model trained once and then adapted for many different tasks, rather than built from scratch each time.
Prompt - The instruction or question you give an AI tool. What you type in is the prompt; what it produces is the output.
Prompt engineering - Deliberately wording a prompt to get a more accurate or useful result from an AI tool.
Fine-tuning - Further training an existing AI model on your own data so its answers better fit your business, instead of using it generically.
Hallucination - When an AI confidently states something false or made up. The single biggest reason AI output needs a human check before it reaches a customer.
Token - The small chunk of text (roughly three-quarters of a word) an AI model reads and generates one at a time; most AI tools are priced by token.
Context window - How much text an AI model can "remember" in one conversation before it starts forgetting the earlier parts.
Inference - The moment an AI model actually produces an answer, as opposed to the earlier process of training it.
Training data - The information an AI model learned from. Its quality and coverage directly shape how good the model's answers are.
Bias (AI bias) - A skew in an AI's output caused by patterns in its training data, which can unfairly favour or disadvantage certain groups or answers.
Multimodal AI - AI that can work with more than one type of input at once, such as text, images and voice together.
General-purpose AI - AI built to handle a wide range of tasks (like ChatGPT), as opposed to AI built for one narrow job (like a fraud-detection model).
Narrow AI - AI designed to do one specific task well, such as reading receipts or scheduling appointments, and nothing beyond that.
AGI (Artificial General Intelligence) - A hypothetical AI with human-level ability across all tasks, not yet achieved by any system in production use.
Neural network - The layered mathematical structure, loosely modelled on the brain, that most modern AI models are built from.
Deep learning - Machine learning using neural networks with many layers, which is what powers most of today's AI advances.
Algorithm - A defined set of steps a computer follows to solve a problem or make a decision; AI models are built and trained using algorithms.
Model - The trained "brain" that an AI product runs on. When people say "which AI model are you using," this is what they mean.
API (Application Programming Interface) - The technical connection point that lets one piece of software (like your booking system) talk to another (like an AI model).
API key - A password-like code that lets your software securely connect to an AI provider's service and be billed for usage.
SaaS (Software as a Service) - Software you access online for a subscription, rather than installing and owning it outright, e.g. most AI tools today.
Open-source AI - AI models whose underlying code (and sometimes training data) is published publicly for anyone to use, inspect or modify.
Closed-source AI - AI models controlled entirely by one company, accessed only through their product or paid API, with the underlying code kept private.
On-premise AI - AI software run on your own servers or computers rather than a third party's cloud, usually chosen for data control or compliance reasons.
Cloud AI - AI software run on a provider's remote servers and accessed over the internet, which is how most SME AI tools work today.
Edge AI - AI that runs directly on a local device (a phone, a camera, a till) instead of sending data to the cloud, usually for speed or offline use.
AI model provider - The company that built and trains the underlying AI model, such as OpenAI, Anthropic or Google, as distinct from the company that builds a product on top of it.
AI wrapper - A product built on top of someone else's AI model with an added interface or workflow, rather than a model built from scratch.
Zero-shot - Asking an AI to do a task it has never specifically been trained or shown examples for, relying purely on its general knowledge.
Few-shot - Giving an AI a small number of examples in your prompt so it understands the pattern you want before it answers.
Explainability - How well an AI's decision can be understood and justified in plain terms, which matters most in regulated or high-stakes decisions.
AI ethics - The set of principles guiding whether an AI use case is fair, transparent and safe for the people it affects.
Synthetic data - Artificially generated data used to train or test an AI model when real data is scarce, sensitive, or expensive to collect.
Latency - The delay between asking an AI a question and getting an answer. High latency is what makes a chatbot feel slow.
Throughput - How many AI requests a system can process in a given time, which matters when many customers use a tool at once.
Sandbox environment - A safe, separate space to test an AI tool or automation before it touches real customer data or goes live.
Proof of concept (POC) - A small, low-cost test build used to prove an AI idea actually works before committing budget to the full version.
MVP (Minimum Viable Product) - The smallest working version of a tool or system, built to test an idea in the real world before investing further.
Vendor lock-in - Being so dependent on one AI provider's specific tools and format that switching to another provider later becomes costly or difficult.
AI literacy - A basic working understanding of what AI can and cannot do, increasingly treated as a core business skill rather than a technical one.
Human-in-the-loop - A system design where a human reviews or approves an AI's output before it takes effect, rather than the AI acting fully on its own.
AI readiness - How prepared a business's data, processes and staff are to actually adopt and benefit from AI, as distinct from simply buying an AI tool.
Digital transformation - The broader shift of a business's operations, tools and culture onto digital and often AI-assisted systems, of which AI adoption is one part.
AI strategy - A business's deliberate plan for where and how it will adopt AI, as opposed to adopting tools one at a time with no overall direction.
AI use case - A specific, defined problem an AI tool is applied to (e.g. "answering after-hours enquiries"), the right starting point before evaluating any tool.
AI maturity model - A framework describing the stages a business typically passes through, from no AI use to fully embedded AI, used to judge where a business currently sits.
Shadow IT - Technology or software adopted by staff without formal approval from whoever manages the business's systems, a common precursor to Shadow AI.
Technology stack (tech stack) - The full set of software and tools a business runs on, which any new AI tool needs to fit into rather than fight against.

Looking for a different topic? Browse the full knowledge base, or tell us what's missing.