LLM

A language model is not a database. It does not look facts up; it generates text from learned patterns. When something is missing, it phrases an invented answer with the same confidence as a correct one.

In a business it therefore usually becomes useful only once it is connected to your own data: the model does the wording, the facts come from your documents. The technique behind this is called retrieval augmented generation.

It is strong at language: summarising, rephrasing, drafting, classifying, pulling details out of unstructured text. It is weak at anything that has to be exact: arithmetic, current figures, legal advice without a source.

So the question is rarely «can we use an LLM?» but «where may a mistake happen, and who checks it?».

Data protection law applies to AI too. In November 2023 the Federal Data Protection and Information Commissioner stated that the current FADP is directly applicable to AI-supported data processing. Anyone feeding customer data into a model needs to know where it is processed and whether it is used for training.

There is now a Swiss option as well: Apertus, an open language model from EPFL, ETH Zurich and the Swiss National Supercomputing Centre CSCS, released in 2025.

We choose the model by the task and by the data path, not by the headlines. Where personal data is involved, we first establish where the model runs and what it may retain – part of our AI consulting and AI governance work.

And we build so that a person can see where an answer came from. An answer without a source is a draft, not information.