AI at LEXchat

This page is the disclosure for the «swiss made software plus AI» label. The contractual commitment on our use of AI is set out in the AI declaration, the processing of personal data in the privacy policy.

LEXchat carries the «swiss made software plus AI» label. The software is developed in Switzerland. AI is not an add-on feature for us but core technology: without language models there is no initial analysis. Our own work lies in everything that surrounds the model.

The label applies to LEXchat Law with the product lines Kanzlei.Recherche and Kanzlei.Pro.

Hosting

LEXchat is a SaaS platform. Application, database and mandate data are held at Infomaniak in Switzerland. The legal corpus consists of officially published statutes and court decisions. It runs on our own infrastructure at Hetzner in Germany; a move to Infomaniak is planned in the medium term.

Integration

We use existing language models and build our own legal processing chain on top of them. We first pseudonymise the facts of the case. We then search our own corpus of federal law, cantonal law and decisions, both semantically and in full text, merge the hits and check the citations against the sources. Only on that basis does the model formulate the analysis (retrieval-augmented generation).

AI used

Generative AI with large language models (LLMs) as well as embedding models for semantic search.

Model origin and type

We work with models from several external providers; we do not operate foundation models of our own. In Switzerland we use open-weight models (including Qwen, Mistral and Gemma), computed at Infomaniak. They handle pseudonymisation, the embeddings for search and, if the law firm selects the Switzerland region, the analysis as well. If the firm selects the EU or USA region, closed-source models from Anthropic (Claude) and OpenAI (GPT) are added, directly or via Amazon Bedrock. Those models receive the facts only in pseudonymised form: names and direct identifiers are replaced beforehand in Switzerland.

Customisation

We do not train models and do not fine-tune them. We steer model behaviour through the choice of legal sources, through prompts and through downstream verification steps. Customer data is not used for training, neither by us nor by the providers. We keep evaluating new models and switch when a model delivers clearly better results.

Which provider processes in which country, and how long it retains data, is set out in our AI declaration and in Annex B of the data processing agreement.