Switzerland’s financial regulator is testing AI to help assess licensing enquiries from fintech startups, using open-source models that include the Swiss-developed Apertus. The pilot aims to speed up regulatory reviews while leaving legal decisions with FINMA’s specialists.<\/p>\n
Switzerland’s financial regulator is testing an AI system that examines fintech business models and identifies the licences or other regulatory requirements they may need. FINMA chair Marlene Amstad presented the project, called Subsuma, in a speech published on 3 November 2026, describing how the authority is extending its use of AI from internal document analysis to preparing responses for financial businesses.<\/p>\n
FINMA receives close to 100 enquiries a year from fintech startups seeking to establish whether their proposed activities require authorisation. The questions frequently concern similar legal provisions, although the business models vary. Subsuma examines each submission, identifies the activities that require regulatory assessment and checks them against Swiss financial legislation.<\/p>\n
The software must support its findings with references to the original case documents, allowing FINMA’s specialists to verify the analysis before responding to an applicant. The regulator has also made clear that every outgoing letter will require an expert’s review and signature.<\/p>\n
Fintech licensing enquiries provide FINMA with a defined starting point for testing the technology. The authority has accumulated previous cases involving payment services, financial activities and other business models, giving its developers material against which to assess the system’s performance.<\/p>\n
A startup’s regulatory position depends on what the business intends to do, how it handles customer money and which activities it performs itself. An application providing technology to a licensed financial institution may face different requirements from a platform that holds client assets or provides regulated services directly.<\/p>\n
Subsuma begins by extracting the details needed for that assessment. It prepares an account of the business model and examines the statutory provisions that govern the activities described, attaching references to the material behind its findings.<\/p>\n
FINMA’s employees previously carried out this work through conventional document review and legal research. The pilot is intended to reduce the time spent preparing the initial assessment, although specialists will continue to examine the facts and legal conclusions.<\/p>\n
The regulator has not published processing-time comparisons or accuracy figures. Its first tests have concentrated on establishing how the models perform and what the software needs before it enters regular use.<\/p>\n
FINMA is running three open-source language models on its own hardware, including Apertus, developed by ETH Zurich, EPFL and the Swiss National Supercomputing Centre.<\/p>\n
The regulator selected this arrangement to compare different models while maintaining control over its computing environment. Fintech enquiries contain information about business plans, financial activities and operational arrangements, making the handling of submissions part of the system’s design.<\/p>\n
Apertus also gives FINMA an opportunity to test Swiss-developed AI in a regulatory setting. ETH Zurich advised the authority on its evaluation methodology, and feedback from FINMA’s specialists may contribute to further development of the model.<\/p>\n
The researchers began with smaller language models to establish the technical approach. Their preliminary findings showed that those models struggled to produce answers of the quality required for regulatory work, prompting plans to introduce more powerful versions.<\/p>\n
Testing involved previous licensing enquiries, followed by new submissions that the models had to assess. FINMA’s specialists reviewed the resulting answers without knowing which model had produced them, while separate AI-based assessments provided another comparison.<\/p>\n
Lawyers, IT specialists and AI researchers worked together to evaluate the results. The review examined how the software interpreted the submitted business models, selected legal provisions and supported its conclusions with references.<\/p>\n
FINMA has developed the evaluation process and user interface, but the project remains at the prototype stage. Amstad expects the system to shorten response times once the regulator has established that its output meets the required standard.<\/p>\n
FINMA’s experiment follows several years of AI adoption across the Swiss financial sector.<\/p>\n
In 2025, the regulator surveyed around 400 institutions, including banks, securities firms, insurers, fund managers and financial-market infrastructure providers. Respondents reported an average of five AI applications in operation, with another nine under development.<\/p>\n
More than 90 percent of institutions using AI had adopted generative technologies. Client communication, chatbots, text generation and general administrative work accounted for much of their activity.<\/p>\n
Banks and other financial businesses have proceeded more cautiously with applications that influence credit decisions, regulatory calculations and other core operations. Errors in these systems can affect customers or the institution’s financial and legal obligations, requiring closer supervision of how decisions are reached.<\/p>\n
The regulator’s survey also found that only about half of institutions using AI had incorporated it into an explicit AI strategy. Most relied on external technology suppliers, with a growing share of applications drawing on models provided by a small number of large technology groups.<\/p>\n
Amstad warned that dependence on common systems creates risks across the financial sector. If several institutions rely on the same model for activities such as fraud detection or risk assessment, an outage or faulty output could affect more than one business at the same time.<\/p>\n
FINMA’s decision to test several open-source models on its own infrastructure addresses some of these concerns within the regulator’s operations. Maintaining and evaluating the technology nevertheless requires expertise that extends beyond developing the application itself.<\/p>\n
Financial institutions are already purchasing software to analyse compliance documents, prepare regulatory reports and support internal controls. FINMA’s pilot provides a detailed example of the testing such products face when their output influences legal or financial assessments.<\/p>\n
A bank evaluating automated compliance software needs to know which records the application examined, whether it applied the correct rules and how employees can investigate an error. The supplier must also explain how the system handles confidential information and incorporates changes to legislation.<\/p>\n
These requirements affect how RegTech companies develop and sell their products. A demonstration using a prepared document shows how quickly the software produces an answer, while testing against actual business records establishes whether employees can rely on it during daily work.<\/p>\n
FINMA’s blind evaluation offers one approach to checking performance. Specialists assess the results without knowing which model generated them, reducing the influence of expectations about individual suppliers. Each answer must also contain references that allow reviewers to examine its legal basis.<\/p>\n
Exporis works with AI and financial technology businesses on product positioning and technical communications for regulated customers. Procurement teams need documented testing methods, clear descriptions of data handling and evidence from deployments that resemble their own operations. These materials allow banks and other financial institutions to examine a supplier’s claims before committing to the software.<\/p>\n
RegTech developers also face the continuing task of maintaining their applications as legislation, business procedures and customer systems change. Buyers need to know who is responsible for updates, how errors are corrected and what happens when the software encounters a case outside its tested capabilities.<\/p>\n
Subsuma has yet to enter routine operation, and FINMA has not announced a deployment date. The next phase will use larger models and build on the evaluation already completed.<\/p>\n
A faster licensing review would benefit startups awaiting clarification of their regulatory position, while FINMA’s specialists could devote more time to cases involving unusual business structures or difficult legal interpretations. The authority will measure that prospect against the accuracy of the software’s analysis and the work employees still need to complete.<\/p>\n
Exporis Insight: How RegTech Buyers Assess AI Software<\/span><\/p>\n FINMA’s testing process offers RegTech vendors a concrete reference for presenting their products to banks, insurers and other regulated businesses.<\/p>\n The regulator is examining models against previous and new cases, asking specialists to assess their answers and requiring references to the documents supporting each conclusion. Suppliers seeking commercial contracts face similar requests for evidence during technical and compliance reviews.<\/p>\n Product materials need to describe the documents an application can process, the systems it connects to and the limits of its automated analysis. Customer case studies have greater credibility when they include testing conditions, error rates and the amount of human review required.<\/p>\n Exporis develops specialist content and B2B communications that help technology suppliers present this information to financial institutions. Clear technical documentation and results from customer deployments give procurement teams a basis for examining the software alongside its contractual and operating requirements.<\/p>","protected":false},"excerpt":{"rendered":" The project offers an early view of how financial supervisors intend to use AI while retaining human control over regulatory decisions.<\/p>","protected":false},"author":5,"featured_media":0,"comment_status":"closed","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[117],"tags":[],"class_list":["post-14295","post","type-post","status-publish","format-standard","hentry","category-content-marketing"],"yoast_head":"\n