Reference 2026
RAG, hallucination, context window, agent, MCP — anyone working with AI in their business in 2026 quickly hits a jungle of jargon. This glossary explains the most important terms without hype and without unnecessary technical depth: a clear definition for each, plus one sentence on why the term matters to you as a decision-maker. Designed as a reference you can return to again and again — and as honest orientation for your process automation.
What this is about
Most failed AI investments come not from poor technology but from misunderstandings about what the terms actually mean. Anyone who can place "hallucination", "context window" or "RAG" asks the right questions in vendor conversations, spots inflated promises and knows where the real costs and risks sit.
This glossary is grouped by theme: first the fundamentals, then the area that matters most to companies, how AI accesses your own data safely (RAG), followed by operations, cost and security, and finally law and governance. You don't have to read it all in one go — jump straight to the term you're dealing with right now.
Part 1
These six terms form the foundation. Once you understand them, you already understand most of what sits behind modern AI products.
A "Large Language Model" is a model trained on enormous amounts of text that understands and generates language itself. It is the engine behind chatbots, text analysis and agents.
Why it matters for you: Almost every AI feature you buy runs on such a model — the choice of model drives quality, cost and data protection.
Inference is the moment when the fully trained model actually answers. Every single request consumes compute — and therefore the ongoing running cost of AI.
Why it matters for you: Unlike classic software, you pay per operation. At high volume, inference becomes the decisive cost factor.
A hallucination is a fabricated but convincing-sounding statement. The model doesn't "know" it is wrong — it simply phrases things plausibly.
Why it matters for you: This is the main reason AI results must be checked on important decisions and a human stays in the loop.
The context window is the amount of text a model can hold "in mind" at once per request. Whatever doesn't fit is not considered.
Why it matters for you: Long contracts or entire knowledge bases don't fit in one go — which is why techniques like RAG are needed to make large data volumes usable.
An AI agent is more than a chatbot: it interprets a task, breaks it into steps, calls tools and makes decisions, rather than just producing text.
Why it matters for you: Agents pay off where understanding and decision-making are required — exactly when is something we explore in our AI agents vs. Zapier & RPA guide.
Agentic AI describes systems of several collaborating agents that handle multi-step tasks largely on their own — ideally with oversight at the critical points.
Why it matters for you: This is the next stage of automation. A practical primer for companies is in the article Agentic AI in the enterprise.
Part 2
For an AI to work with your knowledge instead of just general training knowledge, it needs "Retrieval Augmented Generation" — RAG for short. This group of terms is especially relevant for companies, because this is where data protection, accuracy and verifiability come from. A detailed, vivid explanation is available in RAG explained simply.
RAG connects the language model to your own knowledge base: before answering, relevant documents are retrieved and passed along, so the AI answers from sourced facts.
Why it matters for you: RAG cuts hallucinations sharply and keeps your data in-house. More on this: RAG explained simply.
Hybrid RAG combines classic keyword search with meaning-based (semantic) search, to find both exact terms and related content.
Why it matters for you: Higher match quality, especially with technical terms and part numbers. Details in RAG explained simply.
GraphRAG uses a knowledge graph that maps terms and their relationships. This lets the AI answer questions that connect several documents.
Why it matters for you: Useful with interconnected knowledge (customers, contracts, projects). Context in RAG explained simply.
Corrective RAG checks the retrieved documents for relevance and searches again when matches are weak, before answering — a built-in self-checking step.
Why it matters for you: Fewer wrong answers when sources are patchy. More in RAG explained simply.
Multimodal RAG includes not only text but also images, tables, diagrams or scans in the search and the answer.
Why it matters for you: Important when your knowledge sits in plans, photos or scanned documents. See RAG explained simply.
Embeddings translate text into numerical vectors that represent meaning. Texts with similar content then sit close together — the basis of semantic search.
Why it matters for you: They are the reason an AI finds "related" content even when the exact same word doesn't appear.
A vector database stores embeddings and instantly finds the most meaning-related content for any question. It is the memory of a RAG system.
Why it matters for you: Where this database lives (EU, your own server) co-determines your data protection.
Chunking splits large documents into meaningful sections so the AI can find the right part precisely instead of searching a whole document.
Why it matters for you: Good chunking is a major factor in whether answers turn out precise or vague.
Reranking re-sorts the search results in a second, more accurate step so the truly most relevant passages sit right at the top.
Why it matters for you: An often underestimated lever — it noticeably improves answer quality without switching the model.
Part 3
Getting an AI into production means measuring quality, limiting behavior, controlling cost and defending against attacks. These terms decide whether a promising pilot ever runs reliably.
Evals are systematic tests that measure how correct and reliable an AI system's answers are — comparable to quality control in manufacturing.
Why it matters for you: Without evals, AI quality stays a gut feeling. They are the prerequisite for proving any improvement at all.
Guardrails are technical and organizational rules that define what an AI may and may not do — such as topic limits, approval steps or filters for sensitive data.
Why it matters for you: They are your safety net against unwanted or risky AI outputs in customer-facing situations.
Quantization shrinks a model by slightly reducing the precision of its internal numbers. The result: a smaller, cheaper model with barely noticeable quality loss.
Why it matters for you: Quantized models often run on affordable hardware — enabling operation on your own or European servers.
MCP is an open standard through which AI agents connect to external tools and data sources in a standardized way — essentially a universal plug for AI.
Why it matters for you: MCP reduces expensive custom integrations. Explained in depth in the article MCP servers explained.
Prompt injection is an attack in which hidden instructions inside texts, emails or web pages trick an AI into unwanted behavior.
Why it matters for you: As soon as an agent acts on its own, this becomes a real security risk — background in Prompt injection & AI security.
Shadow AI is the unapproved use of AI tools by employees — without the IT department's knowledge or sign-off. Sensitive data often ends up at external providers unnoticed.
Why it matters for you: A widespread, underestimated risk. How to manage it is shown in Shadow AI & governance.
Part 4
In Europe, AI is not a lawless space. Anyone deploying AI with personal or business-critical data should know the two most important frameworks — they help decide what is allowed in the first place.
The GDPR governs the protection of personal data, the EU AI Act the use of AI by risk class. Together they form the binding legal framework for AI in Europe.
Why it matters for you: Legal basis, data minimization and logging are mandatory — not an afterthought. Explained practically in GDPR-compliant AI in the enterprise.
Governance bundles the rules, responsibilities and controls for AI use. "Human in the loop" means that consequential decisions are signed off by a person.
Why it matters for you: Governance is the prerequisite for a pilot to even become productive and auditable.
Keep reading
This glossary grows with the topic. If you want to know more about an area, our guides offer detailed, practical explanations:
How AI works safely with your own data is explained in RAG explained simply. The legal framework for the mid-market is covered in GDPR-compliant AI in the enterprise. The often-overlooked risk of unapproved tools is addressed in Shadow AI & governance. And which technology category fits your process — from no-code through low-code and RPA to AI agents — is laid out in our overview of process automation technologies.
A complete overview of all articles is in our automation guide. Missing a term, or want it applied to your specific case? Just get in touch — we explain things in plain language and vendor-independently.
Frequently Asked Questions