All articles

AI for Swiss Law and Fiduciary Firms Without Exposing Client Data

What professional secrecy and the Swiss FADP require when law firms and fiduciaries use AI, technically and organisationally, and which set-ups fit.

Privacy & Law8 min read

In many law firms and fiduciary offices, AI is already in use, just not officially. Someone has a file note summarised by a private ChatGPT account; someone else pastes paragraphs from a tax assessment into a translation tool. It is convenient, and in most cases it does not sit well with professional secrecy.

This article sets out the duties that apply to Swiss lawyers and fiduciaries (Treuhänder), what they mean in technical terms, and which ways of running AI are realistic for a small firm. It is general information, not legal advice.

Which confidentiality duties apply to whom

Lawyers face two layers. Criminal professional secrecy under Art. 321 of the Swiss Criminal Code covers, among others, lawyers, notaries and auditors, together with their auxiliaries. On top of that, Art. 13 of the Lawyers Act (BGFA) imposes a duty of confidentiality without time limit and requires lawyers to make sure their auxiliaries keep the secret too.

Fiduciaries and tax advisers are not listed in Art. 321 unless they act as auditors. For them, the professional duty of confidentiality in Art. 62 FADP is the relevant provision: it applies to anyone who learns secret personal data in a profession that requires knowledge of such data. Contractual confidentiality from the engagement letter comes on top.

Both groups are also subject to the Federal Act on Data Protection. In November 2023 the Federal Data Protection and Information Commissioner (FDPIC) stated that the revised Act applies directly to AI-based data processing, because it is written in technology-neutral terms.

What this means for an AI tool

Four practical requirements follow from these duties.

The provider must be bound as an auxiliary

Using an external service means disclosing client data to it. The common view is that this is permissible if the provider is carefully selected, instructed and contractually bound to secrecy as an auxiliary. Legal scholarship is not unanimous on this, as a University of Zurich study on lawyers' use of cloud services shows. A free personal account does not meet the bar: there is no contract with the firm, and the terms of use may allow inputs to be used for training.

Foreign authority access must be assessed

The US CLOUD Act allows US authorities to compel US providers to hand over data regardless of where it is stored. That includes Swiss data centres run by US groups. In Switzerland a structured risk assessment has become common practice, for example the sample clause and method for secrecy holders' cloud contracts described by David Rosenthal.

Access rights must carry over into the AI

If only the inheritance team can open the Keller file in your document store, the AI assistant must show the Keller file only to them. Many search tools put everything into one shared index. Inside a firm, that breaks the need-to-know principle.

You must be able to trace what happened

You should be able to show who asked which question and which documents were used to answer it. That helps with access requests under Art. 25 FADP and with internal reviews.

A note on professional bodies: we could not verify a specific AI guideline from the Swiss Bar Association for this article. Ask your cantonal bar association whether current recommendations exist.

The set-ups compared

For firms of 5 to 50 people there are essentially four options.

Set-upWhere are the documents?What leaves the office?Typical fit
Public chatbot (free or individual plan)Pasted into the chatEverything that is typed inNot for client data
Enterprise cloud (e.g. Copilot, ChatGPT Enterprise)With the cloud providerDocuments and questionsPossible with contract, risk assessment and careful configuration
Hybrid: documents local, language model in the cloudOn your own serverThe question and a few short excerptsGood if excerpts stay small and the provider is contractually bound
Fully localOn your own serverNothingMost control, higher initial investment

Enterprise cloud services are not ruled out as such. They do require a proper contract review, an assessment of foreign authority access and correct permission settings. For a small office, that effort is often larger than expected.

Hybrid set-ups keep the document store, the search index and the logs on your network. The cloud language model only receives the question and the three to five passages needed for the answer. The difference to poorly built hybrid tools, which upload whole documents, matters here. More on that in Not all hybrid RAG is equal.

Fully local systems keep every piece of data in the office. They need your own server with a graphics card and a little more set-up work, and in return the question of foreign authority access disappears.

Organisational measures you need with any set-up

Technology alone is not enough. These points belong in every rollout:

  • A written AI policy for all staff: which tools are allowed, which data may go into which tool, and that private accounts are off limits for client work.
  • Training based on real examples of what is allowed. For instance: a VAT return may go into the internal assistant, but not into a public chatbot.
  • An entry in the record of processing activities and a check whether a data protection impact assessment is needed. There is a checklist in Before you roll out AI in your office.
  • A review of engagement letters: some clients, such as banks or insurers, stipulate where their data may be processed.
  • A rule for AI output: any answer that goes into a legal brief or tax return is checked by a professional against its source.

An everyday example

A fiduciary office with twelve staff, let us call it Keller Treuhand AG, wants to speed up year-end work. Staff should be able to ask: "Which depreciation rates did we use for client X last year, and why?"

With a public chatbot they would have to upload the documents. With a local or properly built hybrid assistant, the system searches the firm's own files, finds last year's file note and the correspondence with the tax authority, and shows the answer with a reference to document and page. Anyone without access to client X gets no answer about client X.

Where KADARAG fits

KADARAG is an AI assistant for your own document store that runs on your server. In local mode everything stays in-house, including the language model. In hybrid mode documents, search and logs stay with you, and only the question with a few short excerpts goes to a cloud language model. Existing file permissions are respected, and every answer names its source.

Next step

Take a quick inventory: which AI tools does your team actually use today, and with what data? The answer tells you whether you need a policy first, an internal tool, or both. If you want to see what a local assistant would look like with your files, book an intro call.

See KADARAG answer your own questions

30 minutes, no commitment. We show you how KADARAG works with documents like yours and whether local or hybrid fits.