AI Terms Glossary for Business: 12 Terms That Shape Cost and Decisions

Short answer: About a dozen terms show up in almost every AI proposal: LLM, token, context window, prompt, RAG, embeddings, vector database, fine-tuning, AI agent, MCP, hallucination, and guardrails with evals. Knowing what each means, and which ones drive cost, makes quotes easier to compare.

Key takeaways

Tokens are the billing unit: model APIs are priced per million tokens, and output tokens cost more than input tokens.
RAG usually comes first: connecting a model to company documents is cheaper to update than fine-tuning it.
A bigger context window has limits: Anthropic's documentation notes that accuracy and recall degrade as token count grows.
Evals are the acceptance test: a fixed set of real questions with expected answers shows whether the system works.

Model and cost terms

Term	Plain meaning	Why it matters
LLM (large language model)	A model trained on large amounts of text, such as GPT, Claude or Gemini.	The model choice sets quality, speed and price per request.
Token	The unit a model reads and writes. In English, one token is roughly four characters or three quarters of a word, per OpenAI.	Usage is billed per token; Hebrew text tokenizes differently, so measure it on real documents.
Context window	How much text the model can consider in one request, including its own answer. Current Claude models reach up to 1M tokens.	More context costs more per request and does not guarantee better answers.
Prompt	The instructions and material sent to the model with each request.	A clear prompt is the cheapest quality improvement available.

Knowledge and data terms

Term	Plain meaning	Why it matters
RAG	Retrieval-augmented generation: the system finds relevant passages in company documents and the model answers from them. See how RAG document search works.	Answers stay current when documents change, with no retraining.
Embeddings	Lists of numbers that represent the meaning of a text, so similar texts get similar numbers.	They let search match meaning even when the words differ.
Vector database	A database that stores embeddings and finds the closest ones to a question.	It is the search layer behind most RAG systems, such as a document Q&A agent like DocBrain.
Fine-tuning	Further training a model on examples to change its style or behavior.	Useful for fixed formats; a poor way to teach changing facts. Compare it in RAG vs fine-tuning vs long context.

Action and safety terms

Term	Plain meaning	Why it matters
AI agent	A system where the model decides on steps and calls tools, such as searching a CRM or sending an email. More in how AI agents work.	Agents need permissions, logs and limits on what they may change.
MCP	Model Context Protocol, an open standard Anthropic introduced in November 2024 for connecting AI tools to data and systems. See what MCP is.	One connector can serve several AI tools instead of a custom integration for each.
Hallucination	A confident answer that is wrong or invented.	The main business risk; reduced by RAG, citations and refusing to answer without a source.
Guardrails and evals	Guardrails limit what the system says or does; evals are repeatable tests that score its answers.	They show whether a system is ready to launch and whether a model change broke it.

Term

Plain meaning

Why it matters

AI agent

A system where the model decides on steps and calls tools, such as searching a CRM or sending an email. More in how AI agents work.

Agents need permissions, logs and limits on what they may change.

MCP

Model Context Protocol, an open standard Anthropic introduced in November 2024 for connecting AI tools to data and systems. See what MCP is.

One connector can serve several AI tools instead of a custom integration for each.

Hallucination

A confident answer that is wrong or invented.

The main business risk; reduced by RAG, citations and refusing to answer without a source.

Guardrails and evals

Guardrails limit what the system says or does; evals are repeatable tests that score its answers.

They show whether a system is ready to launch and whether a model change broke it.

Using the glossary in a vendor conversation

Ask where the data comes from (RAG or fine-tuning), how many tokens a typical request uses, which evals define acceptance, and what the agent is allowed to change. The answers explain most of the price difference between quotes, a topic covered in AI development cost in 2026. An accounting firm that wants answers from client files, for example, needs RAG, permissions and evals far more than fine-tuning. For a project like that, see custom AI system development or book a short scoping conversation.

Frequently asked questions

What is the difference between RAG and fine-tuning?

RAG retrieves documents at question time; fine-tuning changes the model through extra training. RAG suits changing facts, fine-tuning suits a fixed style or format.

Why do AI costs depend on tokens?

Model APIs charge per million input and output tokens, so long documents and long answers raise the bill.

Is an AI agent the same as a chatbot?

No. An agent can plan steps and use tools such as a CRM, so it needs permissions and logging.

Can hallucinations be eliminated?

They can be reduced with RAG, citations, refusal rules and evals, but important answers should stay traceable to a source.

AI Terms Glossary for Business: 12 Terms That Shape Cost and Decisions

Key takeaways

Model and cost terms

Knowledge and data terms

Let's Talk About Your Project

Action and safety terms

Using the glossary in a vendor conversation

Frequently asked questions

What is the difference between RAG and fine-tuning?

Why do AI costs depend on tokens?

Is an AI agent the same as a chatbot?

Can hallucinations be eliminated?

More articles that may interest you

RAG vs Fine-Tuning vs Long Context: Which AI Approach Wins in 2026

How Much Does AI Development Cost in 2026? What Drives the Price

AI Agents Explained: How They Work and How to Build One

Tell us what you need

Ask it.