re:publica 24: Bundesbank provides small insight into AI supervision

Contents

What risks are hidden in the banks' books? This question has occupied banking supervisors for decades - usually with more, sometimes with less success. The Bundesbank is responsible for examining certain risks at smaller German financial institutions - and is increasingly relying on artificial intelligence. But one key requirement is still missing.

Mountains of documents to check

Auditing banks is a complex process: they have to prove to the supervisory authorities that they comply with the applicable rules for banks. These differ depending on the size of the institution. The Bundesbank itself currently supervises around 1,400 banks in Germany. In contrast, the 100 or so largest banks in Europe, known as "significant institutions", are primarily supervised by the European Central Bank (ECB). However, the respective national banks and, in Germany, the Federal Financial Supervisory Authority (Bundesanstalt für Finanzdienstleistungsaufsicht - BaFin) are also involved in the audits.

The core of the supervisory activity is primarily the examination of all documents that are necessary to assess whether a bank is operating on a sound basis. In other words, is it taking excessive risks? How are the risks to be assessed in comparison to the so-called equity capital? To do this, banks must comply with extensive reporting obligations. It is the task of the banking supervisory authority to review these reports, find irregularities and, if necessary, request additional documents or intervene. However, the IT organization in accordance with the regulations for the banking industry is also part of what banks have to regularly prove - all in all: mountains of documents.

This is precisely where the department for digital transformation in banking supervision is hoping to achieve significant efficiency gains through the use of AI models. By 2023, the ECB's banking supervision department had already identified over 40 promising use cases for AI in banking supervision. At the Bundesbank, the main focus is currently on sifting through data volumes, reported Christian Drescher, Head of the Digital Transformation Department at the supervisory authority, at re:publica 24. "The data is not always available in a structured form," he said. According to Drescher, there is potential here to make more documents accessible in a structured way.

Evaluation of press content

The Bundesbank is currently experimenting with a variety of use cases. The tests are not always successful, the economist admits - but there are some promising applications. Among the successes, he counts the improved recognition and automated integration of relevant press content - AI models are already making a lot possible when it comes to topic extraction. According to Drescher, not all relevant processes are first known to the banking supervisory authority and then to the press. A relevant AI application is already being used productively here, which is now to be expanded to include other sources of information such as bank documents.

A large part of the current work is to develop suitable models and applications for the audit processes. For example, a no-code editor is to be made available to bank auditors at the Bundesbank from the middle of the year so that they can integrate modular AI applications into their processes even without their own programming knowledge. This is to be made available via the internal platform for "Text-based Intelligent Assistants" (TIA).

Drescher could not yet say exactly what role the European AI regulation will play for the use of algorithm-driven systems at the Bundesbank. It is not yet in force - but due to its sovereign nature, it is foreseeable that AI can only be used to support banking supervision. Decisions and their justification must always remain comprehensible and be the responsibility of humans, which is also the Bundesbank's guideline for its developments.

Difficult transition to the cloud

Meanwhile, the biggest step towards greater use of AI has yet to be taken: The banking supervisory authority works with strictly confidential material in its audits. So far, only on-premise instances have been used for sensitive data due to IT security regulations. Migration to the cloud is by no means trivial, explains Drescher: so far, it has only been possible to work "with non-bank confidential data" in prototype status. The requirements for more would first have to be negotiated with the providers - this applies equally to the Bundesbank and the supervised banks. In particular, the localization and data security requirements for hyperscalers are still problematic at present.

(mho)