BANKING · INDUSTRY HUB

AI for Swiss banks under FINMA supervision: governance, credit scoring, AML and client triage

How Swiss banks use AI in credit scoring, AML/KYC, sentiment analysis and client mail triage – within FINMA 08/2024, Banking Act Art. 47 and the revised FADP.

Researched & fact-checked by: DuneDive LLC · As of: 2026-05

Swiss banking and AI: an overview

As of early 2026 the Swiss banking sector comprises around 230 FINMA-supervised banks. These include the two large banks, the cantonal banks, the Raiffeisen group, regional banks, private banks and foreign-controlled institutions. Over 100,000 people work in the sector; the Swiss Bankers Association (SBA) and the Association of Swiss Commercial and Investment Banks (VAV) are the central trade bodies.

AI is no longer experimental in this sector. The April 2025 FINMA survey of 400 supervised institutions showed that more than half of responding banks already use AI productively – mainly in client mail triage, AML transaction monitoring, trading sentiment analysis and compliance research. Supervisory Notice 08/2024 from 18 December 2024 spells out the FINMA expectation: AI applications must be inventoried, risk-classified and continuously monitored.

In parallel, discussions continue around the FINMA Ordinance on Operational Risk (OperOps) and the revised Circulars 2024/05 "Risk management when processing critical data" and 2008/21 "Operational risks for banks". Both contain AI-specific control references in 2026. The industry collective bargaining agreement for Swiss banking (AGV Banken) additionally covers the labour-law side of AI use – for example when employee performance data feed into AI models.

The question therefore is no longer "should we use AI" but "how do we govern it so that banking secrecy (Banking Act Art. 47), revised FADP, FINMA 08/2024 and internal risk policies fit together cleanly".

Why a deliberate position is mandatory in 2026

Four pressure points hit every bank simultaneously in 2026.

First: Banking Act Art. 47 banking secrecy. Data on customer relationships are protected by criminal law. A careless prompt to a US-hosted language model with identifiable customer data is not just a data-protection incident but potentially a criminal offence with personal liability of the persons involved. Storage of prompts at the provider without a DPA and a no-training guarantee is similarly critical.

Second: FINMA 08/2024 mandates governance, inventory, data quality and monitoring. A bank in 2026 without an AI inventory, an accountable AI owner at executive level and a risk-classification matrix is no longer aligned with supervisory practice. In a supervisory review this is one of the first items examined.

Third: bias risks in credit scoring and AML. Models trained on historical credit decisions can inherit old skews – for example systematically higher rejection in certain postcodes or for self-employed applicants. FINMA 08/2024 expects operational risks (robustness, correctness, bias, explainability) to be assessed per application. Bias audits are no longer a nice-to-have.

Fourth: employee and customer expectations. In 2026 younger staff expect productive AI tools at work; shadow AI (private ChatGPT accounts used for internal research) spreads uncontrolled. On the customer side, corporate and high-net-worth private clients expect faster first responses to email and phone queries. A bank that ignores both will lose staff to more digitally advanced competitors and mandates to robo-advisors.

The strategic answer is neither "block" nor "open the floodgates" but a controlled platform: multi-LLM gateway with data-classification routing, pseudonymisation before model call, EU/CH hosting with DPA and no-training, audit-grade trail per Art. 957a CO plus FINMA-aligned reporting to the board.

Where AI works productively in a Swiss bank in 2026

Five application clusters cover the bulk of realistically automatable work today. Each is risk-classified under FINMA 08/2024.

Credit scoring with human-in-the-loop. AI models can derive a risk score from application documents, account movements, sector data and external solvency indicators. Important: the score is a recommendation, not a decision. The credit decision itself remains with the credit committee. A bias audit, a documented model description and traceable decision reasoning are mandatory here. For consumer credit, the revised FADP adds access rights and reasoning duties on rejection.

AML and KYC screening. Dynamic checks of new clients against PEP lists, sanctions, negative press and the beneficial-owner transparency register. For ongoing relationships, automated triage of transaction patterns; suspicious cases are escalated to compliance with justification. See AI-GwG-KYC-Screening and GwG-Revision-2026 for duties from mid-2026.

Trading and market sentiment analysis. News and social-media feeds are classified (positive/negative/neutral per ticker or asset class), aggregated and delivered to the investment committee or trading desk. The investment decision itself stays with the human – AI only delivers a synthesis across thousands of sources no human reads in real time.

Client mail triage and pre-qualification. Incoming emails are classified (account matter, card block, mortgage, investment advice, complaint), summarised and linked to the known client file. A draft reply is suggested, the relationship manager decides on dispatch. Routine flows (address change) can be automated with four-eye sampling.

Compliance and policy research with RAG. Internal regulations, FINMA circulars, CISA guidance, AGV banking notices and legal-department memos are indexed into a proprietary knowledge base. Compliance staff can ask "Which due-diligence steps must be documented for a trust with a Liechtenstein trustee?" and receive grounded answers with citations. See Retrieval-Augmented-Generation.

Across all applications: a multi-LLM gateway. Customer data and account movements go exclusively to EU- or CH-hosted models (Mistral Large EU, Anthropic Claude via AWS Frankfurt, local Llama 3.x); general research may go to cheaper US models – always without identifiable customer data in the prompt.

How a bank starts with AI – in 6 steps

01Run an AI inventory: capture all AI applications already deployed or embedded, including those silently in standard software (core banking, CRM, OCR modules, voice bots). Per application: model, provider, data classification, hosting region, DPA status. Without this inventory, FINMA compliance is not possible in 2026.
02Nominate an AI owner at executive level and anchor in the organigram. Define a reporting line to the board, an annual AI status report. For larger banks, run a parallel AI risk committee with compliance, IT security, risk management and legal.
03Draft an internal AI guideline based on FINMA 08/2024, Banking Act Art. 47, revised FADP and AGV Banken. Minimum content: permitted models and hosting regions, pseudonymisation duties, ban on shadow AI, procedure for new applications, audit-trail duties.
04Build a hosting and routing architecture: multi-LLM gateway with routing by data classification. Personal data and client links exclusively to EU/CH-hosted models with DPA and no-training. Local fallback (Llama 3.x, Mistral) for especially sensitive applications.
05Start a first low-risk pilot: compliance research with RAG or client mail triage with pseudonymisation. Clear KPIs (handling time, error rate, staff acceptance). Eight to twelve weeks implementation, three months accompanied production.
06High-risk applications (credit scoring, AML models) only after pilot validation: bias audit, documented model description, periodic validation, model-drift monitoring. Quarterly reporting to the board. Audit trail per Art. 957a CO for every model decision.

Where a bank should start with AI in 2026

FINMA duties shift the sequence. Governance work comes before each productive use case.

Stage 0 – FINMA compliance baseline. AI inventory (including AI silently embedded in standard software), nomination of an executive-level AI owner, risk-classification matrix per FINMA 08/2024, reporting line to the board. This work is non-negotiable and should be completed before any new pilot.

Stage 1 – Light audit and internal guideline. An external stocktake: which software is in use (Avaloq, Finnova, Temenos, in-house core banking), how customer data flow through the network today, which US-hosted tools are already integrated. Output: a report with three pilot candidates and a legal-status assessment. In parallel: an internal AI guideline based on FINMA 08/2024 and Banking Act Art. 47.

Stage 2 – First low-risk pilot. Realistic for a regional or private bank: compliance research with RAG (no customer data, low risk), or client mail triage with pseudonymisation. Eight to twelve weeks of implementation, three months of accompanied production with FINMA-compliant monitoring.

Stage 3 – High-risk applications with own infrastructure. Only after a successful first use case and proven monitoring setup are credit scoring or AML model triage appropriate. These demand bias audits, periodic validation, model-drift monitoring and documented decision foundations – typically with a dedicated in-house team or a qualified managed service.

For smaller private banks or securities firms, shared use of sector-specific solutions (e.g. via association initiatives or specialist FinTech providers with FINMA experience) is a sensible option. Important: each bank remains accountable – outsourcing is not a transfer of responsibility.

Where AI does not belong in a bank in 2026

Three areas where reservation in 2026 is not caution but a supervisory duty.

Autonomous credit decisions without a human. No Swiss supervisory expert in 2026 recommends transferring credit decisions entirely to a model. Bias risk, lack of reasoning capability and the non-delegable accountability of the institution make this supervisorily risky. A score may recommend, a human decides. For consumer credit, the reasoning must additionally be accessible to the rejected customer.

Automatic blocking of client relationships without human confirmation. An AML disclosure to MROS, a freeze of an account or the termination of a business relationship are intrusive measures with banking-secrecy and evidentiary consequences. Triage may be automatic – the intervention itself must originate from an authorised compliance officer with documentation.

Use of employee performance data in non-anonymous models. The banking collective agreement (AGV Banken) and revised FADP rules limit profiling of employees. Models processing individual performance data – for example to identify "top performers" or "risk staff" – require involvement of social partners and are delicate without staff representation. The 2026 EU AI Act classifies workplace-related AI as "high-risk".

Particularly delicate and not finally settled in 2026: fully automated investment recommendations without an advisor (robo-advisor with LLM advisory). FINSA suitability duties and FADP profiling rules interlock here; a fully automated recommendation without documented suitability checks is not advisable.

Trade-offs

STRENGTHS

Client mail triage and draft replies cut handling time noticeably
AML transaction monitoring more systematic than threshold rules alone
Compliance research with RAG brings policy knowledge quickly to staff
Trading sentiment analysis aggregates thousands of sources no human reads in real time
FINMA Supervisory Notice 08/2024 creates a clear frame – the sector can act with structure

WEAKNESSES

Banking Act Art. 47 and FINMA 08/2024 require a careful hosting and routing architecture
Bias risks in credit scoring and AML need continuous monitoring and periodic validation
Governance overhead (inventory, AI owner, risk committee) demands dedicated resources
Shadow AI via private staff accounts is a real data-leak vector
Model drift in production applications forces retraining and revalidation

FAQ

How do Banking Act Art. 47 and AI use fit together?

Banking Act Art. 47 protects customer-relationship data under criminal law. Sharing with third parties – and a cloud model provider is a third party – requires either customer consent or a statutory basis. In practice: customer data go only to models with a data-processing agreement, no-training guarantee, EU or CH hosting and a documented data-flow trail. Pseudonymisation before the model call reduces risk further but does not replace the DPA.

What does FINMA expect in a 2026 supervisory review on AI?

Four points per Supervisory Notice 08/2024: (1) an AI inventory with risk classification; (2) an executive-level AI owner; (3) documented data-quality and bias reviews per application; (4) continuous monitoring including model drift and explainability. For high-risk applications such as credit scoring or AML models, periodic validation, a documented model description and a per-decision audit trail are added. Deviations must be justified – silence is not a position.

May we automate credit decisions if the algorithm is sound?

Even a technically convincing model does not relieve the institution of accountability. FINMA 08/2024 is explicit: responsibility for AI decisions remains with the institution, not with the model provider. The 2026 state of practice: credit scoring is a recommendation, the credit committee decides. For consumer credit, the revised FADP additionally requires a reasoning duty toward the customer – "the AI decided" is sufficient neither in law nor in reputation.

What does the AGV Banken collective agreement say about AI use for staff?

The AGV Banken collective agreement covers labour conditions in Swiss banking and includes provisions on data and personality protection of staff. AI applications processing individual performance data – for example for evaluation, promotion or risk assessment – require involvement of the staff committee and FADP-compliant information of those concerned. Pure productivity tools (an AI assistant for email drafts) are less critical but should still be part of the internal AI guideline.

Sources

FITS YOUR STACK?

What this looks like in your business – a 30-minute intro call.

Book a call