Reference
AI Glossary: Plain-English Definitions
The AI industry produces jargon faster than anyone can absorb it, and a lot of it is used loosely by people selling things. This glossary defines 52 of the terms you are most likely to encounter as a UK business, in plain English, written by people who build AI systems rather than market them. Where a term maps to something we actually do, we have linked to it.
A
Agentic AI
AI systems built to pursue objectives across multiple steps rather than answering once and stopping. The defining trait is acting, not just responding.
AI Agent
Software given a goal rather than a question. It plans the steps needed, carries them out using tools and systems, checks its own results and adjusts, often completing a task without a person approving each step.
AI Governance
The policies, checks and accountability that keep a business’s AI use safe, lawful and under control: what the AI may do, what data it may touch, and who answers for it when something goes wrong.
AI Integration
Connecting AI to the systems a business already runs so it can use real data and take action, rather than operating in isolation in a chat window.
AI Literacy
A working understanding across a workforce of what AI can and cannot do, and how to use it sensibly and safely. Usually the highest-value first training investment for a business.
Algorithmic Bias
When an AI system produces systematically unfair outcomes for particular groups, usually because patterns in its training data reflected existing inequities. A central concern in responsible AI.
API
Application Programming Interface. The standardised door one piece of software offers so other software can talk to it. Most AI integrations work by connecting through an API.
Artificial Intelligence (AI)
Software that performs tasks normally requiring human intelligence, such as understanding language, recognising patterns, or making judgements. In business use today it usually means machine learning systems, particularly large language models.
Automation
Having a task carried out by software rather than a person. Traditional automation follows fixed rules; AI automation can also handle steps that need a degree of judgement.
B
Business Process Automation (BPA)
Automating an entire business process end to end rather than a single task, typically spanning several systems and several steps.
C
Chatbot
Software that holds a conversation with a person, answering questions and responding to requests. Distinguished from an AI agent by being conversational rather than task-completing.
Churn Prediction
Using historical data to identify which customers are likely to stop buying or cancel, ideally early enough to intervene.
Classification
A machine learning task where the system sorts inputs into defined categories, such as routing an enquiry to the right department or flagging a transaction as suspicious.
Computer Vision
AI that interprets images and video: reading documents from photographs, detecting defects on a production line, or recognising objects in a scene.
Context Window
The amount of text an AI model can consider at once. A larger context window means the model can work with longer documents in a single pass without losing earlier detail.
D
Decision Intelligence
Turning data and forecasts into specific recommendations at the moment a decision is made, rather than leaving a person to interpret a dashboard.
Deep Learning
A branch of machine learning using layered neural networks. It underpins most modern AI capability, including language models and computer vision.
Demand Forecasting
Predicting how much of something will be needed, and when, so stock, staff and capacity can be planned around what is coming rather than what happened last year.
E
Embedding
A numerical representation of text or other content that captures its meaning, allowing software to find genuinely related information rather than just matching keywords. The mechanism behind most AI search over company documents.
EU AI Act
The European Union’s comprehensive law regulating artificial intelligence, sorting AI uses into four risk tiers with obligations that scale accordingly. It can apply to UK businesses whose AI reaches people in the EU.
Explainability
How clearly the reasoning behind an AI system’s output can be understood and communicated. Important wherever decisions affect people and must be justified to a regulator or a customer.
F
Fine-tuning
Further training an existing AI model on your own data so it performs better on your specific tasks. Often unnecessary; grounding a model in your documents usually achieves more for less.
G
General-Purpose AI (GPAI)
AI models capable of a wide range of tasks rather than one narrow function. A specific regulated category under the EU AI Act, with transparency obligations for providers.
Generative AI
AI that produces new content, text, images, code, rather than only classifying or predicting. The technology behind tools like ChatGPT and Claude.
Grounding
Connecting an AI model to your own documents and data so it answers from your information rather than general knowledge. The single most important factor in whether a business AI tool is actually useful.
Guardrails
Deliberate limits on what an AI system may do or say, built in during development. Distinct from policy, which governs how people use AI; guardrails constrain the system itself.
H
Hallucination
When an AI model produces confident, plausible output that is simply wrong. The defining failure mode of generative AI, and the reason grounding and human checks matter.
High-Risk AI System
Under the EU AI Act, AI used in areas such as recruitment decisions, credit scoring, critical infrastructure or medical devices, carrying the heaviest compliance obligations.
Human in the Loop
A design where a person reviews or approves an AI system’s output before it takes effect. Standard practice wherever a mistake would be costly or hard to reverse.
I
Inference
The act of an AI model producing an output from an input. Most AI running costs are inference costs, charged per use, which is why usage volume drives ongoing spend.
ISO 42001
The international management standard for artificial intelligence, increasingly asked about during client and partner due diligence.
L
Large Language Model (LLM)
An AI model trained on very large amounts of text, able to understand and produce language. The technology behind most current business AI tools.
M
Machine Learning (ML)
Software that improves at a task by finding patterns in data rather than following explicitly written rules. The foundation of nearly all modern AI.
Model Context Protocol (MCP)
An open standard for connecting AI models to external tools and data sources in a consistent way, reducing the custom work each new integration requires.
Multimodal AI
AI that works with more than one type of input, such as text and images together. Useful for tasks like reading a scanned form or interpreting a photograph alongside a written description.
N
Natural Language Processing (NLP)
The field concerned with software understanding and working with human language. Large language models are the current dominant approach.
O
Orchestration
Coordinating multiple AI steps, tools and systems into a single reliable workflow, including handling what happens when one part fails.
P
Predictive Analytics
Using historical data to forecast what is likely to happen next, so decisions can be made with foresight rather than hindsight.
Predictive Maintenance
Using equipment data to predict failures before they occur, converting unplanned downtime into scheduled work.
Prompt
The instruction given to an AI model. How a prompt is written materially affects the quality of the result, which is why prompting is a genuine, teachable skill.
Prompt Engineering
The practice of writing instructions that reliably get useful results from AI models, including structuring the request and supplying the right context.
Prompt Injection
An attack where hidden instructions in content an AI reads cause it to behave in ways its owner did not intend. A security risk specific to AI systems that traditional controls do not catch.
R
Responsible AI
Building and using AI in ways that are fair, transparent and accountable, going beyond the minimum a regulation requires.
Retrieval-Augmented Generation (RAG)
A technique where an AI model retrieves relevant information from your own documents before answering, so responses are based on your material rather than general knowledge.
Robotic Process Automation (RPA)
Software that mimics human interaction with computer interfaces, clicking and typing through a process. Older than AI automation and more brittle, but useful where no API exists.
S
Shadow AI
Staff using AI tools that the business has not approved or does not know about. Common, and a frequent source of data-governance risk.
Structured Data
Information organised in a consistent, machine-readable format, such as a database or spreadsheet, as opposed to free text, emails or documents.
Supervised Learning
Training a model on examples where the correct answer is already known, so it learns to produce that answer for new cases.
T
Token
The unit AI models process text in, roughly a word fragment. Pricing is usually per token, so longer documents cost more to process than short questions.
Training Data
The information a model learns from. Its quality and coverage largely determine what the model is good at, and where its biases and blind spots lie.
U
Unstructured Data
Information without a consistent machine-readable format: emails, documents, PDFs, recordings. Most business information is unstructured, which is why AI that can read it is valuable.
W
Workflow Automation
Automating a sequence of connected steps across systems and people, rather than a single isolated task.
Frequently asked
Questions about AI terminology
Machine learning is a branch of artificial intelligence. AI is the broad idea of software performing tasks that normally require human intelligence; machine learning is the specific approach of achieving that by finding patterns in data rather than following written rules. Nearly all AI in business use today is machine learning.
A large language model is an AI model trained on very large amounts of text, able to understand and produce language. LLMs are the technology behind tools like ChatGPT and Claude, and behind most business AI applications involving reading, writing or answering questions.
Hallucination is when an AI model produces output that is confident, plausible and wrong. It is the defining failure mode of generative AI, and the main reason business systems need grounding in real data and a way for people to check important outputs.
Retrieval-augmented generation is a technique where an AI model looks up relevant information from your own documents before answering. It matters because it is the difference between a tool that answers from general internet knowledge and one that answers from your actual policies, records and documentation.
Models process text in units called tokens, roughly word fragments, and providers charge by how many are processed. This is why running costs scale with usage and why long documents cost more to handle than short questions. It is worth estimating before committing to a build.
Rather just ask someone?
Definitions only get you so far. Book a free consultation and ask whatever you want about AI in your business, in plain English, with no obligation and no jargon.