AI TERMINOLOGY FOR DECISION-MAKERS
A practical guide to the language used in enterprise AI — what each term means, where it creates value and how we use it in client work.
Explore the terms ↓OUR RULE
RAG may solve a knowledge problem. An agent may execute several steps. MCP may connect tools. MoE may influence model economics. None of these is a business outcome by itself.
We use terminology to make architecture, risk and cost choices explicit — then translate them into what your team will actually do differently.
THE WORKING DICTIONARY
Each definition includes our practical use. That distinction prevents a technical capability from being mistaken for a finished workflow.
Plain EnglishAI that can pursue a goal through several steps, use tools, inspect results and adjust its next action.
How we use itFor bounded workflows with clear permissions, stop conditions, human approval points and evidence of what happened.
Plain EnglishA system that combines a model with instructions, context, tools and an execution loop.
How we use itWe define the job, allowed actions, owner, evaluation and escalation path before discussing autonomy.
Plain EnglishThe system searches approved information first and gives the relevant material to the model before it produces an answer.
How we use itFor policy, product, client or operational knowledge that must be current, private and traceable to a source.
Plain EnglishA model architecture with multiple specialised internal expert networks. A routing mechanism activates only a subset for each input.
How we use itAs part of model-selection literacy: capacity, latency, quality and cost matter. Most clients do not need to train an MoE model; they need to choose and evaluate the right existing model.
Plain EnglishAn open standard that lets AI applications discover and use external data sources and tools through a consistent interface.
How we use itTo connect approved business systems while making tool boundaries, authentication and permissions visible.
Plain EnglishA model trained to understand and generate language, code and structured information.
How we use itAs one component in a workflow. The surrounding context, tools, evaluation and human decision usually determine business reliability.
Plain EnglishA model that can work across more than text — for example images, audio, documents, video or interface screenshots.
How we use itFor document intake, visual inspection, research and workflows where information arrives in mixed formats.
Plain EnglishNumerical representations that allow systems to compare meaning and retrieve similar content.
How we use itInside search and RAG systems, with testing for relevance, access control and outdated content.
Plain EnglishAdditional model training on selected examples to change behaviour or improve performance on a repeatable task.
How we use itOnly when prompts, context and retrieval cannot produce stable performance and enough quality training examples exist.
Plain EnglishRepeatable tests that measure whether an AI system produces acceptable results under expected and difficult conditions.
How we use itWe turn real cases, failure modes and reviewer expectations into a scorecard before scaling.
Plain EnglishGiving the model reliable context so its answer is based on the organisation's actual information.
How we use itThrough approved documents, connected systems, structured instructions and source references.
Plain EnglishA person reviews, changes or approves the system's work at a defined point.
How we use itWhere judgement, accountability, financial impact, people decisions or customer commitments require a named owner.
Plain EnglishTechnical and operational controls that restrict what the system may access, produce or do.
How we use itPermissions, input/output rules, approval gates, budgets, audit records and escalation paths matched to risk.
Plain EnglishThe amount of information a model can consider at one time.
How we use itWe select and structure relevant context instead of assuming that sending more documents will create a better result.
Plain EnglishThe moment a trained model processes an input and generates a result.
How we use itInference quality, latency and cost become operating measures when a workflow runs repeatedly.
Plain EnglishSelecting different models for different tasks based on capability, risk, speed or cost.
How we use itUse the strongest model where judgement is hard and a smaller, cheaper model where work is predictable and testable.
PUTTING THE TERMS TO WORK
Only relevant, permitted information enters the task.
The model follows the reporting method and identifies uncertainty.
It uses authorised tools and stops at the defined approval point.
Quality checks and accountable review determine whether it moves forward.
TECHNOLOGY IN SERVICE OF THE WORK
We will explain what you need, what you do not need and how your team will operate the result.