Glossary
AI terms for business decisions.
AI Agent
A system that can interpret a goal, use tools and take bounded actions with controls.
Agentic AI
AI systems designed around goal-directed actions, tool use and feedback loops.
LLM
A large language model that predicts and generates text or other content from context.
RAG
Retrieval-augmented generation, where a model answers using retrieved business documents.
MCP
Model Context Protocol, a standard pattern for connecting AI systems to tools and context.
Embedding
A numerical representation of text, images or other content used for search and similarity matching.
Vector Database
A database designed to store embeddings so related information can be found by meaning rather than exact keywords.
Context Window
The amount of information a model can consider in one request.
Tool Calling
A pattern where an AI system requests a defined tool action, such as searching a database or preparing a draft.
Function Calling
A structured form of tool calling where the model returns arguments for a predefined function.
Fine-Tuning
Additional model training for a specific style, task or dataset, used only when simpler methods are not enough.
Inference
The moment a model produces an output from an input prompt and context.
Guardrail
A control that limits, checks or routes AI behaviour.
Human-in-the-loop
A process where people review or approve AI output before important actions.
Hallucination
A confident AI output that is unsupported, inaccurate or invented.
Multimodal AI
AI that can work with more than one type of input, such as text, images, audio or video.
AI Governance
The roles, policies and controls used to manage AI use, risk and accountability.
Prompt Injection
An attack or failure mode where hostile or hidden instructions try to change what an AI system does.
Structured Output
AI output returned in a predictable format, such as JSON, so software can validate and use it.
Knowledge Assistant
An assistant that answers questions from approved internal documents or systems.
Model Selection
The process of choosing a model based on task quality, cost, risk, data handling and integration needs.
Data Minimisation
Using only the information required for a task, reducing unnecessary exposure of sensitive data.
Evaluation
Testing AI outputs against expected quality, safety and business criteria before relying on them.
Audit Trail
A record of inputs, outputs, tool calls and approvals so AI-assisted work can be reviewed.
Deterministic Automation
Software automation that follows fixed rules rather than model judgement.