What does BaFin guidance on ICT risks in the use of AI say about AI in banking?

What are the key points of BaFin guidance on ICT risks in the use of AI?

• Non-mandatory advice to help financial entities implement DORA when using AI; based on discussions with financial entities; 'does not constitute a binding interpretation of DORA' and 'does not define any supervisory expectations'. Described as a living document.
• Scope: entities subject to the ICT risk management requirements in Articles 5 to 15 DORA, with ICT third-party risk management, the Delegated Regulation (EU) 2024/1774 (RTS RMF) and the Delegated Regulation (EU) 2025/532 (RTS Subcontracting); Article 16 DORA entities are excluded.
• Definition: AI system as in Article 3(1) of the EU AI Act, understood for this guidance as a combination of ICT assets and ICT infrastructure in which a complex mathematical model (itself an ICT asset) is implemented; the guidance does not address autonomy, adaptiveness, modelling methodology or data validation.
• Proportionality (Article 4 DORA): AI used in critical or important functions needs more extensive security and control measures than, for example, a human-supervised self-service assistant.
• Chapter II: AI risk is assessed like other ICT risk; management body responsibility under Article 5(2)(a) and 5(4) DORA; AI systems belong in the ICT risk management framework (Articles 6, 8 and 9), which must be reviewed at least annually (Article 6(5)).
• Chapters III and IV: apply standard software development, testing (RTS RMF Articles 16 and 17), asset identification (Article 8(4) DORA with RTS RMF Articles 4 and 5), monitoring and logging (Article 10), business continuity and recovery (Articles 11 and 12), and secure retirement of models and data; examples include adversarial testing, model-drift monitoring and versioning of model artefacts.
• Cloud and third parties: risk assessment and due diligence (Article 28), subcontracting transparency (Article 30(2)), SLAs for critical or important functions (Article 30(3)), audit rights, exit strategies and portability including models, training data and configuration (Article 28(7) and (8)); references BaFin's and the Bundesbank's cloud outsourcing statement.
• Chapter V and annex: cybersecurity and data security across the lifecycle, and reporting of major ICT-related incidents under Article 19 DORA; the annex case study of an LLM-based assistant (on-premise, own cloud tenant, or provider outside the tenant) is illustrative and expresses no supervisory expectations.

What did BaFin guidance on ICT risks in the use of AI change for banks?

Before this guidance, BaFin's AI papers addressed model governance (2021, 2022). This is its first guidance linking AI directly to DORA: it tells banks that AI systems — including LLM assistants embedded in standard software and AI-generated code — are ICT assets whose lifecycle, third-party dependencies and incidents fall under existing DORA duties, without adding new legal obligations.

What does BaFin's guidance on ICT risks in the use of AI require of banks under DORA?

BaFin's guidance does not create new requirements: it explains how existing DORA duties apply to AI systems. A bank's management body remains ultimately responsible for ICT risk (Article 5 DORA), AI systems must be identified, classified and managed as ICT assets inside the ICT risk management framework (Articles 6 to 11), and the framework must be reviewed at least annually. AI used in critical or important functions requires more extensive controls under the proportionality principle (Article 4). The guidance applies the same logic to development and testing, operation and retirement, third-party and cloud use (Articles 28 to 30), cybersecurity and reporting of major ICT-related incidents (Article 19). It is non-mandatory advice and, in BaFin's words, does not define supervisory expectations.

The guidance is an interpretation aid from BaFin's CTF 5 division. It reuses the AI Act's definition of an AI system but deliberately limits itself to ICT risks, leaving modelling methodology, data validation, autonomy and adaptiveness to other regimes such as the EU AI Act and BaFin's model-risk papers. It therefore complements the 2021 BDAI principles and 2022 results paper on model governance.

For supervision, BaFin signals no new examination standard: the requirements it describes derive from DORA, the RTS on ICT risk management and the RTS on subcontracting, and the guidance's own text says it does not define supervisory expectations. Its case study on an LLM-based assistant is explicitly illustrative.

At EU level, the ESAs' July 2026 statement on ICT risks from frontier AI models (JC 2026 25) and the ECB's work cover related ground; they are tracked separately on this site.

WHAT THIS MEANS IN PRACTICE

• Add AI systems, models, training data sets and libraries to the DORA register and asset inventory with criticality classification.
• Include AI incidents, model drift and data poisoning scenarios in ICT incident management, testing and continuity plans.
• Re-run third-party due diligence for AI services consumed via API or embedded in standard software, including subcontractors and data locations.
• Plan exit and portability for AI services supporting critical or important functions, including export formats for models and training data.
• Apply the same change management, code review and testing to AI-generated code and to AI built outside the ICT function.