AI glossary
AI terms, explained plainly
The important terms around artificial intelligence — explained for decision-makers and teams in mid-sized companies, in practical language and with a link to everyday work.
Fundamentals
Artificial Intelligence (AI)
An umbrella term for systems that carry out tasks which normally require human intelligence — for example, understanding text, recognizing images, or making decisions. In a business context, AI today usually means machine learning and large language models, not the "general AI" of science fiction. The practical benefit lies in automation, pattern recognition, and decision support.
Machine Learning
A branch of AI in which algorithms learn from data instead of being programmed explicitly. An ML model finds patterns in training data and can apply them to new data — for example, to predict customer churn or to classify documents automatically. Quality depends directly on the quality of the data.
Deep Learning
A specialized form of machine learning that uses artificial neural networks with many layers. Deep learning makes it possible to recognize complex patterns — in images, speech, or text. The technology behind most modern AI breakthroughs, from speech recognition to image generation.
Neural Network
A computational model that imitates biological nerve cells: many simple units (neurons), organized in layers, that communicate through weighted connections. Training adjusts the weights until the network solves the intended task. The basis of all modern deep learning systems.
Natural Language Processing (NLP)
Technologies that let computers understand, analyze, and generate human language. Uses include automatic summaries, sentiment analysis, chatbots, translation, and text classification. Especially relevant for companies that process large volumes of text — emails, contracts, and customer feedback.
Computer Vision
A field of AI that teaches computers to interpret images and video — recognizing objects, classifying scenes, or identifying anomalies on production lines. Used in mid-sized companies for quality control, document processing (OCR), and visual inspection.
Token
The smallest unit into which a language model splits text — roughly a word or part of a word. "Unternehmensstrategie" (corporate strategy), for example, is split into several tokens. This matters in practice because AI services bill by token, and models can process only a maximum number of tokens per request (the context window).
Supervised Learning
A learning method in which the model is trained on labeled data — that is, input–output pairs. Example: emails marked as "spam" or "not spam." The model learns the mapping and can classify new emails. The most widely used approach, and the easiest to understand.
Unsupervised Learning
A learning method without labeled data. The model finds structures and patterns on its own. Typical applications: customer segmentation, anomaly detection, or topic clustering in large volumes of text. Useful when no pre-classified training data is available.
Reinforcement Learning
A learning method in which an agent learns by trial and error to maximize rewards. Used when there are no "correct" training examples, but there is an evaluation function — for example, in robot control or in the optimization of language models (RLHF). Also the basis of the impressive results in chess and Go.
Transfer Learning
The principle of reusing a model pretrained on a large dataset for a more specific task. Instead of starting from scratch, the model's existing knowledge is used as a base and adapted only for the use case at hand. This saves a large amount of compute time and training data. The underlying principle of fine-tuning.
Models & Architecture
Large Language Model (LLM)
Large language models such as GPT-4, Claude, or Llama, trained on very large volumes of text. LLMs can understand, generate, summarize, and translate text. They are the basis of most current AI applications in companies — from chatbots and document analysis to code generation.
Transformer
The architecture behind all modern language models (GPT, Claude, Llama). The decisive innovation is the "attention mechanism," which lets the model recognize relationships across long passages of text. Introduced by Google in 2017, and since then the basis of virtually all progress in NLP.
Embedding
A mathematical representation of text (or images, or audio) as a vector of numbers. Similar content sits close together in vector space — "Geschäftsführer" (managing director) and "CEO" have similar embeddings. The basis for semantic search, recommendation systems, and retrieval augmented generation (RAG).
Vector Database
A specialized database that stores embeddings and can search efficiently for similar vectors. It makes it possible to find the most relevant documents, product descriptions, or FAQ entries in fractions of a second — by meaning, not by keywords. A core component of modern RAG systems.
Practice & Implementation
Prompt
The input (an instruction, a question, or context) sent to a language model. The quality of the prompt largely determines the quality of the answer. In business use, prompts are often standardized as templates so that results stay consistent across different users and use cases.
Prompt Engineering
The craft and method of writing prompts so that language models deliver optimal results. It includes techniques such as role instructions, structured output formats, chain-of-thought prompting, and few-shot examples. For companies, it is often the fastest way to improve the quality of AI results substantially — without changing the model.
Chain of Thought (CoT)
A prompting technique in which a language model is asked to think step by step before it gives an answer. It improves result quality considerably on complex tasks — for example, analyses, calculations, or multi-step decisions. Easy to apply: "Think step by step" in the prompt is often enough.
Few-Shot Learning
A way of showing a language model what you expect by including a few examples in the prompt. Instead of a long explanation, you provide 2–5 examples of input and the desired output. Especially useful for standardized tasks such as classification, formatting, or data extraction.
Fine-Tuning
Further training of a pretrained model on your own data, to optimize it for a specific use case. More costly and more involved than prompt engineering, but necessary when the base model needs to master domain-specific knowledge or a particular writing style permanently. It requires high-quality training data.
Retrieval Augmented Generation (RAG)
An architecture pattern in which a language model retrieves relevant documents from a knowledge base before it generates an answer. It addresses two central problems: the model can access current, company-specific information, and hallucinations are reduced because answers are based on real sources.
Hallucination
When an AI model generates information that sounds plausible but is factually wrong — invented citations, laws that do not exist, or incorrect figures. A core risk when companies use LLMs. Countermeasures: RAG, fact-checking, structured outputs, and human-in-the-loop processes.
Inference
The process in which a trained AI model takes an input and produces an output — production use, as distinct from training. Inference costs (compute time, API fees) are a significant factor when assessing the economics of an AI solution.
AI Agent
An AI system that plans subtasks on its own, carries them out, and evaluates the results — beyond a single request to a model. AI agents can use tools (web search, APIs, databases), assess intermediate results, and adjust their approach. Currently one of the most dynamic areas of development in AI.
Agentic Engineering
The discipline of embedding AI agents systematically in work processes — not as passive assistants, but as actively guided members of the team. It covers task decomposition, context management, quality assurance, and the orchestration of several agents. The difference from simply using a tool: methodical guidance rather than ad hoc use.
MLOps
Practices and tools for running machine learning models in production — analogous to DevOps for software. It covers versioning of data and models, automated training, monitoring of model quality, and deployment pipelines. It becomes relevant once AI models are in continuous use, not only in experiments.
Governance & Law
EU AI Act
The European AI regulation — the world's first comprehensive AI law. It classifies AI systems by risk level and sets requirements for transparency, documentation, and human oversight. Relevant for companies: certain AI applications (for example, in HR or lending) are subject to strict obligations. It has been entering into force in stages since 2024.
AI Governance
The rules, processes, and responsibilities for the responsible use of AI in a company. It covers data quality, bias checks, transparency of decisions, access controls, and compliance with the EU AI Act. Not a paperwork exercise — it is the basis that lets AI projects scale without creating risk.
AI Maturity
An assessment of how far a company has come in adopting AI — from "first experiments" to "AI anchored in core processes." Important for strategy planning: measures that suit an AI-mature company will overwhelm a company still in the early stage. An honest assessment of the current position avoids misallocated resources.
AI Strategy
A structured plan for how a company intends to put AI to profitable use — not "a little bit of everything," but prioritized by impact and feasibility. A good AI strategy connects business goals with technical possibilities, takes existing data and skills into account, and defines concrete next steps.
Model Sovereignty
A company's ability to keep freedom of choice over AI models — and not depend on a single provider. It includes strategies such as multi-provider approaches, the use of open-source models, and clear API abstractions. Especially relevant in the European context, with a view to data protection and geopolitical risk.
AI terms in practice?
Understanding is the first step — let’s talk about how these ideas apply in your company.