Generative KI: was sie ist, wie sie funktioniert und wo ihre Grenzen liegen

Generative AI refers to models that create new content, such as text, image, code, audio or video, rather than only classifying existing data. They learn patterns from large datasets and predict the most likely next element on that basis. The leap over older AI is that natural-language operation becomes possible.

From technique to value. In just one week your team builds a working prototype for a real process with generative AI. Book a discovery call, or enter your email below.

Artificial intelligence has existed as a field since the 1950s. For a long time the dominant AI classified and predicted, spam filters or credit scoring, deciding about existing inputs without creating anything new. Generative models predate the hype, but only the transformer architecture from 2017 and large language models from 2020 made them useful. The real shift came with usability. Once you could instruct a model in ordinary language, generative AI left the research labs and became a tool for every function. That is why the term is business-relevant today, not because the maths is new, but because using it no longer requires programming.

A generative language model does not understand text as a human does. It splits language into small units called tokens and learns which token is most likely to follow a given sequence. Producing a sentence is a chain of such predictions.

  input (prompt)
   "Summarise this contract ..."
        |
        v
  [ token by token ]    each step picks the
   the      -->  most     most likely next token
   contract -->  likely   from learned patterns
   ...      -->  token
        |
        v
  output = chain of chosen tokens
   (no lookup, no built-in fact check)

The most important line is in brackets at the bottom. The model looks nothing up and checks no facts on its own. It produces the most plausible pattern. That is why it can sound fluent and convincing and still be wrong, which is called hallucination. Anyone deploying generative AI in a company must know this mechanism, because it explains both the strength (language and creativity) and the weakness (no built-in truth). A company brain or a human check closes exactly that gap.

An illustrative model, not a client figure. A team writes 10 proposal texts a week, about 45 minutes each for writing and formatting.

The numbers are chosen for structure, not as a promise. The point is the division of roles. Generative AI delivers the fast draft, the human checks and owns it. The time gain sits in the draft, quality and liability stay with the human, and that division is what makes the use responsible.

The value is broad because almost every industry works with language, image or code. In marketing agencies, our first target market, generative AI sits directly on value creation, because content is the product. In IT and SaaS it speeds development and documentation. In industry and engineering it helps make documentation knowledge accessible and draft quotes. In finance and insurance use is more cautious and more checked, because regulation and liability apply. The common thread is that generative AI produces drafts a human quickly turns into results.

In short, generative AI is the part of AI that creates. Everything else classifies, automates by rules, or builds on top.

It fits tasks with language, drafting, summarising, translation and creative variation, anywhere a human can check the result. It does not fit as the sole source for hard facts, exact calculation, or decisions without human control, because the model sounds plausible but has no built-in truth. The right question is not whether generative AI can, but whether a human can own the result.

The EU AI Act regulates generative AI through transparency duties, such as labelling AI-generated content, and through rules for general-purpose AI models. For deployers Article 4 also applies, requiring AI competence, precisely because generative AI is easy to operate but easy to overestimate. A deliberate, checked use can support compliance but does not guarantee it. The specific classification is for the company to assess, with qualified counsel where needed.

Article 4 scope note: this page explains the technique and how deliberate use can support competence and governance. It is not legal advice and not proof of automatic conformity.

Company brain · AI adoption · AI literacy · AI enablement · AI skills gap matrix

Two ways, depending on where you are.