All insights
ORBITRA INSIGHTS

AI Glossary for Business: 40+ Terms Explained in Plain English

Plain-English definitions of the AI terms business leaders meet most: LLMs, tokens, RAG, embeddings, AI agents, MCP, prompt injection, the EU AI Act, ROI, and more.

By Orbitra AIPublished 10 min read
AI GlossaryEnterprise AIAI LiteracyAI Governance
AI Glossary for Business: 40+ Terms Explained in Plain English

AI projects come with their own vocabulary, and meetings stall when half the room is unsure what "RAG", "agentic", or "context window" means. This glossary explains the terms business leaders, managers, and project teams meet most often, in plain language and without hype. Where a term deserves more than a paragraph, we link to a full guide.

The short answer

  • Start with five terms: large language model, token, retrieval-augmented generation (RAG), AI agent, and hallucination. Most AI conversations build on them.
  • A chatbot answers; an agent acts. The difference between them decides how much risk, integration, and oversight a project needs.
  • Most business AI does not train models. Companies usually combine existing models with their own data through RAG, tools, and integrations.
  • Governance terms matter as much as technical ones. Data residency, AI literacy, and the EU AI Act shape which designs are allowed.

Foundations

Artificial intelligence (AI)

Software that performs tasks that usually need human judgement, such as understanding language, recognising images, or making predictions. In business today, "AI" mostly means machine-learning systems, and increasingly generative AI.

Machine learning

A way of building software that learns patterns from examples instead of following hand-written rules. A spam filter trained on thousands of labelled emails is a classic example.

Deep learning and neural networks

Machine learning that uses neural networks with many layers. Deep learning powers modern speech recognition, image recognition, and language models.

Generative AI

AI that creates new content, such as text, images, audio, video, or code, rather than only classifying or predicting. Chat assistants and image generators are generative AI.

Large language model (LLM)

A neural network trained on very large amounts of text to predict the next token. At scale, that simple objective produces models that can write, summarise, translate, reason through problems, and follow instructions. GPT, Claude, Gemini, Llama, and Qwen are well-known LLM families.

Foundation model

A large model trained on broad data that can be adapted to many tasks. Most LLMs are foundation models, and businesses build on them instead of training their own from scratch.

Multimodal model

A model that works with more than one type of input or output, for example reading an image and answering a question about it, or transcribing a call and summarising it.

Small language model (SLM)

A language model small enough to run on modest hardware. There is no fixed cut-off; the term usually means models with up to a few billion parameters. Small models are cheaper and faster, and can run on your own servers, which helps with privacy and data residency.

Open-weight model

A model whose trained parameters, or weights, are published so you can download and run it yourself. Open weights are not always open source: the licence may restrict use, and the training data is often not released.

Parameters

The numbers a model learns during training. The parameter count, such as 7B for seven billion, is a rough guide to a model's size and hardware needs, not a direct measure of its quality.

Token

The unit of text a language model reads and writes, often a word or part of a word. Usage-based pricing and context limits are measured in tokens. Turkish and Arabic text often needs more tokens than the same message in English, which raises cost and latency.

Context window

The maximum amount of text, measured in tokens, that a model can consider at once. It includes your instructions, any documents you supply, the conversation so far, and the model's answer.

Training, pre-training, and inference

Training is how a model learns its parameters, and pre-training is the first and most expensive stage, on broad data. Inference is running the trained model to produce an answer. Businesses mostly pay for inference.

Fine-tuning

Further training of an existing model on your own examples to change its behaviour, style, or output format. Fine-tuning teaches a model how to respond; it is usually not the best way to teach it facts that change, which is what retrieval is for.

Hallucination

When a model produces fluent but false or unsupported information, such as an invented policy or citation. Grounding answers in approved sources, showing those sources, and testing with evaluation sets reduce hallucinations, but nothing removes them completely.

Prompt, system prompt, and prompt engineering

A prompt is the input you give a model: the question, instructions, examples, and context. A system prompt is the standing instruction that sets an assistant's role, rules, and tone. Prompt engineering is the practice of writing and testing prompts so the output is reliable.

Building AI systems

Embedding

A list of numbers that represents the meaning of a piece of text or an image, so that similar meanings get similar numbers. Embeddings make semantic search possible, such as finding a passage about refund deadlines when a customer asks about returning an order.

Vector database

A database built to store embeddings and find the most similar ones quickly. It is the usual search layer behind RAG systems.

Retrieval-augmented generation (RAG)

A pattern in which the system first retrieves relevant passages from your own documents, then asks the model to answer using them. RAG keeps answers current and tied to approved sources without retraining the model. See RAG in plain English.

Grounding

Tying a model's answer to a trusted source, such as your knowledge base or a database, and ideally showing that source to the user.

Tool use (function calling)

The ability of a model to request an action from other software, such as looking up an order, querying a database, or opening a ticket. The model chooses the tool and its inputs; your system runs the call and returns the result.

AI agent

A system in which a model works towards a goal by choosing its next steps, calling tools, checking results, and repeating until the task is done, within limits you set. See AI chatbot vs. AI agent.

Agentic workflow

A business process in which AI agents carry out some steps on their own, alongside fixed automation and human approvals. See agentic AI workflow architecture.

Multi-agent system and orchestration

A design in which several specialised agents work together, for example one that researches, one that drafts, and one that checks. Orchestration is the layer that routes work between them, keeps track of state, and enforces rules. See multi-agent orchestration.

Model Context Protocol (MCP)

An open protocol, introduced by Anthropic in November 2024, for connecting AI applications to tools and data sources in a standard way. One MCP server can make a system such as a CRM available to many AI clients. See the MCP enterprise guide.

Agent2Agent protocol (A2A)

An open protocol, announced by Google in April 2025 and now a Linux Foundation project, that lets agents from different vendors discover each other and hand off tasks. MCP connects agents to tools; A2A connects agents to other agents.

Guardrails

Rules and checks around a model that limit what it receives, says, or does, such as blocking personal data, declining off-topic requests, or requiring approval before an action.

Human in the loop

A design in which a person reviews or approves the AI's output or action at defined points, especially for decisions that affect customers, money, or people's rights.

Evaluation (evals)

Systematic testing of an AI system against realistic examples with known good answers, so that quality is measured before launch and after every change.

LLM observability

Recording what an AI system does in production, including prompts, retrieved sources, tool calls, costs, latency, and errors, so that problems can be traced and fixed. See LLM observability with OpenTelemetry.

Prompt injection

An attack in which instructions hidden in user input, or in content the AI reads such as a web page, email, or document, try to override the system's rules. It ranks first in the OWASP Top 10 for LLM applications. See LLM security hardening.

Self-hosted model

A model that runs on infrastructure you control, on your premises or in a private cloud, instead of through a provider's API. It keeps data inside your environment, in exchange for running the hardware yourself. See self-hosted LLMs for enterprises.

Quantisation

Storing a model's parameters with fewer bits, for example 4 bits instead of 16, so that it needs less memory and runs on smaller hardware, usually with a small loss of quality.

Business and governance

AI governance

The policies, roles, and controls that decide which AI systems an organisation uses, for what purpose, with which data, and who is accountable for them.

AI acceptable use policy

A written policy that tells employees which AI tools they may use, what data they may enter, and who reviews the output. See our AI acceptable use policy template.

AI literacy

The knowledge staff need to use AI effectively and safely: what it can and cannot do, how to check its output, and how to handle data. The EU AI Act requires providers and deployers of AI systems to take measures that support it. See corporate AI training that works.

EU AI Act

The European Union's risk-based law on artificial intelligence, Regulation (EU) 2024/1689. It applies in stages and also reaches companies outside the EU whose AI is used in the EU. See the EU AI Act for companies in Türkiye.

KVKK

Türkiye's Personal Data Protection Law, Law No. 6698. It governs how personal data is processed, including data entered into AI tools and transferred to providers abroad. See KVKK and AI.

PDPL

Short for Personal Data Protection Law, the name used for the data protection laws of Saudi Arabia and the UAE, among others. Both affect where an AI assistant may send and store conversation data. See Saudi and UAE data protection for AI assistants.

Data residency

The requirement or preference that data is stored and processed in a particular country or region. It shapes which models and cloud regions a company can use.

Total cost of ownership (TCO)

The full cost of an AI system over its life: build and integration, model usage or hosting, monitoring, maintenance, and the human review it still needs. See what an AI assistant really costs.

Return on investment (ROI) and realisation rate

ROI compares the value an AI system creates with its total cost. The realisation rate is the share of freed time that actually becomes savings or valuable work, and it is the assumption that moves ROI most. Try it with your own numbers in the AI ROI calculator.

Generative engine optimisation (GEO)

Making content easy for AI search tools, such as ChatGPT search and Google's AI features, to find, understand, and cite. See GEO for companies.

Frequently asked questions

What is the difference between AI, machine learning, and generative AI?

AI is the broad goal of software that performs tasks needing human judgement. Machine learning is the main way to build it today: learning patterns from examples. Generative AI is the part of machine learning that creates new content, such as text or images.

What is the difference between RAG and fine-tuning?

RAG gives a model the right information at the moment it answers, by retrieving it from your documents. Fine-tuning changes the model itself through extra training. Use RAG for facts that change, such as prices and policies; consider fine-tuning for a consistent style or output format.

Which AI terms should managers learn first?

Large language model, token, context window, RAG, AI agent, hallucination, and human in the loop. Together they explain what modern AI systems can do, what they cost, and where they need oversight.

Next steps

A shared vocabulary is the first step towards good AI decisions. Orbitra runs corporate AI training that builds this vocabulary on your own processes, in English, Turkish, or Arabic, and helps teams turn it into working systems through AI automation consulting.

LET'S GET STARTED
PROJECT INTAKE

Let's put your AI automationinto orbit

Tell us the process, customer journey, or team capability you want to improve. We will map the right next step for AI automation consulting, an AI assistant, corporate training, or a measurable pilot.

Email Us[email protected]For inquiries, strategy, and demos
ORBITRA / PROJECT BRIEFSECURE CHANNEL
01
02
03
04