What Changed
- FINMA's 2025 survey of ~400 Swiss institutions showed institutions average 5 active AI applications and 9 in development, predominantly using generative AI for non-critical workflows.
- FINMA reiterated that legal accountability cannot be delegated to algorithms and warned of systemic risks from vendor concentration and AI-driven fraud.
- FINMA unveiled 'Subsuma', an AI pilot utilizing open-source models—including Swiss-developed Apertus—to summarize and assess fintech licensing requests while keeping final decisions strictly with human experts.
### Background: AI in the Swiss Financial Market
Financial services operate at the intersection of intensive data processing, execution speed, and public trust. Speaking at the Swiss Bankers Association (SBA) Digital Finance Day, Marlene Amstad, Chair of the Swiss Financial Market Supervisory Authority (FINMA), emphasized that the financial sector is more exposed to artificial intelligence (AI) than most other industries. Because banking relies heavily on processing unstructured data—such as text in mortgage files, corporate documentation, client onboarding videos, and emails—the rapid advancement of generative AI tools represents a structural transformation in decision-making.
Switzerland's financial market, comprising global institutions, regional banks, insurers, asset managers, and fintech start-ups, serves as a natural laboratory for financial AI applications. FINMA has monitored industry adoption through surveys since 2021. Its 2025 survey covered approximately 400 supervised entities—including banks, securities firms, insurers, fund management companies, asset managers, and financial market infrastructures—to evaluate the current landscape of AI deployment.
### Key Survey Findings and Supervisory Expectations
FINMA's 2025 survey revealed three primary trends across the Swiss financial sector:
1. **High Momentum**: Supervised entities operate an average of five active AI applications while actively developing nine more, with larger firms moving at a significantly faster pace.
2. **Broad Deployment in Low-Criticality Use Cases**: Over 90 percent of institutions using AI employ modern generative methods. Implementation focuses heavily on early-stage, less critical tasks such as client-facing chatbots, text generation, and general process optimization. Core processes, such as automated credit decisions or regulatory metric calculations, are being approached with greater caution.
3. **Governance and Dependency Gaps**: Only about half of the institutions using AI have embedded it into an explicit AI strategy. Most rely on external software vendors, increasingly concentrated around a small group of large technology companies.
FINMA warned that heavy reliance on a limited number of technology vendors introduces systemic concentration risks; an outage or flaw in a single widely used model for risk scoring or fraud detection could jeopardize overall financial stability. Furthermore, as institutions explore "agentic AI"—systems that independently plan, execute, and chain tasks—FINMA reaffirmed a core principle: **accountability cannot be delegated to an algorithm**. Financial institutions remain fully accountable for all AI tools they deploy, whether developed in-house or acquired externally.
### Dual-Use Risks: Cyber Attacks and Digital Fraud
FINMA emphasized that a significant portion of AI-related risks facing financial institutions does not originate from their own operational use, but from malicious third parties leveraging the technology. As a dual-use technology, AI enhances both cybersecurity defenses and cyber attack tactics. Advanced models can discover unknown software vulnerabilities and construct working exploits rapidly, drastically reducing the window institutions have to patch known software weaknesses.
Concurrently, AI tools have expanded digital fraud vectors. Citing data from Interpol's global threat assessment on financial fraud, FINMA noted that AI-enabled fraud operations are approximately four and a half times more profitable than traditional methods. Common criminal applications include synthetic voice impersonation on payment orders, identity fraud in digital onboarding, and real-time phishing. Despite FINMA issuing supervisory guidance on digital fraud risks in spring 2026, the 2025 survey revealed ongoing defense gaps, with a notable share of institutions lacking dedicated digital fraud policies or responsible oversight bodies.
### AI in FINMA Supervision: The 'Subsuma' Licensing Prototype
FINMA is also integrating AI internally to enhance supervisory efficiency and effectiveness. While existing internal tools focus on document analysis pipelines and pre-inspection anomaly detection, FINMA unveiled a new prototype called **Subsuma**, which assists with external-facing regulatory evaluations.
Each year, close to 100 fintech start-ups consult FINMA to determine whether their business models require a statutory licence or regulatory setup. The Subsuma prototype assists supervisors by:
* Analyzing written requests to extract key business model elements relevant to legal assessments.
* Evaluating the business model against statutory licensing provisions and providing traceable source citations to the case file.
* Operating on FINMA's internal hardware using three distinct open-source language models to preserve technical independence.
One of the core models tested within Subsuma is **Apertus**, a fully open, transparent, and multilingual language model developed in Switzerland by ETH Zurich, EPFL, and the Swiss National Supercomputing Centre. Evaluation of the prototype relies on blind testing by FINMA experts alongside cross-model evaluations.
FINMA emphasized that Subsuma functions strictly as a supporting tool. Under Swiss principles-based financial law, administrative decisions require human legal interpretation and context. FINMA confirmed that no official communication or regulatory decision will be issued without expert human review and signature.
### Frequently Asked Questions
**What were the main findings of FINMA's 2025 industry AI survey?**
The survey of around 400 Swiss institutions showed high adoption momentum, with entities operating five AI applications on average and developing nine more. Over 90 percent of AI-using entities rely on generative methods for low-criticality tasks like chatbots and text generation. However, only half have an explicit AI strategy, and reliance on a small set of tech vendors poses systemic concentration risk.
**Can Swiss financial institutions delegate legal accountability to AI algorithms?**
No. FINMA explicitly stated that accountability cannot be delegated to an algorithm or agentic AI system. Institutions remain fully responsible for the purpose, limitations, and outcomes of all AI applications they deploy.
**What is FINMA's new 'Subsuma' AI prototype?**
Subsuma is an internal AI prototype designed to help FINMA specialists evaluate licensing inquiries submitted by approximately 100 fintech start-ups each year by summarizing business models and checking them against statutory licensing requirements with full source traceability.
**Which Swiss open-source AI model is being tested within the Subsuma project?**
FINMA is testing **Apertus**, an open, transparent, and multilingual language model developed by ETH Zurich, EPFL, and the Swiss National Supercomputing Centre.
**What risks does FINMA highlight regarding AI-driven financial fraud?**
FINMA highlighted that malicious actors use AI for real-time phishing, voice forgery on payment orders, and identity fraud during digital onboarding. According to Interpol data cited by FINMA, AI-enabled fraud is four and a half times more profitable than traditional fraud.
Who May Be Affected
Swiss banks, insurers, asset managers, securities firms, fund management companies, financial market infrastructures, fintech start-ups, and technology vendors.
Cross-Border Context
Applies within Switzerland under Swiss federal financial market supervision, referencing international threat data from Interpol and dependencies on global technology providers.
What to check next
- Monitor future FINMA publications regarding updates to the Subsuma AI prototype and its transition to larger models.
- Review FINMA supervisory guidance on digital fraud risks published in spring 2026 for specific compliance expectations.
This article provides general information and does not constitute legal advice. Consult the official text and obtain advice appropriate to your circumstances where needed.