What does FINMA expect from AI agents in Swiss finance?

Before the first Claude agent of a Swiss bank, insurer or asset manager touches a client file, its compliance team will ask which text FINMA will read the project against. The answer is a seven-page guidance note dated 18 December 2024, in which FINMA describes what its supervisors examined in discussions and on-site reviews, and the gaps they found.

FINMA Guidance 08/2024, “Governance and risk management when using artificial intelligence”, is the supervisory communication in which the Swiss regulator names the main AI risks (model, IT, cyber, third-party, legal and reputational) and lists the controls it assessed in supervised institutions: inventory, data quality, tests, documentation, explainability and independent review.

Existing law, applied in proportion

Guidance 08/2024 starts from the absence of AI-specific legislation in Switzerland: the technology-neutral, principle-based requirements of financial market law on governance and risk management already cover the risks of AI. FINMA expects each institution to align its governance, risk management and controls with its own use of AI. How far that goes depends on the size, complexity and risk profile of the institution, on the materiality of each application and on the probability that its risks materialise.

A footnote lists what raises that probability, and it reads like a description of an agent: autonomy and process integration, linkage of several models, unstructured data, exposure increased by outsourcing. Another footnote adds that AI is not a high-risk application per se.

Where things stand on 16 November 2026

FINMA has not turned the guidance into binding rules. Its published lists of guidance notes, circulars and consultations show no AI-specific text after December 2024. The most recent guidance, 05/2026 of 9 July 2026, deals with quantum computing, and none of its open consultations concerns AI (Circular 2017/6 on direct transmission; Circular 2019/2 on interest-rate risks). AI appears in other texts: Guidance 02/2026 of 9 April 2026 on digital fraud notes that advances in AI, video manipulation and deepfakes make fraudulent account openings harder to detect.

FINMA also measured adoption: between late November 2024 and mid-January 2025, it questioned around 400 banks, insurers, fund managers and other licensed institutions. According to FINMA, around half of them use AI or have applications in development, and 91% of those using AI also use generative AI. The press release ends on a recommendation: contact FINMA in good time before using AI in critical processes or to calculate regulatory parameters.

What FINMA assessed, area by area, applied to an AI agent

The core of the guidance follows the areas FINMA examined. Each one becomes a concrete question for an agent project.

Governance: a named owner, and clauses with the provider

FINMA observed that institutions focus on data protection and less on model risks such as lack of robustness, correctness, bias and explainability, and that decentralised development makes clear responsibilities hard to assign. Where applications are numerous or significant, it looked for a centrally managed inventory, defined responsibilities for development, implementation, monitoring and use, testing and documentation standards, and training. For outsourced applications, it expected additional tests and controls and contractual clauses on responsibility and liability. For a Claude agent, this means a business owner named in writing and a contract that says who answers for what when the model or the provider changes. Groups that have created the role described in our article on the Chief AI Officer in the enterprise usually attach this ownership to it.

Inventory: a broad definition of AI

Some institutions defined AI narrowly to focus on the newest risks. FINMA checked that the definition of AI was broad enough, citing the OECD approach, and that inventories were complete, which supervisors found hard since generative AI tools became accessible to every employee. Risks tend to be higher, the guidance adds, when AI serves supervisory compliance or critical functions, or strongly affects customers or employees. For an agent, the inventory entry covers the model, its instructions and every tool it can call.

Data quality when the input is text

Some institutions had defined no requirements or controls on data quality for AI. FINMA notes that more unstructured data, texts and images, is now analysed, which makes quality harder to assess, and that buyers of external solutions often know nothing about the underlying data. An agent that reads client correspondence, contracts or fund documentation needs internal rules on which sources it may read, how current they must be and who answers for their integrity.

Tests, monitoring and human corrections

FINMA looked for tests of accuracy, robustness, stability and, where relevant, bias; for performance indicators defined in advance with experts of the business area; for thresholds that keep outputs correct over time; and for the detection of data drift. The guidance adds that monitoring includes analysing the cases where users ignored or changed the output, because these manual corrections reveal weaknesses. For an agent, each edit a reviewer makes to a draft is therefore worth logging. The guidance also expects fallback mechanisms for an application that drifts away from its original objectives.

Documentation and explainability

For material applications, documentation should cover purpose, data, model selection, performance measures, assumptions, limitations, tests, controls and fallback solutions, written for its reader. On explainability, FINMA observed that results often cannot be understood, explained or reproduced, and it looked harder wherever decisions had to be justified to clients, the supervisor or the audit firm. For a language model, we think explaining means citing the document, rule or data behind each proposal, and logging the model version, the instructions and the sources consulted.

Independent review

FINMA did not always see a clear line between development and independent review, and only a few institutions had qualified staff review the whole model development process. For material applications, it looked for an objective, informed and unbiased opinion whose results were taken into account. In a large bank this reviewer sits in the second line or internal audit; a small asset manager may need an external party.

Who is accountable for the model provider?

The guidance lists growing dependence on third parties, including providers of models and cloud services in an increasingly concentrated market, among the risks of AI. FINMA’s Risk Monitor 2025 (p. 17), published on 17 November 2025, keeps outsourcing among the principal risks of the Swiss financial centre: eight out of ten banks outsource significant parts of their ICT infrastructure and its operation, and nine out of ten report having outsourced at least part of a critical function.

Circular 2018/3 and critical functions

Under FINMA Circular 2018/3 “Outsourcing”, which covers banks and insurers, responsibility for an outsourced function cannot be delegated and stays with the institution (margin no. 23), as the Risk Monitor recalls. For banks and securities firms, Circular 2023/1 on operational risks and resilience for banks, in force since 1 January 2024, defines critical functions and sets resilience requirements for them. When an agent supports a critical function, its model provider belongs in the outsourcing and resilience file. We recommend planning an exit in that file; our article on architecture choices that limit model lock-in covers the technical side.

Banking secrecy and personal data

Data sent to a model leaves the institution’s systems. For a bank, client confidentiality under Article 47 of the Banking Act applies; for every institution, the Federal Act on Data Protection governs personal data, including disclosure abroad. Compliance should know which data reaches the provider, where it is processed and how long it is kept, and these answers, which depend on the access channel and contract chosen, belong in the agent’s documentation.

What does Swiss data protection law require of an AI agent?

The revised Federal Act on Data Protection (FADP) has been in force since 1 September 2023. On 8 May 2025, the Federal Data Protection and Information Commissioner (FDPIC) confirmed that the FADP applies directly to AI-supported processing: the purpose, functionality and data sources must be transparent, users have the right to know whether they are talking to a machine, and high-risk processing calls for a data protection impact assessment under Article 22.

Article 21 and automated individual decisions

Article 21 FADP requires the controller to inform the data subject of any decision based exclusively on automated processing that has a legal consequence for them or a considerable adverse effect on them. On request, the person may express their point of view and ask for the decision to be reviewed by a natural person. These duties fall away when the decision is directly connected with a contract and grants the person’s request, or when the person has explicitly consented. Article 25 adds that the right of access covers whether such a decision was taken and the logic behind it.

The Swiss model differs from Article 22 GDPR, which gives a right not to be subject to such a decision; the FADP relies on information and human review. Wilful failure to inform under Article 21(1) can lead, on complaint, to a fine of up to CHF 250,000 under Article 60, imposed on private persons. An agent that rejects a claim or a credit request without human review falls under Article 21. An agent that prepares the file for a person who then decides stays outside it, in our reading, as long as that person really weighs the case.

Does the EU AI Act apply to a Swiss financial institution?

Switzerland is outside the EU, yet Article 2(1) of the AI Act covers providers placing AI systems on the EU market wherever they are established (point a), and providers and deployers established in a third country where the output produced by the system is used in the Union (point c). A Swiss private bank whose agent’s output reaches clients in the EU, or a Swiss group whose EU subsidiaries rely on an agent run from Zurich or Geneva, has to check its exposure. An EU subsidiary that uses the agent is a deployer in its own right under point b.

Two EU dates matter for a Swiss firm in scope. The transparency duty of Article 50, under which a chatbot or agent must disclose to the people it addresses that it is an AI unless that is obvious, has applied since 2 August 2026. The Annex III high-risk regime starts on 2 December 2027, the date set by the Digital Omnibus on AI (Regulation (EU) 2026/1744, in force since 27 July 2026); in finance it targets credit scoring of individuals (fraud detection aside) and pricing and risk assessment in life and health insurance. The other deadlines are listed on our AI Act compliance page.

The federal AI bill is not yet in consultation

On 12 February 2025, the Federal Council decided to ratify the Council of Europe Framework Convention on AI, which Switzerland signed on 27 March 2025, and asked the Federal Department of Justice and Police for a consultation draft by the end of 2026 covering transparency, data protection, non-discrimination and supervision. In its answer of 20 May 2026 to interpellation 26.3414, the Federal Council said the draft would be submitted to it by the end of 2026. Its press release of 28 October 2026 states that the consultation is scheduled to begin in spring 2027. Until a bill is adopted, FINMA’s guidance and the FADP are the texts that apply to AI in Swiss finance.

How should a Swiss financial institution start with Claude agents?

Before a first agent goes live, a Swiss institution should have four things in place: the agent in its AI inventory with a materiality rating, tests and indicators fixed before go-live, a written opinion from someone who did not build it and, for a critical process, early contact with FINMA. The first use case should leave the decision with a person.

The FINMA file, built in the order of the guidance

The inventory comes first because it drives the rest: the materiality rating sets how much testing, documentation and review the agent needs. Tests follow, with indicators chosen by business experts and thresholds written down before the first real case; reviewers’ corrections then feed the monitoring FINMA describes. The independent opinion closes the file, and for a critical process the call to FINMA comes before go-live, as its April 2025 press release recommends.

A first agent that feeds the outsourcing file

Supplier contract review is a sensible first case because it serves that same file: an agent for contract review can extract liability, audit, data-location and exit clauses from outsourcing contracts, and a lawyer validates each extraction. In asset management, an agent that pre-fills fund due diligence questionnaires works on documents the selection committee already reads, and the committee keeps its decision. Sector safeguards are described for Claude agents in banking and Claude agents in insurance.

Reading FINMA’s areas through LOOP™

The method we use, LOOP™ governance for AI agents, sorts each action of an agent into one of 4 trust zones: green for autonomous execution, orange for human validation, red for mandatory escalation, black for immediate block and CISO alert. Behind the zones sit 3 escalation levels and, for each agent, a living registry of 7 dimensions. In our reading, the registry’s identity, owners, data accessed and versioned instructions cover much of what FINMA expects from an inventory and documentation, and orange-zone validations produce the record of human corrections its monitoring section asks firms to analyse. Independent review lies outside the method: we recommend a reviewer who did not build the agent. The principles behind LOOP™ explain why the human decision stays at the centre.

Koneetiv, a French consultancy and official Anthropic partner, works exclusively on Claude for companies of all sizes and for regulated sectors. For groups with entities in Switzerland and in the EU, an ISO/IEC 42001 management system can carry one set of AI controls for entities on both sides. Before starting the inventory work described above, a baseline helps: governance is one of the 6 axes measured by our AI maturity assessment.