AI Terms, Without Jargon

When comparing AI offerings, you will find three words for the same thing in three proposals — and the same word for three different things. This is not due to bad faith; the market is younger than its vocabulary. However, six terms are sufficient to categorize the majority of what is currently offered. If you can distinguish them, you can break down any proposal into three questions: Where does the knowledge come from, who makes the decision, and who verifies the result?

The Six Terms at a Glance

Language Model (LLM, Large Language Model)

A program that has learned to process and generate language from very large amounts of text. It possesses general knowledge and language patterns — it does not know your price list, your processes, or your customer history. Anything it should know about your company must be provided to it in a controlled manner at runtime.

In practice, this means: Choosing the model is rarely the most important decision in an AI project. The model is the interchangeable part — the connection to your data and the control over it are not. A proposal in which the model name plays the leading role is dealing with a side issue.

RAG (Retrieval-Augmented Generation)

The method by which a language model actually knows something about your company. Before every answer, the system searches your documents, databases, or ticket histories for the relevant sections and presents them to the model along with the question. The answer is thus based on your content — and can cite its source instead of guessing.

Two characteristics make this method attractive for companies: A new document takes effect immediately, without anything having to be retrained. And permissions remain enforceable — anyone who is not allowed to see a file will not receive it in the answer either. However, the latter only applies if the search knows the permissions of the questioner. This is where RAG projects fail, not the model.

AI Agents (Agentic AI)

A language model answers questions. An agent completes tasks: It is allowed to call tools, look up information in systems, execute multiple steps in sequence, and write back results — closing a ticket, creating an order, booking a receipt.

This changes the risk class. A wrong answer is annoying; a wrong booking costs money. An agent therefore needs the same things as a newly hired employee: clearly limited rights, a log of what it has done, and defined points where a human approves. The viable scope is almost always narrower than the demo — an agent for one workflow beats one that is supposed to do everything.

MCP (Model Context Protocol)

An open standard for how AI applications are connected to data and tools: a uniform socket instead of a proprietary adapter per device. A connection to your ERP built once will still work if you change the model or the interface.

The benefit is less technical than contractual in nature. Standardized connections reduce switching costs — they are the countermeasure against a proposal whose integrations only work within the provider’s platform.

Fine-Tuning (Retraining a Model)

An existing language model is retrained with its own examples until it reliably hits a specific technical language, format, or tone. What it learns in the process is then embedded in the model — with the state of the training.

This is where the most common confusion lies. Fine-Tuning is suitable for style, format, and recurring tasks, not for teaching the model up-to-date company knowledge: Every new price list would require a new training run, and what is once inside cannot be selectively removed again. RAG is responsible for knowledge. In addition, a carefully formulated prompt often achieves the same result — without training costs. Fine-Tuning is the answer if that has been proven insufficient.

Hallucination

This is what it is called when a language model generates an answer that sounds convincing but is wrong or fabricated: a regulation that does not exist, a part number that looks plausible.

The crucial point is the categorization. This is not a defect that disappears with the next model generation, but a property of the technology — a system that generates language also generates language when it does not know the answer. An AI solution becomes reliable not through a better model, but through architecture: Basing answers on documented sources, making sources visible, having critical steps approved by a human, and spot-checking results. This belongs in the design and is not a net that can be cast afterwards.

What Fits Which Concern?

Five Questions for Every AI Proposal

1. Where does the knowledge come from? From your sources (RAG), from the training (Fine-Tuning), or only from the general knowledge of the model? If the answer is “the model knows that,” the connection is missing.
2. Where are the data processed? Which provider, which region, how long stored — and is it trained with your inputs?
3. What is the system allowed to do itself? Only read or also write, in which systems, up to which limit, and at which point does a human approval intervene?
4. How do you notice that an answer was wrong? Without source citations, logs, and spot checks, there is no quality statement, only an impression.
5. What happens when switching providers? Who owns the prompts, knowledge base, and evaluations — and do the connections also work elsewhere?

What Follows From This

The three terms that are most often confused in proposals are RAG, Fine-Tuning, and Agents. In short: RAG brings knowledge, Fine-Tuning brings style, Agents bring action. If you can distinguish them, you can recognize in a few minutes whether a proposal fits your use case — and which of the five questions remains unanswered in it.

How we build a productive solution from this — from the assessment through architecture and data to operations — is on our page for AI Services.