AI governance for German banks, insurers and financial firms

The short answer

German banks, insurers and other financial firms do not have an AI rulebook of their own. They govern AI under three sets of rules that already apply to them, each now with a clear German reading.

• The EU AI Act, which BaFin monitors for AI in their regulated business under the KI-MIG.
• DORA, whose ICT risk requirements BaFin has applied to AI in its guidance on ICT risks of AI.
• The GDPR and the German Federal Data Protection Act, which from 20 November 2026 add section 37a on scoring.

BaFin’s role under the AI Act

Section 2(3) of the KI-MIG makes BaFin the market surveillance authority for AI systems used in direct connection with regulated financial activities by a list of financial entities, including credit institutions, insurers, pension funds and crypto-asset service providers. It took up the role on 29 July 2026.

In a July 2026 interview, BaFin described its mandate under the AI Act as monitoring rather than prudential supervision. For now it is checking the transparency duties, the prohibited practices and whether firms promote AI literacy, taking a risk-based approach that looks at widely used applications rather than every system at every firm. From December 2027 it adds the high-risk systems most relevant to the sector.

• Banks and lenders. AI used to assess the creditworthiness and credit scores of natural persons.
• Insurers. AI used for risk assessment and pricing in life and health insurance, such as calculating individual surcharges.

AI that is not tied to a regulated activity is not BaFin’s under the AI Act. Recruitment and HR systems at a bank, for example, are supervised by the Bundesnetzagentur. Classifying systems correctly is the first step, and the high-risk classification guide covers how.

What BaFin’s DORA guidance asks

BaFin published its guidance on ICT risks in the use of AI on 18 December 2025. It is non-binding and does not set supervisory expectations, but it shows how BaFin reads DORA for AI. The main points for a firm are these.

• AI sits inside the ICT risk framework. AI systems and their components, including training data sets, models and libraries, are identified, classified and monitored as ICT assets.
• Governance starts at the top. Many firms adopt an AI strategy approved by the management body, which keeps ultimate responsibility for ICT risk and needs the knowledge to understand it.
• Testing is proportionate. Depending on criticality, testing can include adversarial attacks, AI-specific penetration testing and stress tests. BaFin flags unannounced model changes by third-party providers as a particular testing challenge.
• Third parties are mapped. For AI supporting critical or important functions, firms keep track of every party involved in processing, including subcontractors such as model providers and GPU capacity, and agree clear service levels.
• Operations are logged. Risk-based logging of AI decisions, model versions and training data, thresholds for misbehaviour and inclusion in business continuity plans.

The scoring rules from November 2026

The law implementing the EU consumer credit directive, published in May 2026, rewrites the BDSG’s scoring rules from 20 November 2026. Section 37a bars scores built on special category data, age, gender, name, social network data, bank account flows or address data, and excludes minors. On request, the firm that creates a score must explain the data and criteria used and the weighting of the factors that most influence it. A person can contest a score-based decision and have a natural person decide.

For consumer credit, the new section 30 BDSG also gives borrowers whose creditworthiness assessment involves automated processing the right to ask for human intervention, an explanation of the assessment and its logic, and a review of the lending decision. A lender that relies on an AI model needs to be able to provide that explanation and route the review to a person. The BDSG section 37a page covers the detail.

A working approach

1. Know what you run. Register every AI system and model, including generative AI tools staff use and AI inside vendor products, with the provider each one depends on.
2. Classify and assign the authority. Mark credit scoring and life and health insurance pricing systems as high-risk from December 2027, and note whether BaFin or the Bundesnetzagentur supervises each system.
3. Check scoring models against section 37a. Confirm the data each model uses, that it is not used for minors, and that you can explain the main factors and offer a human decision.
4. Bring AI into DORA. Treat models and AI providers as ICT assets and third parties, with testing, logging and exit plans scaled to criticality. The vendor AI risk guide sets out the steps.
5. Report to the board. Keep the AI strategy and its risks in front of the management body. The board reporting guide covers what to include.

Aicura supports this work with its AI register, vendor management, risk, bias and impact assessments, security testing and monitoring. It assesses against the EU AI Act from the Essentials plan, and ISO/IEC 42001 and the NIST AI RMF from Pro. The decisions, and the dialogue with BaFin and your data protection authority, stay with you.