Banking secrecy and AI in Switzerland: banks, FINMA and asset managers

At a private bank in Geneva or an asset manager in Zurich, AI assistants found their place very quickly: summarising a research note, preparing a portfolio report, drafting a letter to a client in German, French or English. The time saved is real, and teams know it.

But the Swiss financial centre rests on an old promise: discretion. The day an adviser pastes a client statement into a chatbot, that promise is put to the test. Here is what banking secrecy, the Financial Institutions Act and FINMA's expectations mean in practice for the use of AI.

Banking secrecy also applies to negligence

Article 47 of the Banking Act punishes breaches of professional secrecy by a bank's officers, employees, agents or liquidators. An intentional breach carries a custodial sentence of up to three years or a monetary penalty, and up to five years where the person seeks a financial advantage. The point many people forget is in paragraph 2: negligence is punishable by a fine of up to 250,000 francs. The duty continues even after employment ends.

For portfolio managers and fund management companies, Article 69 of the Financial Institutions Act (FinIA) sets out a comparable offence. A careless copy and paste into an external AI tool is therefore not just a procedural slip: it can have criminal consequences for the person who did it.

The AI tool is a service provider like any other

For FINMA, handing processing to an external provider is outsourcing. Circular 2018/3 on outsourcing requires, among other things, careful selection and monitoring of the provider, security and confidentiality of data, and preserved audit rights. An online AI assistant that processes client data comes close to this logic, even if nobody presented it as an outsourcing project.

FINMA made this clearer in its Guidance 08/2024 of 18 December 2024 on governance and risk management when using artificial intelligence. It expects institutions to keep an inventory of their AI applications, classify them by materiality and risk, pay attention to data quality and carry out serious due diligence on solutions supplied by third parties. It also flags growing dependence on a few AI providers as a risk to watch.

Critical data and operational resilience

FINMA Circular 2023/1 on operational risks and resilience for banks has been in force since 1 January 2024. It requires institutions to identify critical data, meaning data that calls for increased security measures, and to protect it throughout its life cycle. FINMA has observed that many institutions limited this concept to client identifying data (CID) and expects a broader approach.

For AI, the consequence is simple: you need to know which critical data can reach a prompt, through which channels, and what happens if it does. A portfolio, an investment strategy or a draft transaction can be as sensitive as a client's name.

The situations that cause problems

At banks and asset managers, the risk is concentrated in a few repeated actions:

• pasting a positions statement to get a performance commentary;

• translating an exchange with a client that contains their name, country of residence and assets;

• summarising a compliance file or a source of funds review;

• asking the AI to analyse a spreadsheet exported from the portfolio management system;

• using a coding assistant with snippets that contain account identifiers.

External people matter too: consultants, developers, IT service providers. Article 47 expressly covers the bank's agents. If a contractor pastes client data into an AI assistant, the risk is the same, and the contract should address it.

Traditional network monitoring tools struggle to see these flows, as we explain in our article on why email and network DLP fail on ChatGPT. And when usage escapes approved tools, the hidden cost climbs fast: see the real cost of shadow AI.

A four step approach

1. Inventory the AI tools actually in use, including browser extensions and personal accounts, as FINMA expects.

2. Classify use cases by the data involved: public, internal, confidential, subject to banking secrecy.

3. Govern the chosen tools by contract, applying outsourcing requirements, and ban the others for client data.

4. Control at the workstation: detect client names, IBANs, account numbers and amounts before sending, then anonymize or block.

These steps do not require a large project. The inventory and classification can progress within a few weeks, as long as compliance, IT security and the business lines are involved from the start.

The fourth step also provides the evidence expected during a review: a log of incidents prevented, useful for internal audit, the external audit firm and FINMA.

Sovereignty and the nFADP, the foundation

These sector rules come on top of the revised Federal Act on Data Protection (nFADP), in force since 1 September 2023, and questions about transfers abroad. We covered them in detail in our first article on the nFADP, generative AI and data sovereignty in Switzerland. For a bank, the nFADP sets the minimum; banking secrecy and FINMA raise the bar.

In summary

Banking secrecy also punishes negligence, FINMA treats external AI tools with the same rigour as any outsourcing, and Circular 2023/1 requires critical data to be protected. Institutions that want to benefit from AI need to know what goes into prompts and be able to prove it.

ZeusLock detects and anonymizes client data in prompts and files sent to AI assistants, directly on the workstation, and logs every incident. Request a demo for your institution.