AI in smaller financial institutions: what FINMA expects on operational risk, and how to create value with it

AI has long arrived in financial institutions

In April 2025, FINMA published the results of its first survey on artificial intelligence. Around 400 banks, securities firms, insurers, fund management companies and managers of collective assets took part. About half already use AI or are developing first applications. On average, these institutions run up to 5 applications and are working on 9 more.

In smaller institutions, this often looks unspectacular: drafts of client letters, commentary in reporting, checking onboarding documents, translations. Much of it runs on standard tools such as Copilot or ChatGPT, some of it without a formal decision. That is exactly where the supervisory question begins.

FINMA does not prohibit AI. It expects you to know where it works, and to manage the risks.

What FINMA expects

There is no dedicated AI law for the financial market. FINMA places AI within existing obligations and has set out its expectations in guidance.

Rulebook	What it means for AI	Who it applies to
FINMA Circular 2023/1 «Operational risks and resilience», in force since 1 January 2024	Risks from information and communication technology and critical data must be managed. The board defines critical functions and their tolerances for disruption. Since 1 January 2026, after a two-year transition, institutions must ensure their operational resilience. If an AI application touches a critical function or critical data, it falls within this framework.	Banks, securities firms and financial groups. Smaller institutions in FINMA categories 4 and 5 are exempt from parts of it, for example on IT, cyber and critical data. Critical functions and resilience apply to them too.
FINMA Circular 2018/3 «Outsourcing»	Anyone outsourcing a material function to a provider needs an inventory, careful selection, a contract with audit rights and an exit plan. Cloud AI services fall under this once they are material. Banking secrecy applies unchanged.	Banks, securities firms, insurers
FINMA Guidance 08/2024 «Governance and risk management when using AI»	Names the risks: models (robustness, correctness, explainability, bias), data, IT and cyber, dependence on third parties, legal and reputational. Expects an inventory, clear responsibility, documentation, testing, monitoring and training.	All supervised institutions
FINMA Guidance 05/2025 «Operational resilience»	Shows where institutions stand: of 267 institutions surveyed, half have defined no more than three critical functions.	Banks, securities firms, financial market infrastructures

For asset managers supervised by a supervisory organisation, Circular 2023/1 does not apply directly. It is still a useful benchmark for a defensible practice, especially regarding dependence on providers.

Six questions your executive team should be able to answer

1. Where does AI run today? An inventory of all applications, including the unofficial ones. The definition must be broad enough, or the inventory stays incomplete.
2. What is critical? Which applications touch a critical function, client data or decisions affecting clients?
3. Who is responsible for what? One responsible person per application, and who decides on new applications.
4. How do we check the results? Tests before going live, spot checks afterwards, and for anything going to clients, always a person who approves it.
5. What does the contract with the provider cover? Place of processing, no training with your data, subcontractors, audit rights and exit.
6. What do we do if the service fails? A tolerance for the disruption and a manual procedure that takes over.

Where AI creates value in smaller institutions

FINMA’s expectations are no reason to avoid AI. They help you choose areas of use by risk. Many worthwhile applications carry limited risk because a person checks the result.

Use	Benefit	Control
Drafts of client correspondence	Faster answers, consistent tone	The advisor checks and sends
Commentary in client reporting	Days less effort per quarter	Approval before sending
Onboarding: checking documents for completeness	Shorter lead time	The decision stays with compliance
Internal search in directives and product documents	Answers in seconds instead of minutes	Source references, internal data only

Higher requirements apply where AI decides on its own, for example on loans or investments. There, the data protection rules on automated individual decisions and the duties on suitability come into play. Such applications do not belong at the start.

Three steps to a practice that stands up to scrutiny

1. Inventory and classification: Capture all AI applications and classify them by risk. For a smaller institution, this usually takes two to three weeks.
2. Guardrails and contracts: A short directive on permitted tools and data, reviewed contracts with providers and integration into your existing risk management.
3. A first application with measurement: Choose a process with clear benefit and limited risk. Record the effort beforehand and report to the board on impact and risks.

For an overview of data protection, FINMA and the EU AI Act for the board, see our article «AI governance for boards». Detailed legal questions belong with specialised lawyers. Which AI work pays off and how you lead it is what we clarify with you.