ProzessX
Academy ↗DEBook a free consultation
AI Glossary

AI terms explained simply.

From prompt to EU AI Act: the key terms for conversations about AI at your company, kept short and free of jargon.

30 minutes, online, free and non-binding.

In the glossary

The ProzessX AI glossary explains 20 terms around AI in the Mittelstand, including RAG, AI agent, hallucination, human in the loop, and the EU AI Act. Each entry is short, factual, and links to pages that go deeper. Published by ProzessX, the AI consultancy from Nagold.

A to Z

The key AI terms.

This glossary explains the terms that come up most often in conversations about AI at a company. The entries are kept deliberately short and listed alphabetically. Where deeper pages exist, you'll find the link right underneath.

Agentic RAG

Agentic RAG extends RAG with an agent. Instead of searching once and answering, the AI plans several steps: it rephrases the search, reads results from different sources, checks whether the answer really addresses the question, and searches further if there are gaps. This helps with questions that can't be resolved with a single passage of text, such as comparisons across multiple contracts or quotes. The cost is higher effort per query and more elaborate safeguards. As with RAG, the answer cites its sources, and a human decides what follows from it.

Agentic RAG in the Deep Dive →

Data Sovereignty

Data sovereignty means: you decide where your data is stored, who may access it, and what it's used for. In AI projects, this mainly concerns whether requests are sent to an external provider and in which country its servers are located. Options include services in the EU, operation in Germany, or a solution on your own hardware. Which level fits depends on the type of data. Personal data and trade secrets require more control than an ad text. We settle this together with you before the project starts.

Responsible AI at ProzessX →

Embedding and Vector Database

An embedding is the representation of a text as a sequence of numbers that captures its meaning. Texts with similar meaning sit close together in this numerical space, even if they use different words. A vector database stores these number sequences and quickly finds the most similar passages for a given question. Together, the two form the search foundation for RAG (see below): your documents are converted into embeddings once, after which the AI finds matching passages without having to reread every document for every question. Where this database runs is a matter of data sovereignty.

EU AI Act

The EU AI Act, the AI Regulation, is Regulation (EU) 2024/1689 and applies directly in all EU member states. It doesn't regulate the technology itself but its use: the higher the risk to health, safety, and fundamental rights, the stricter the requirements. It distinguishes four tiers, from prohibited practices through high-risk AI and transparency obligations to applications with minimal risk. Most obligations fall on providers of AI systems; operators within a company have fewer, but clear, duties. This entry is not a substitute for legal advice.

The EU AI Act for the Mittelstand →

Fine-Tuning

Fine-tuning takes an already trained language model and retrains it further with your own examples so it more reliably hits a particular style, format, or technical language. It changes the model itself and requires well-prepared training examples. New factual knowledge is harder and more expensive to add this way than with RAG, where the AI looks things up in your current documents. For many tasks in the Mittelstand, good prompting and RAG are therefore enough. Fine-tuning only pays off once a task occurs very frequently, is very uniform, and quality isn't reached with the other means. We check this case by case and often recommend simpler means first.

Generative AI

Generative AI creates new content: text, images, audio, program code, or summaries. It learns from very large numbers of examples which patterns belong together and assembles a new answer from them. Language models like the ones behind chat assistants are the best-known case. At a company, they help with drafts for quotes, emails, or minutes. They deliver suggestions, not verified facts, and they sound convincing even when they're wrong. That's why a human review belongs before any consequential use. At ProzessX, the principle is: the AI prepares, the human approves.

Hallucination

A hallucination occurs when a language model invents a statement that sounds plausible but isn't true. This happens because the model strings words together by probability rather than checking whether something is true. Typical cases are invented sources, wrong figures, or references to laws that don't exist as stated. The risk can be reduced, for example through RAG with source citations, clear task framing, and review steps, but not eliminated entirely. That's why any AI output that a decision relies on should be checked by a person. Figures, deadlines, and legal questions in particular are worth checking against the cited source.

Human in the Loop

Human in the loop means that at decisive points in an automated process, a person reviews and approves. The AI gathers information, makes a suggestion, or drafts an answer; the person confirms, changes, or rejects it. The extent is up to you: for routine work, a spot check may be enough; for quotes, contracts, or customer commitments, every output should be approved. This protects against errors, builds trust within the team, and makes it easier to involve the works council. It's at the core of our approach: the AI prepares, the human approves.

AI and the works council →

AI Agent

An AI agent is a system that doesn't just answer a question but completes a task in several steps. It plans the approach, uses tools such as email, calendar, databases, or your CRM, checks intermediate results, and continues until the task is done. One example: the agent reads an inquiry, finds matching documents, drafts a quote, and submits it for approval. The more tools an agent uses, the more important clear boundaries, logging, and human approval before consequential steps become.

AI automation →

AI Phone Assistant

An AI phone assistant takes calls, understands spoken language, and responds in a natural voice. It can log concerns, answer questions about opening hours or services, book appointments, and forward the call to a person when needed. Technically, it combines speech recognition, a language model, and speech output with your calendar or software. It makes the most sense where calls go unanswered outside business hours or during peak times. What matters is a clear notice that the caller is speaking with an AI, and an easy path to a real person.

AI phone assistant with audio samples →

Artificial Intelligence

Artificial Intelligence (AI) is the umbrella term for computer systems that solve tasks that would otherwise require human thought: understanding language, recognizing patterns, preparing decisions, or writing text. The systems common today learn from data instead of having every step fixed in advance. They're powerful at clearly defined tasks, but not all-knowing and not free of errors. For companies, what matters less is the technology than the benefit: where does AI ease the daily workload, and who checks the result? That's exactly where our consulting comes in.

AI consulting at ProzessX →

Large Language Model (LLM)

A large language model (LLM) is the technology behind chat assistants. It was trained on very large amounts of text and calculates, for a given text, which word is likely to come next. This produces answers, summaries, translations, and program code. A language model only knows what's in its training and in the current request. It doesn't know your internal documents unless you provide them, for example via RAG. Models differ in performance, cost, speed, and where they can be run.

MCP (Model Context Protocol)

The Model Context Protocol (MCP) is an open standard through which AI applications talk to tools and data sources. Instead of building a custom connection for every piece of software, an MCP server exposes its functions in a uniform way, for example searching contacts, creating a note, or reading a document. An AI agent can then use these functions. For companies, this means fewer individual integrations and more interchangeability. What remains important is permission management: an agent should only be allowed to do what the person it acts on behalf of would be allowed to do.

On-Premise AI

On-premise means the AI runs on your own hardware or in a self-controlled environment, not at an external service. Your data never leaves your building or your chosen data center location. This is the furthest-reaching form of data sovereignty and can be the deciding factor for sensitive data or works council requirements. In return, hardware, operations, and updates need to be planned for, and open models don't match the quality of the largest cloud models on every task. We place this within a tier model, from the EU cloud to your own hardware.

Responsible AI and the tier model →

Prompt

A prompt is the instruction given to a language model: the question, the task, and everything the model needs to know about it. Good prompts name the role, the goal, the tone, the desired format, and examples where helpful. They clearly separate instruction from material, such as the text that needs summarizing. Teams that share and maintain prompts get more consistent results. In automations, prompts are fixed in the workflow; in chat assistants, the person writes them themselves. Phrasing them well is something you can learn from your own examples, for instance in a workshop.

AI training courses →

Process Automation

Process automation means designing recurring workflows so that software takes them over fully or in part: transferring data, sorting inquiries, creating documents, or coordinating appointments. Classic automation follows fixed rules. With AI, you can also automate processes that include unstructured input, such as freely written emails or phone calls. It always starts with process discovery: which workflow occurs often, is clearly describable, and is worth the effort? The people involved also belong at the table early, from the relevant department to the works council. A small pilot follows, then expansion of what works in daily use.

Guide: process automation in the Mittelstand →

RAG (Retrieval Augmented Generation)

RAG means, roughly, "text generation enriched through search." Before answering, the system searches your own documents for matching passages and feeds them to the language model. The answer is then based on your material rather than the model's general training, and it can cite its sources. This keeps knowledge current without having to retrain the model. RAG lowers the risk of hallucinations but doesn't rule it out. Typical uses are knowledge bases for service, sales, and technical teams, as well as chatbots built on your own documents.

AI chatbot with RAG →

Responsible AI

Responsible AI describes the aim of using AI so that it works traceably, safely, and with consideration for people. In practice, that means: clear accountability, human approval for consequential steps, protection of personal data, transparency toward employees and customers, and compliance with legal requirements like the EU AI Act. For us, Responsible AI isn't an extra page, it's the working foundation. Your data stays in the EU, and we classify every project by risk and data type before it starts.

Responsible AI →

RPA (Robotic Process Automation)

With RPA, software operates program interfaces like a "robot" following fixed rules: it clicks, copies fields, and transfers data between applications. RPA suits stable, rule-based processes with consistent screens where no interface exists. It doesn't understand content and hits its limits as soon as a screen changes or inputs are unstructured. AI can complement RPA, for example by reading an email and preparing the data the robot then enters. Whether RPA, AI, or both is the right fit is something we clarify during process discovery.

Choosing RPA or AI →

Workflow

A workflow is a sequence of steps executed one after another or depending on conditions: for example "inquiry received, check details, follow up on missing information, create draft, person informed." In automation tools, it's built as a chain of building blocks and runs unattended once a trigger occurs. Individual steps can include AI, for example to read text or draft a reply. What matters is error handling and logging: what happens if a step fails, and who can see what ran? In XOS, workflows can be set up and shared as templates.

Note: This glossary gives a general overview and is not legal or investment advice. For legal questions, such as about the EU AI Act or data protection, please consult your own legal counsel.

AI prepares it. People sign off on it.

Tell us which process is costing you the most time right now. We will get back to you within 24 hours with a suggestion for an initial call.

Patrick Sager
Patrick SagerYour contact

By submitting this form you agree to the processing of your data under our privacy policy .