AI glossary in plain English
Short definitions of the AI terms that come up most in business conversations, written for founders and operations leaders. Each one links to the chapter that explains it in full.
A
- Agentic AI
- Agentic AI is AI that carries out multi-step tasks on its own, deciding what to do next and using other software, instead of only answering a question.
- AGI (artificial general intelligence)
- AGI (artificial general intelligence) is a hypothetical AI that could match or beat people at most kinds of intellectual work. There is no agreed definition.
- AI agent
- An AI agent is software that uses an AI model to carry out a task step by step, choosing its own actions and using other software such as email or a database.
- AI assistant
- An AI assistant is an app that helps people with tasks through conversation, such as writing, summarising and answering questions. ChatGPT is an example.
- AI automation
- AI automation means using AI for the steps of an automated process that need reading or judgement, such as sorting emails or drafting replies.
- AI benchmarks
- AI benchmarks are standard tests used to compare AI models, such as coding problems, maths questions or exams. Labs publish scores when models launch.
- AI bias
- AI bias is when an AI system treats some groups of people unfairly, usually because it learned patterns from biased past data or human writing.
- AI evals
- AI evals are tests that check how well an AI system does a specific job, by running it on cases with known correct answers and scoring the results.
- AI governance
- AI governance is the rules, roles and records an organisation uses to control how AI is used, who is accountable for it and how decisions can be traced.
- AI guardrails
- AI guardrails are the controls that limit what an AI system can say or do, such as rules it is told to follow and software that checks its answers.
- AI hallucination
- AI hallucination is when an AI tool states something false as if it were true, such as a made-up figure, quote, source or policy.
- AI memory
- AI memory is a feature of AI chat apps that saves facts about a user between conversations so later chats can use them. The model itself remembers nothing.
- AI model
- An AI model is the trained core of an AI system: a program that has learned from data to turn an input into an output, such as a question into an answer.
- AI wrapper
- An AI wrapper is a product built on top of another company's AI model, adding its own interface, instructions and features around it.
C
- Chatbot
- A chatbot is software that holds a conversation with people in text or voice. Modern chatbots such as ChatGPT use AI language models to write their replies.
- Computer vision
- Computer vision is the area of AI that lets computers understand images and video, such as reading scanned documents, recognising objects or spotting defects.
- Context engineering
- Context engineering is choosing what information an AI model is given with each request: instructions, documents, search results, past messages and notes.
- Context rot
- Context rot is the drop in an AI model's accuracy as more text is fed into it at once, even when the text is well within the model's limit.
- Context window
- A context window is the maximum amount of text an AI model can take in and work with at one time, measured in tokens.
D
- Deep learning
- Deep learning is a type of machine learning that uses neural networks with many layers. It powers image recognition, speech-to-text and AI chatbots.
E
- Embeddings
- An embedding is a list of numbers that represents the meaning of a piece of text or an image, so items with similar meanings can be found by search.
- Explainable AI
- Explainable AI is the set of methods for showing why an AI system produced a particular result, such as which factors raised or lowered a score.
F
- Few-shot prompting
- Few-shot prompting means including a few worked examples in a prompt so the AI copies their pattern, format and tone. Zero-shot means giving none.
- Fine-tuning
- Fine-tuning is further training an existing AI model on examples so it behaves differently every time, such as writing in a house style or a strict format.
- Foundation model
- A foundation model is a large, general-purpose AI model trained on broad data so it can be adapted to many tasks. GPT, Claude and Gemini are examples.
- Frontier model
- A frontier model is one of the most advanced AI models available at a given time, usually the newest, largest and most expensive model from a leading lab.
- Function calling (tool calling)
- Function calling, or tool calling, lets an AI model use other software, such as a search, a calculator or a database, by requesting it in a set format.
G
- Generative AI
- Generative AI is AI that creates new content, such as text, images, audio, video or code, in response to a request written in plain language.
- Golden dataset
- A golden dataset is a fixed set of real examples with verified correct answers, used to test an AI system's accuracy before launch and after every change.
- GPT
- GPT is the family of AI language models made by OpenAI, the models behind ChatGPT. GPT stands for generative pre-trained transformer.
- Grounding
- Grounding means giving an AI model trusted source material, such as a document or a database record, and having it base its answer on that material.
H
- Human in the loop
- Human in the loop means a person reviews and approves each AI output before it is sent or acted on. The AI drafts and the person decides.
I
- Inference
- Inference is using a trained AI model to produce an answer. Every time someone asks an AI tool a question, the model is running inference.
K
- Knowledge cutoff
- A knowledge cutoff is the date when an AI model's training data ends. The model knows nothing about later events unless it is given that information.
L
- Large language model (LLM)
- A large language model (LLM) is an AI system trained on huge amounts of text to understand and write language. GPT, Claude and Gemini are LLMs.
- LLMOps
- LLMOps is the practice of running AI language-model applications reliably in daily use: tracking versions, testing changes and monitoring cost and quality.
M
- Machine learning
- Machine learning is a way of building software that learns from past examples instead of following rules written by a person.
- MCP (Model Context Protocol)
- MCP (Model Context Protocol) is an open standard that lets AI assistants connect to other software, such as CRMs, databases and file stores, in one common way.
- MCP server
- An MCP server is a small program that lets AI assistants use another system, such as a CRM or a file store, through the Model Context Protocol.
- Model drift
- Model drift is when an AI system's results change or get worse over time even though nobody changed it, because the data or the underlying model shifted.
- Multimodal AI
- Multimodal AI is AI that can work with more than one type of content, such as text, images, audio and video, in the same system.
N
- Natural language processing (NLP)
- Natural language processing (NLP) is the area of AI that helps computers understand and produce human language, such as reading, translating and summarising.
- Neural network
- A neural network is a type of computer program, loosely inspired by the brain, made of connected layers that learn patterns from data. Most modern AI uses one.
O
- Open-source AI
- Open-source AI usually means AI models anyone can download and run. Strictly, it also requires publishing the training code and details of the training data.
- Open-weight model
- An open-weight model is an AI model whose trained files are published for anyone to download and run on their own computers, within its licence terms.
P
- Private LLM
- A private LLM is an AI language model set up so a company's data stays under its control, usually run on the company's own servers or private cloud.
- Prompt engineering
- Prompt engineering is the skill of writing clear instructions for an AI tool so that it gives useful, consistent answers.
- Prompt injection
- Prompt injection is an attack that hides instructions in content an AI reads, such as an email or web page, to make it do something its owner did not intend.
R
- RAG (retrieval-augmented generation)
- RAG (retrieval-augmented generation) is a method where an AI system first searches a set of documents, then writes its answer from the passages it found.
- Reasoning model
- A reasoning model is an AI model that works through a problem step by step before answering. It is slower and costs more, and is better at multi-step problems.
- Responsible AI
- Responsible AI means building and using AI in ways that are fair, safe, transparent, private and accountable, backed by practical controls.
- RLHF (reinforcement learning from human feedback)
- RLHF (reinforcement learning from human feedback) is a training step where people rate an AI model's answers and the model learns to give the kind they prefer.
S
- Semantic search
- Semantic search finds results by meaning rather than exact words, so a search for money back can find a page about refunds.
- Shadow AI
- Shadow AI is employees using AI tools their company has not approved, such as personal ChatGPT accounts, often with company data.
- System prompt
- A system prompt is a set of hidden instructions given to an AI model before every conversation, setting its role, tone, rules and answer format.
T
- Token
- A token is a small piece of text, a short word or part of a longer one, that AI models read and write. AI usage limits and prices are counted in tokens.
- Token pricing
- Token pricing is how AI providers charge: by the amount of text sent to and returned by the model, counted in tokens, with output text costing more.
- Training data
- Training data is the information an AI model learns from, such as text, images or past records. It shapes what the model knows and how it behaves.
V
- Vector database
- A vector database stores text and images as lists of numbers that capture meaning, so it can quickly find the items closest in meaning to a search.
- Vibe coding
- Vibe coding is building software by describing it to an AI in plain language and using the code the AI writes, often without reading it.