Sovara Labs
Book demo
All articles

AI regulation in Switzerland: an overview for banks

A broad overview of AI regulation for Swiss banks, covering AI-specific guidance and existing banking regulations.

FINMA logo and the Swiss Financial Market Supervisory Authority's multilingual name on a glass sign

Where Swiss AI regulation stands today

Swiss banks already operate within a broad regulatory framework when using AI. Existing rules on IT, governance, operational risk, outsourcing and data protection apply alongside FINMA’s AI-specific supervisory expectations. Its Guidance 08/2024 explains how financial institutions should identify, manage and monitor AI-related risks within this framework.

While AI regulation is evolving quickly, recurring patterns are emerging across Swiss and international approaches, often around familiar themes. For banks, these recurring themes offer a starting point for putting governance, monitoring and documentation in place that can be adapted as the rules evolve.

Core themes in AI governance

Recurring themes are emerging across Swiss and international approaches to AI governance, giving banks a starting point for preparing and adapting as regulation evolves. The controls needed in each area depend on the bank, the application and its risks.

Oversight continues after deployment. Changes to models, tools, data or instructions can affect behavior, so testing, monitoring and review need to keep pace.

How Sovara supports ongoing oversight

Sovara logs the agent’s entire activity as it executes, including LLM calls, tool calls, their inputs and outputs, available reasoning traces, and, for accountability, the identity of the person who triggered the run. This creates an audit trail that reviewers can use to reconstruct the agent’s work, investigate failures and see what happened at each step.

At high volumes, manually reviewing every agent step is rarely practical. To solve this problem, Sovara monitors behavior in real time: Sovara maintains a record of approved behavior and known failures, with tools that help teams review examples, correct mistakes and update the record efficiently. It compares new runs against this record to assess whether they follow approved patterns, resemble known failures or show unfamiliar behavior. The resulting failure resemblance score and novelty score let banks trigger warnings, request human review or stop runs before outputs are delivered.

Past run 1
Past run 2
Past run 3
Approved examplesKnown failures
New AI decision
Sovara
Compare with past behavior
Failure resemblancescoreNovelty score
Real-time controls
ContinueWarnAbort
Audit trailEntire agent activityFull audit log

Scroll horizontally to follow the full flow.

Sovara supports banks’ oversight in four areas:

  • Logging and audit trail. Sovara records LLM calls, tool calls, inputs, outputs and available reasoning traces so reviewers can reconstruct agent activity.
  • Monitoring and intervention. Failure resemblance and novelty scores flag potential failure patterns and unfamiliar behavior without requiring someone to inspect every step. Banks define when these signals trigger warnings, human review or stopping a run.
  • Testing corrections. Sovara compares evaluation runs and links results to traces, helping teams check fixes and detect regressions. Banks define the test cases and acceptance criteria.
  • Evidence for independent review. Auditors can inspect agent traces and evaluation results directly in Sovara. Banks appoint qualified reviewers and maintain their independence.

Deep dive: expectations, controls and how Sovara helps

Our review draws on 460 cited entries across 44 source documents, including Swiss legislation, FINMA circulars and guidance, self-regulation and incorporated standards. We group them into 16 topics and combine entries addressing the same control objective into 155 consolidated points, retaining source-specific conditions and citations.

The applicable requirements depend on the institution and the AI use case, with additional rules for activities such as lending, AML, trading and digital identification. The scope and applicability section sets out the relevant conditions, exemptions and effective dates.

Expand a topic to see the detailed points, their source references and how Sovara helps.

Scope and applicabilityDetermine which rules apply to the bank and AI use, including materiality, exemptions and effective dates. Distinguish binding requirements from guidance.

  • SCP-01: Technology-neutral coverage. Use a sufficiently broad AI definition and apply the relevant technology-neutral banking and identification framework.

    Sources: AI §1, p. 3 (FINMA guidance); AI §2.2, pp. 4-5 (FINMA guidance); VID mn 53, pp. 11-13 (FINMA circular).

  • SCP-02: Risk-proportionate application. Scale controls to the institution, actual use and application materiality, while observing each regime’s scope and relief provisions.

    Sources: AI §1, p. 3 (FINMA guidance); AI §2.2 fn 7, p. 5 (FINMA guidance); AI §3, p. 7 (FINMA guidance); OPS mn 19, p. 5 (FINMA circular); GOV mn 1, 8, p. 3 (FINMA circular).

  • SCP-04: Operational-risk exemptions. Apply the specific category and small-regime exemptions before selecting supplementary operational-risk tools.

    Sources: OPS mn 20, p. 5 (FINMA circular); OPS mn 21, p. 5 (FINMA circular); OPS mn 33, p. 6 (FINMA circular; not category 4–5 banks).

  • SCP-05: Outsourcing perimeter and relief. Determine whether a significant function is outsourced, respect reserved functions, and substantiate any intragroup adjustment, FINMA exception or small-bank relief.

    Sources: OUT mn 1-6.3, p. 3 (FINMA circular); OUT mn 7, p. 3 (FINMA circular); OUT mn 22, p. 5 (FINMA circular); OUT mn 36, p. 7 (FINMA circular); OUT mn 36.1-36.3, p. 7 (FINMA circular; small-bank regime).

  • SCP-06: Personal-data perimeter. Determine whether the AI processes natural-person data within the FADP’s territorial and material scope.

    Sources: FADP Arts. 2-3 (legislation).

  • SCP-07: Processing-record exemptions. Check the limited recordkeeping exemption; high-risk profiling is not covered by the small-organisation relief.

    Sources: FADP Art. 12(5); DPO Art. 24 (legislation); DPO Art. 24 (legislation).

  • SCP-08: Automated-decision exceptions. Assess the specific statutory exceptions before applying the automated-decision information and review duties.

    Sources: FADP Art. 21(3) (legislation; solely automated decisions).

  • SCP-09: DPIA exceptions. Use a DPIA exemption only when its particular statutory conditions are satisfied.

    Sources: FADP Art. 22(4)-(5) (legislation).

  • SCP-11: Identification rule interaction. Read the bank due-diligence agreement and digital-identification routes within the applicable AML framework.

    Sources: CDB Arts. 1-2, p. 10 (self-regulation); VID mn 1-3, p. 3 (FINMA circular).

  • SCP-12: Digital-identification amendments. Apply the circular as amended with effect from 1 November 2026, observing the one-year transition until 1 November 2027 for the mn 39 adjustments to residence verification and liveness detection.

    Sources: VID cover; list of modifications, pp. 1, 15 (FINMA circular); VID mn 54, p. 14 (FINMA circular).

  • SCP-13: Financial-service and client classification. Determine the financial service and client category before applying the corresponding conduct duties.

    Sources: FINSA Arts. 3-5, 7, 20 (legislation); CONDUCT mn 1-3, p. 3 (FINMA circular).

  • SCP-14: Trading and algorithmic-participant perimeter. Distinguish market-wide conduct prohibitions, supervised-bank duties and venue-imposed algorithmic participant controls.

    Sources: FMIO Art. 31(1)-(2), chapeau (legislation); MARKET mn 3-7, p. 3 (FINMA circular).

  • SCP-18: Official-material copyright exclusion. Distinguish the official enactments and reports excluded from copyright protection from other material used in AI workflows.

    Sources: COPY Art. 5 (legislation).

Governance and competenceAssign accountable owners, approval rights and independent oversight, and retain responsibility for AI-supported decisions. Ensure staff, management, reviewers and providers have appropriate skills, training and resources.Sovara supports governance through reviewed guidance, second-person approvals and records available to auditors.

Responsibilities and oversight

Expertise, resources and training

  • SKL-01: Role-appropriate AI and technology training. Ensure sufficient role-appropriate AI expertise and provide suitable AI, ICT and cyber awareness and training across relevant roles and hierarchies, with documented ICT responsibilities.

    Sources: AI §2.1, p. 4, para. 2 (FINMA guidance; banks with many or significant AI applications); OPS mn 26 and fn 7, pp. 5-6 (FINMA circular); OPS mn 49, p. 8 (FINMA circular); CY24 §2.3, pp. 6-7 (FINMA guidance); AI23 p. 24, Governance and responsibility, expectations paragraph (published supervisory expectations).

  • SKL-02: Qualified outsourced-AI providers. Use providers with the qualifications needed for the outsourced AI function.

    Sources: AI §2.1, p. 4, para. 2 (FINMA guidance; outsourced AI).

  • SKL-03: Qualified independent reviewers. Ensure reviewers have suitable expertise throughout the AI development process.

    Sources: AI §2.7, p. 7 (FINMA guidance).

  • SKL-06: Management competence. Ensure the board collectively and executive management have the knowledge and experience needed for their responsibilities.

    Sources: GOV mn 16, p. 5 (FINMA circular); GOV mn 51, p. 8 (FINMA circular).

  • SKL-07: Adequate operating and control resources. Provide adequate personnel, infrastructure and IT, and empowered control resources.

    Sources: GOV mn 13, p. 4 (FINMA circular); GOV mn 64, p. 10 (FINMA circular).

Inventory and risk assessmentMaintain relevant AI, data and dependency inventories and assess AI and operational risks. Apply the additional registers, classifications, risk methods and tolerances required for particular activities and regulatory regimes.Sovara supports risk assessment with evidence from recorded failures and evaluations.

AI inventories, classification and dependencies

General AI and operational-risk assessment

Outsourcing, privacy and trading risk assessments

  • RSK-09: Pre-contract outsourcing risk analysis. Specify the outsourced function and assess its risks before entering the contract.

    Sources: OUT mn 16, p. 5 (FINMA circular).

  • RSK-10: Privacy and security risk assessment. Determine personal-data protection needs and security risk and perform a prior DPIA when the statutory high-risk trigger is met.

    Sources: FADP Art. 22(1)-(2) (legislation); DPO Art. 1 (legislation).

  • RSK-18: Algorithmic trading limits. Apply the required trading thresholds and limits to prevent algorithmic disruption.

    Sources: FMIO Art. 31(2)(b)-(c) (legislation).

Data qualityUse accurate, current, complete and suitable AI data, with evidence of its sources, preparation and potential biases. Apply additional identity, client and accounting data requirements where the relevant activity is in scope.Sovara directly helps reviewers inspect and correct the data sources agents select.

General AI data quality and provenance

  • DAT-01: Complete, correct and suitable inputs. Set data-quality requirements so that AI data, including personal data, is complete, correct, current and of integrity, with secured availability and access.

    Sources: AI §1 fn 2, likelihood (FINMA guidance); AI §2.3, p. 5, para. 2 (FINMA guidance); FADP Art. 6(5) (legislation).

  • DAT-02: Representative data and bias safeguards. Assess representativeness, bias and the suitability of purchased or manipulated data.

    Sources: AI §2.3, p. 5, para. 1 (FINMA guidance); AI23 p. 24, Robustness and reliability, risk paragraph (published supervisory expectations).

  • DAT-03: Data provenance and preparation evidence. Document data sources, preparation, relevance, bias and stability for material AI applications.

    Sources: AI §2.5, p. 6 (FINMA guidance; material AI applications).

Identity, client and accounting data

  • DAT-05: Identity evidence and KYC currency. Collect the required identity data and evidence, verify onboarding information and suspicious patterns, and update KYC information on a risk basis.

    Sources: AMLA Art. 7(1bis) (legislation); CDB Art. 7, p. 13 (self-regulation); VID mn 11, p. 4 (FINMA circular).

  • DAT-06: Client-profile accuracy. Collect necessary client-profile information and investigate inaccurate, doubtful or contradictory information used in advice or creditworthiness assessment.

    Sources: FINSO Art. 17(4) (legislation); CONDUCT mn 13, p. 4 (FINMA circular); KKG Art. 31 (legislation).

  • DAT-15: Auditable bookkeeping inputs and outputs. Keep AI-supported bookkeeping complete, accurate, evidenced and auditable.

    Sources: CO Art. 957a (legislation).

Privacy and banking secrecyProtect personal data and banking secrets. Observe purpose limits, consent where required, notices, foreign-transfer safeguards and individual rights.

  • PRI-01: Banking secrecy. Protect banking secrets, including against negligent disclosure, and observe continuing secrecy and the actual statutory exceptions.

    Sources: BANK Art. 47(1)-(2) (legislation); BANK Art. 47(4)-(5) (legislation).

  • PRI-02: Lawful and purpose-limited processing. Limit AI processing to lawful, proportionate and purpose-compatible uses; employee data has its additional employment-purpose restriction.

    Sources: FADP Art. 6(1)-(3) (legislation); CO Art. 328b (legislation).

  • PRI-03: Data minimisation and deletion. Delete or anonymise personal data that is no longer needed.

    Sources: FADP Art. 6(4) (legislation).

  • PRI-04: Valid consent where required. Ensure consent is informed and explicit where required; video identification has a specific consent requirement for video and recording.

    Sources: FADP Art. 6(6)-(7) (legislation); VID mn 12, p. 4 (FINMA circular).

  • PRI-05: Privacy by design and default. Build privacy safeguards into the processing lifecycle and use privacy-protective default settings.

    Sources: FADP Art. 7(1)-(2) (legislation); FADP Art. 7(3) (legislation).

  • PRI-06: Foreign-transfer safeguards. Use the applicable adequacy or safeguard route for foreign transfers and rely on an exception only within its narrow conditions.

    Sources: FADP Art. 16 (legislation); FADP Art. 17 (legislation).

  • PRI-07: Clear processing notices. Provide clear, accessible collection and purpose information, including foreign-transfer disclosures and indirect-collection timing, subject to the actual information-duty exceptions.

    Sources: FADP Art. 19(1)-(3) (legislation); FADP Art. 19(4)-(5) (legislation); FADP Art. 20 (legislation); DPO Art. 13 (legislation).

  • PRI-08: Residual-high-risk consultation. Carry out the prescribed prior consultation where a DPIA leaves residual high risk, subject to the applicable statutory procedure.

    Sources: FADP Art. 23(1)-(4) (legislation).

  • PRI-10: Subject access and response. Provide the required processing information and understandable access, verify identity, assist as required and observe the normal 30-day response period and justified restrictions.

    Sources: FADP Art. 25(1)-(2) (legislation); FADP Arts. 25(3)-(7), 26 (legislation); DPO Arts. 16-17 (legislation); DPO Art. 18 (legislation).

  • PRI-11: Conditional data portability. Provide qualifying portable data in a usable format within the applicable deadlines; assess reasoned restrictions and the exclusion of controller-generated inferences.

    Sources: FADP Arts. 28-29 (legislation); DPO Art. 20 (legislation); DPO Arts. 21-22 (legislation).

  • PRI-12: Personality protection and remedies. Avoid unlawful personality infringements, assess any claimed justification and respect rectification and other civil remedies.

    Sources: FADP Art. 30 (legislation); FADP Art. 31(1) (legislation); FADP Art. 32 (legislation); CC Arts. 28-28a (legislation).

  • PRI-13: Creditworthiness justification limits. Observe the specific limits on the statutory justification for creditworthiness processing.

    Sources: FADP Art. 31(2)(c) (legislation).

  • PRI-14: Research-processing safeguards. Apply the specific safeguards and anonymisation conditions for the research-purpose justification.

    Sources: FADP Art. 31(2)(e) (legislation).

Testing and independent assuranceTest AI and system performance, validate controls and critically question data, models and results. Keep independent audits distinct from AI testing, with their own purposes and responsibilities.Sovara directly supports testing and independent assurance through evaluation records and auditor platform access.

AI and system evaluation

  • TST-01: Fit-for-purpose implementation before release and change. Test and validate AI/system requirements before production and relevant changes.

    Sources: AI §2.1, p. 4, para. 2 (FINMA guidance; banks with many or significant AI applications); AI §2.4 fn 9, p. 6 (FINMA guidance); OPS mn 52, p. 8 (FINMA circular).

  • TST-02: Data and functionality testing. Plan tests of data quality and functionality.

    Sources: AI §2.4, p. 5 (FINMA guidance).

  • TST-03: Performance criteria and regular validation. Ensure sufficiently accurate, robust and reliable AI results; predefine suitable performance indicators and validate accuracy, robustness and stability with appropriate metrics.

    Sources: AI §2.4, p. 5 (FINMA guidance); AI §2.4, pp. 5-6 (FINMA guidance); AI23 p. 24, Robustness and reliability, expectations paragraph (published supervisory expectations).

  • TST-04: Bias testing. Perform AI bias tests where necessary.

    Sources: AI §2.4, p. 5 (FINMA guidance).

  • TST-05: Expert cases, adversarial tests and benchmarks. Use appropriate expert questions, expected answers, cases, adversarial tests and benchmarks.

    Sources: AI §2.4, pp. 5-6 (FINMA guidance); AI §2.4 fn 8, p. 5 (FINMA guidance); AI §2.4 fn 10, p. 6 (FINMA guidance).

  • TST-09: Independent model acceptance and validation. Separate development from objective independent review and obtain qualified validation or approval evidence initially, periodically and after relevant changes.

    Sources: AI §2.7, p. 7 (FINMA guidance); AI §2.7, p. 7 (FINMA guidance; material AI applications); GOV mn 72, p. 10 (FINMA circular).

  • TST-10: Document test evidence. Document performance, tests and controls for material AI applications.

    Sources: AI §2.5, p. 6 (FINMA guidance; material AI applications).

  • TST-26: Critical challenge of data, models and results. Keep AI data, models and results open to critical questioning.

    Sources: AI23 p. 24, Robustness and reliability, expectations paragraph (published supervisory expectations).

Backtesting, sensitivity and stress tests

  • TST-06: Backtesting and out-of-sample validation. Validate predictions with tests where the correct result is known, such as backtesting and out-of-sample tests.

    Sources: AI §2.4 fn 8, p. 5 (FINMA guidance).

  • TST-11: Limits, boundary behaviour and repeatability. Test application limits, behaviour in borderline cases and the repeatability of results.

    Sources: AI §2.4 fn 8, p. 5 (FINMA guidance).

  • TST-12: Sensitivity and stress tests. Run appropriate sensitivity analyses and stress tests and review their implications.

    Sources: AI §2.4 fn 8, p. 5 (FINMA guidance); GOV mn 69, p. 10 (FINMA circular).

Operational-risk scenarios and control effectiveness

  • TST-13: Operational-risk scenarios and comparison. Use the applicable operational-risk scenario and loss analysis and compare findings across business areas.

    Sources: OPS mn 36, p. 7 (FINMA circular; not category 4–5 banks); OPS mn 37, p. 7 (FINMA circular; not category 4–5 banks).

  • TST-14: Control design and operating effectiveness. Assess organisational, risk-management and key-control design and operating effectiveness, independently where prescribed; demonstrate AML effectiveness beyond evidence that a control merely ran.

    Sources: OPS mn 31, p. 6 (FINMA circular); GOV mn 37, p. 7 (FINMA circular; banks with an audit or risk committee); GOV mn 44, p. 7 (FINMA circular; banks with an audit or risk committee); GOV mn 91, p. 12 (FINMA circular); CY24 §2.2, p. 6 (FINMA guidance); AML23 §2.1(b), p. 5 (FINMA guidance); AML26 §2.1(b)(ii), p. 9 (FINMA guidance).

Independent audit and assurance

Explainability and human oversightMake relevant AI decisions, limitations and uncertainty understandable, and make AI use appropriately transparent. Provide required human review and analyse overrides and corrections.Sovara directly connects agent actions to the guidance and human corrections behind them.

  • EXP-01: Understandable decisions and strategies. Make AI results explicable according to the recipient, relevance and process integration; explain decisions requiring justification, understand their drivers and conditional behaviour, and make algorithmic trading-strategy documentation understandable to third parties.

    Sources: AI §2.6, p. 6 (FINMA guidance; decisions that must be justified); MARKET mn 63, p. 9 (FINMA circular); AI23 p. 25, Transparency and explicability, expectations paragraph (published supervisory expectations).

  • EXP-02: Plausibility, robustness and uncertainty. Evidence robustness, reliability and traceability and assess plausibility.

    Sources: AI §2.5, p. 6 (FINMA guidance; material AI applications); AI §2.6, p. 6 (FINMA guidance; decisions that must be justified).

  • EXP-03: Solely automated decision rights. For qualifying solely automated decisions, provide the required notice, requested human review and opportunity to state a viewpoint, together with the applicable right to information about decision logic.

    Sources: FADP Art. 21(1) (legislation; solely automated decisions); FADP Art. 21(2) (legislation; solely automated decisions); FADP Art. 25(2)(f) (legislation; solely automated decisions).

  • EXP-06: Overrides and corrections. Analyse cases where users ignored or changed AI outputs, as these corrections can reveal weaknesses.

    Sources: AI §2.4, p. 6 (FINMA guidance).

  • EXP-07: Transparency about AI use. Make the use of AI transparent according to the recipient, relevance and process integration, separately from explaining how an output was produced.

    Sources: AI23 p. 25, Transparency and explicability, expectations paragraph (published supervisory expectations).

AI lifecycle, monitoring and recordsControl AI development, deployment, change and retirement, and monitor performance and risks. Keep the documentation, logs and records required by each applicable regime for its prescribed periods.Sovara directly supports ongoing monitoring and records changes to agent guidance for review and reporting.

Development and change

  • LIF-01: Risk assessment for innovation and change. Assess development/monitoring weaknesses and adaptation frequency, consult risk control before relevant innovations, assess major changes and new-technology AML risk beforehand, and update AML analysis after material business changes.

    Sources: AI §1 fn 2, likelihood (FINMA guidance); OPS mn 32, p. 6 (FINMA circular); GOV mn 73, p. 10 (FINMA circular); AMLO Art. 23 (legislation); AML23 §2.3, pp. 5-6 (FINMA guidance).

  • LIF-02: Lifecycle controls through retirement. Maintain controls throughout development and procurement and the critical-data lifecycle, and manage end-of-life ICT risks.

    Sources: OPS mn 50, p. 8 (FINMA circular); OPS mn 57, p. 8 (FINMA circular; not category 4–5 banks); OPS mn 74, p. 10 (FINMA circular; not category 4–5 banks).

  • LIF-03: Written framework and operating procedures. Set documentation standards and maintain central, audience-appropriate documentation of the approved risk framework, material controls and applicable processing, AML and identification procedures.

    Sources: AI §2.1, p. 4, para. 2 (FINMA guidance; banks with many or significant AI applications); AI §2.5, p. 6 (FINMA guidance); OPS mn 108, p. 13 (FINMA circular; not small-bank regime); GOV mn 52, p. 8 (FINMA circular); DPO Art. 5(1)-(3) (legislation; large-scale sensitive data or high-risk profiling); AMLO Art. 26 (legislation); VID mn 9, p. 4 (FINMA circular).

  • LIF-04: Purpose and design documentation. Document the purpose, data selection and preparation, model selection and controls of material AI applications.

    Sources: AI §2.5, p. 6 (FINMA guidance; material AI applications).

  • LIF-05: Assumptions and limitations. Document material AI assumptions and limitations.

    Sources: AI §2.5, p. 6 (FINMA guidance; material AI applications).

  • LIF-06: Review and incident feedback. Feed independent-review findings and cyber-incident lessons into development and continuous improvement of resilient design.

    Sources: AI §2.7, p. 7 (FINMA guidance; material AI applications); CY24 §2.4, p. 7 (FINMA guidance); RES §4, p. 12 (FINMA guidance).

Ongoing monitoring

  • MON-01: Output thresholds, drift and adaptation. Define appropriate output thresholds and validation methods, recognise automatic-optimisation drift risk, detect and address data drift, and assess adaptive AI systems on an ongoing basis.

    Sources: AI §2.4, p. 6 (FINMA guidance); AI §2.4 fn 9, p. 6 (FINMA guidance); AI23 p. 24, Robustness and reliability, risk paragraph (published supervisory expectations).

  • MON-02: Risk and control indicators. Monitor risk profiles and limits using suitable risk and control indicators.

    Sources: OPS mn 35, p. 6 (FINMA circular; not category 4–5 banks); GOV mn 45-46, pp. 7-8 (FINMA circular; banks with an audit or risk committee); GOV mn 71, p. 10 (FINMA circular).

  • MON-06: Outsourcing risk in ongoing controls. Integrate relevant outsourcing risks into the bank’s ongoing control framework.

    Sources: OUT mn 20, p. 5 (FINMA circular).

  • MON-08: Relationship and transaction monitoring. Maintain effective AML relationship and transaction monitoring.

    Sources: AMLO Art. 20(1) (legislation).

Evidence and retention

  • AUD-06: Traceable decisions, assumptions and findings. Keep verifiable tolerance/limit documentation, auditable risk decisions and monitoring, and institution-specific cyber-exercise evidence.

    Sources: GOV mn 58, p. 9 (FINMA circular; category 1–3 banks); BO Art. 12(3) (legislation); CY24 §4, pp. 9-10 (FINMA guidance).

  • AUD-07: Document and retain the DPIA. Document processing, risks and safeguards in the DPIA and retain it for at least two years after processing ends.

    Sources: FADP Art. 22(3) (legislation); DPO Art. 14 (legislation).

  • AUD-08: Trace data operations. Ensure traceability of relevant personal-data entry and disclosure.

    Sources: DPO Art. 3(3)(a)-(b) (legislation).

  • AUD-09: Conditional personal-data processing logs. Where the statutory sensitive-data/profiling logging trigger applies, record the required operations and identities and keep the logs separate, access-restricted and retained for at least one year.

    Sources: DPO Art. 4(1) (legislation); DPO Art. 4(3)-(4) (legislation); DPO Art. 4(5) (legislation).

  • AUD-11: Assessable AML and due-diligence evidence. Keep accessible, independently assessable AML/due-diligence records, document timely completion and the required controls/risk-analysis structure, and observe the AMLA ten-year retention rule.

    Sources: AMLA Art. 7(1) (legislation); AMLA Art. 7(2)-(3) (legislation); AMLO Art. 22 (legislation); AML23 Annex, pp. 8-9 (FINMA guidance); AML26 §2.1(b)(i), p. 9 (FINMA guidance); CDB Arts. 44-45, pp. 35-36 (self-regulation).

  • AUD-12: Identification recordings and evidence. Record the complete relevant identification interview and archive its photographs/recordings; document E-ID identification where that route is used.

    Sources: VID mn 8, p. 4 (FINMA circular); VID mn 17, p. 5 (FINMA circular); VID mn 44.4, p. 9 (FINMA circular).

  • AUD-13: Advice and conflict documentation. Document client assessments, needs, recommendation reasons and actual/potential conflicts, with evidence enabling the required account within ten working days.

    Sources: FINSA Art. 15 (legislation); FINSO Art. 18 (legislation); FINSO Art. 28 (legislation).

  • AUD-14: Trading audit trail and communication records. Flag algorithmic orders and record cancellations, document clearly suspicious trades and retain the specified trading communications for at least two years.

    Sources: FMIO Art. 31(2), chapeau (legislation); MARKET mn 59, p. 9 (FINMA circular); MARKET mn 60-61, p. 9 (FINMA circular).

Security and accessProtect AI systems and data from unauthorised access, manipulation, leakage and attacks. Cover access controls, secure environments, threat detection and relevant supplier safeguards.

  • SEC-01: Threat and vulnerability assessment. Assess AI attack and failure exposure, outsourcing/cloud IT-security risks and vulnerabilities, and keep emerging digital fraud threats under review.

    Sources: AI §1 fn 2, likelihood (FINMA guidance); OPS mn 63, p. 9 (FINMA circular); FRAUD §3.1.2 Findings/Recommendations, pp. 6-8 (FINMA guidance); AI23 p. 24, Robustness and reliability, risk paragraph (published supervisory expectations).

  • SEC-02: Risk-appropriate security objectives. Protect AI-related data and production assets for confidentiality, integrity and availability, with appropriate access and the traceability required for personal data.

    Sources: AI §2.3, p. 5, para. 2 (FINMA guidance); OPS mn 55, p. 8 (FINMA circular); OPS mn 64, p. 9 (FINMA circular); OPS mn 75, p. 10 (FINMA circular; not small-bank regime); FADP Art. 8 (legislation); DPO Art. 2 (legislation).

  • SEC-04: Separate environments. Separate production, development and testing where the applicable operational-risk provisions require it.

    Sources: OPS mn 51, p. 8 (FINMA circular; not category 4–5 banks).

  • SEC-05: Protect development and test data. Protect critical data used during AI development and testing.

    Sources: OPS mn 76, p. 10 (FINMA circular; not category 4–5 banks).

  • SEC-06: Least privilege and access review. Apply data, premises and user-access controls, maintain appropriate central authorisations informed by the threat landscape, and review privileges regularly where required.

    Sources: OPS mn 78, p. 10 (FINMA circular; not category 4–5 banks); CY24 §2.2, p. 5 (FINMA guidance); DPO Art. 3(1)(a)-(c) (legislation).

  • SEC-07: Security monitoring and logs. Protect and monitor critical-data systems and use timely logging, monitoring and analysis to detect attacks.

    Sources: OPS mn 65, p. 9 (FINMA circular); OPS mn 77, p. 10 (FINMA circular; not category 4–5 banks); CY24 §2.4, p. 7 (FINMA guidance).

  • SEC-08: Data-loss prevention. Cover all critical data in the applicable data-loss prevention measures.

    Sources: CY24 §2.3, p. 6 (FINMA guidance).

  • SEC-09: Supplier security obligations. Agree and monitor the security requirements for outsourced functions.

    Sources: OUT mn 24, p. 5 (FINMA circular).

  • SEC-11: Deepfake and manipulated-video detection. Address deepfakes and manipulated video in relevant digital identification and fraud controls.

    Sources: FRAUD §3.2 Recommendations, p. 10 (FINMA guidance).

  • SEC-12: Identification liveness checks. Apply the route-specific liveness and address checks for online identification by photograph, qualified electronic signature and E-ID; the mn 39 adjustments must be in place by 1 November 2027.

    Sources: VID mn 32, p. 7 (FINMA circular); VID mn 39, p. 8 (FINMA circular); VID mn 44.3, p. 9 (FINMA circular).

  • SEC-14: Insider-information barriers. Apply information barriers and confidentiality controls to prevent prohibited access to or influence from insider information.

    Sources: MARKET mn 49-52, pp. 7-8 (FINMA circular).

Third parties and outsourcingAssess AI providers and define contractual duties, security, oversight and audit access. Control subcontracting and retain the bank’s responsibilities.Sovara supports additional testing and review of outsourced AI when its use is instrumented.

  • VEN-01: Understand purchased AI and vendor methods. Identify AI embedded in purchased services and understand relevant vendor data and methods.

    Sources: AI §2.1, p. 4, para. 1 (FINMA guidance).

  • VEN-02: Provider capability and concentration due diligence. Assess provider capability, resources and concentration, including third-party AI concentration and enhanced due diligence for critical-data providers.

    Sources: AI §2, p. 3 (FINMA guidance); OPS mn 82, p. 11 (FINMA circular); OUT mn 17, p. 5 (FINMA circular).

  • VEN-03: Additional outsourced-AI assurance. Apply the additional tests and controls appropriate to outsourced AI.

    Sources: AI §2.1, p. 4, para. 2 (FINMA guidance; outsourced AI).

  • VEN-04: Document contractual responsibilities and safeguards. Use the required written/text-evidenced agreement to define responsibilities, liability, duties and interfaces and implement the specified outsourcing safeguards.

    Sources: AI §2.1, p. 4, para. 2 (FINMA guidance; outsourced AI); OUT mn 19, p. 5 (FINMA circular); OUT mn 32, p. 6 (FINMA circular); OUT mn 34, p. 6 (FINMA circular).

  • VEN-05: Provider instruction and control rights. Retain the required rights to instruct and control the service provider.

    Sources: OUT mn 21, p. 5 (FINMA circular).

  • VEN-06: Foreign data and remote access safeguards. Protect critical data stored abroad or accessed remotely under the applicable operational-risk provision.

    Sources: OPS mn 79, p. 11 (FINMA circular; not small-bank regime).

  • VEN-07: Effective supervision and foreign audit access. Preserve effective FINMA supervision and enforceable foreign inspection and audit rights.

    Sources: OUT mn 28, p. 6 (FINMA circular); OUT mn 30, p. 6 (FINMA circular).

  • VEN-08: Subcontractor safeguards with distinct consent rules. For outsourcing, require early subcontractor notice, flow-down safeguards and orderly termination options; where the FADP processing rule applies, also ensure processor security and the required subprocessor approval.

    Sources: OUT mn 33, p. 6 (FINMA circular); FADP Art. 9(2)-(3) (legislation).

  • VEN-09: Supplier security assurance and remediation. Agree security obligations, assess suppliers regularly and secure prompt remediation of supplier vulnerabilities.

    Sources: CY24 §2.1, pp. 4-5 (FINMA guidance); CY24 §2.1, p. 4, bullets 1-2 (FINMA guidance).

  • VEN-10: Provider incident information and responsibility. Ensure relevant provider cyber information reaches the bank, including from non-material providers, and ensure processors promptly notify controllers of breaches; retain the bank’s outsourced-cyber responsibility.

    Sources: CY20 §2 fn 5, p. 3 (FINMA guidance); CY24 §3, p. 8, bullet 7 (FINMA guidance); FADP Art. 24(3) (legislation).

  • VEN-13: Permitted data processing and confidentiality. Use processors only within the permitted instructions and confidentiality constraints.

    Sources: FADP Art. 9(1) (legislation).

  • VEN-14: Qualified and controlled financial-service delegates. Select, instruct and supervise qualified delegates, retain required responsibility and evidence/plausibility review, and preserve provider-chain responsibility for client information.

    Sources: AMLO Art. 28 (legislation); AMLO Art. 29 (legislation); CDB Art. 43, p. 34 (self-regulation); VID mn 51, p. 10 (FINMA circular); FINSA Art. 23 (legislation); FINSA Art. 24 (legislation).

Incidents and regulatory notificationDefine AI exception handling and recognise, investigate, contain and escalate incidents. Route them into the existing FINMA, NCSC/BACS and FDPIC reporting processes.Sovara directly supports AI exception handling, incident review and reporting with recorded evidence.

AI exceptions and internal escalation

FINMA and cyber reporting

Personal-data breach notices

  • INC-14: FDPIC high-risk breach notice. Notify the FDPIC promptly when the statutory high-risk breach trigger applies, with the required content and progressive completion.

    Sources: FADP Art. 24(1)-(2) (legislation); DPO Art. 15(1)-(2) (legislation).

  • INC-15: Affected-person breach information. Inform affected persons where required and make the information comprehensible.

    Sources: FADP Art. 24(4)-(5) (legislation); DPO Art. 15(3) (legislation).

Continuity and resiliencePrepare fallbacks for material AI applications and include AI that supports critical functions in continuity planning. Plan for supplier change and exit.

  • RES-01: AI fallback and continuity arrangements. Prepare fallback mechanisms and document fallback solutions for material AI applications. Where AI supports a critical function, include it in the bank’s continuity and operational-resilience arrangements.

    Sources: AI §2.4 fn 9, p. 6 (FINMA guidance); AI §2.5, p. 6 (FINMA guidance; material AI applications); OPS mn 86, p. 11 (FINMA circular); OPS mn 102, p. 13 (FINMA circular); RES §3.2 Notes, p. 9 (FINMA guidance).

  • RES-07: Supplier continuity, transfer and exit. Assess provider change and continuity, enable orderly insourcing or transfer and maintain the joint emergency security framework.

    Sources: OUT mn 18, p. 5 (FINMA circular); OUT mn 18.1, p. 5 (FINMA circular); OUT mn 25, p. 6 (FINMA circular).

AML and identificationApply the relevant identity, beneficial-owner, transaction-monitoring and investigation requirements where AI supports AML controls. Preserve the conditions, evidence and approvals specific to each identification route.Sovara makes the work of AI assistants in AML investigations reviewable.

Monitoring and investigation

  • AML-07: High-risk cases and transaction flags. Apply mandatory high-risk cases, internal risk flags and the required transaction criteria and flags.

    Sources: AMLO Art. 13(3)-(6) (legislation); AMLO Art. 14 (legislation).

  • AML-08: Alert evaluation and investigation. Investigate unusual/high-risk transactions and PEP circumstances, evaluate alerts promptly, obtain reliable evidence and document investigation and decision handling.

    Sources: AMLA Art. 6(2)-(4) (legislation); AMLO Arts. 15-16 (legislation); AMLO Art. 20(3) (legislation); AMLO Art. 25(1)(c)-(e) (legislation).

  • AML-09: IT-supported transaction monitoring. Use the required IT-supported monitoring unless the limited exception applies.

    Sources: AMLO Art. 20(2),(4) (legislation).

  • AML-19: AML duties unchanged by AI. Where AI supports onboarding, monitoring or investigations, the bank’s AML risk analysis, MROS reporting and asset-freeze duties continue to apply in full.

    Sources: AMLO Art. 25(2) (legislation); AMLA Art. 9(1),(1quinquies) (legislation); AMLA Art. 10 (legislation).

Identity, beneficial owners and due diligence

Digital-identification routes and evidence

  • AML-14: Identification methods and document authenticity. Observe the applicable in-person/correspondence authentication rules and digital photo/document checks, which now also cover identification documents with a QR code.

    Sources: CDB Arts. 9-11, pp. 14-15 (self-regulation); VID mn 13-15, p. 4 (FINMA circular); VID mn 32, p. 7 (FINMA circular); VID mn 6, 14-15, pp. 3-4 (FINMA circular).

  • AML-15: Digital identification quality and high-risk gates. Stop unreliable identification and obtain required high-risk approvals, while meeting the relevant quality, evidence and secure-transmission conditions.

    Sources: VID mn 18-22, p. 5 (FINMA circular); VID mn 31.1-31.5, pp. 6-7 (FINMA circular).

  • AML-16: Qualifying digital identification routes. Use the chosen electronic-copy, bank-transfer, biometric-chip, qualified-signature or digital-authentication route only with its specific conditions and evidence.

    Sources: VID mn 29-31, p. 6 (FINMA circular); VID mn 33-33.1, pp. 7-8 (FINMA circular); VID mn 38-39, p. 8 (FINMA circular); VID mn 40-41, pp. 8-9 (FINMA circular).

  • AML-18: E-ID validity and holder verification. For identification by E-ID, verify that the E-ID is valid and was issued to the contracting party.

    Sources: VID mn 44.1-44.2, p. 9 (FINMA circular).

Client conduct and tradingApply appropriateness, suitability, disclosure, non-discrimination, order-handling and market-conduct safeguards to AI-supported advice and trading. Preserve the requirements specific to the service and client category.

Client information and advice

  • CON-01: Clear, timely service and product information. Provide the required service, instrument, risk, cost, affiliation and product-universe information and documents before the relevant contract/service, and identify advertising appropriately.

    Sources: FINSA Art. 8(1)-(2) (legislation); FINSA Art. 8(3)-(6) (legislation); FINSA Art. 9 (legislation); FINSO Arts. 7, 9-10, 12 (legislation); CONDUCT mn 23, p. 5 (FINMA circular).

  • CON-03: Identify the advice category. Distinguish transaction advice from portfolio advice and apply the corresponding assessment.

    Sources: FINSA Art. 10 (legislation); CONDUCT mn 4, p. 3 (FINMA circular).

  • CON-04: Client knowledge and suitability assessment. Assess knowledge and experience for the relevant transaction advice and the required suitability/financial information for portfolio advice, using representative investment-category data and proportionate inquiry detail.

    Sources: FINSA Art. 11 (legislation); FINSA Art. 12 (legislation); FINSO Arts. 16-17 (legislation); CONDUCT mn 14, p. 4 (FINMA circular).

  • CON-05: Execution-only exceptions and failed assessments. Use the execution-only exemption only with its required warning, and warn or abstain where the applicable assessment cannot be satisfied.

    Sources: FINSA Art. 13 (legislation); FINSA Art. 14 (legislation).

  • CON-09: Residual and proprietary-product conflicts. Disclose unavoidable residual or proprietary-product conflicts in the required durable form.

    Sources: FINSO Art. 26 (legislation); CONDUCT mn 25, p. 5 (FINMA circular).

Market-conduct and algorithmic trading controls

Fair marketing and AI non-discrimination

Financial risk, lending and valuationApply the specialised requirements for the financial model, lending decision, valuation or risk calculation in which AI is used. Keep model approval, calibration, accounting and capital requirements distinct.

Lending and creditworthiness

  • PRU-07: Product-specific creditworthiness and affordability. Apply the relevant mortgage, consumer-credit, leasing or credit-card creditworthiness/affordability rules, including required information, repayment assumptions, limits and repeat checks.

    Sources: CREDIT Art. 111 (legislation); KKG Arts. 22, 27a (legislation); KKG Art. 28(2)-(5) (legislation); KKG Art. 29 (legislation); KKG Art. 30 (legislation).

Regulated risk and valuation models

Abbreviations used in sources

Core FINMA publications

FINMA guidance has no independent legal effect. Circulars explain how FINMA applies existing law.

The wider framework also includes the Banking Act and Banking Ordinance, the Data Protection Act and its Ordinance, and cyber reporting under the Information Security Act. FINMA has also published guidance on cyber reporting and supervision (05/2020 and 03/2024), operational resilience and digital fraud risks. Further rules on anti-money laundering, financial services and capital requirements may apply to the particular AI use case.

Applicability depends on the institution and its use of AI. Sources checked on 1 November 2026.

Note: AI regulation is evolving. We will monitor developments and update this article as new rules and guidance are published. This includes the federal AI legislative project, for which the Federal Council’s 28 October 2026 update confirms a consultation in spring 2027. The update also includes a guide for companies and associations drafting voluntary AI codes of conduct.