Dusseldorf. In its new role as an AI market surveillance authority, the financial regulator Bafin wants to increasingly check whether the systems used by insurers and other financial companies contain prohibited practices. “We will not make any compromises here,” said Sebastian Schnitzler, head of the risk modeling and artificial intelligence (AI) department at Bafin, at the Handelsblatt event “AI in Insurance”.
On July 29th, the German law implementing the European AI regulation came into force. Since then, Bafin has been responsible for market surveillance of AI systems that are related to regulated financial activities. For example, so-called “social scoring”, i.e. the classification of people based on their behavior, personal characteristics or certain other characteristics, is prohibited under the AI regulation.
As a rule, insurers have given themselves clear guidelines that the AI systems they develop and use do not carry out such a classification, explained Schnitzler. When implementing specific use cases, however, it is not always entirely clear whether the combination of different data could potentially result in a forbidden classification. Bafin therefore primarily looks at how a system behaves in unusual or critical borderline situations, he said.
After initial data viewing, the financial supervisory authority sees a need for improvement in one point: “We were not satisfied with the variety of data that goes into AI models everywhere,” emphasized Schnitzler.
It was not always clear why sensitive customer data would be required for the purpose of the system. Insurers should prepare for relevant supervisory questions and be able to justify the data appetite of their AI systems - especially where they pass on customer data to third parties.
Ergo board member Mark Klein made it clear that insurers were taking the issue seriously. Departments that want to use certain AI applications at the Düsseldorf insurer would have to have them approved by a central internal office. To do this, they would have to, among other things, complete a comprehensive list of questions.
Consultant Jonathan Larsen, former chief innovation officer at China's Ping An Group, added that insurers could avoid most problems if they trained their AI systems with anonymized data. At the same time, Ergo manager Klein emphasized that regulation should not be an enemy of innovation. Otherwise, Europe will continue to be left behind in the area of AI.
According to its own statement, Bafin also expects an understandable justification for disadvantageous automated decisions, such as the rejection of a benefit application or a risk-based premium increase. But this is not an AI-specific requirement, emphasized Schnitzler: It applies regardless of whether a human, a statistical model or an AI system prepared the decision. AI just makes them harder to fulfill, he said.
Schnitzler sees great potential for how insurers can use AI profitably in process automation, risk assessment and pricing as well as at the interface with the customer. The AI enables shorter response times and more precise information, he said.
Klein from the insurer Ergo, which has so far used AI in customer communication, document management and image and video generation, reported efficiency increases of at least 20 to 30 percent. However, companies must place greater focus on realizing the advantages of AI and scaling use cases.
Bafin expert Schnitzler also emphasized that not all insurers benefit equally from the potential of AI. He said: “If you apply them to bad, fragmented data, you won’t get progress, you’ll get automated chaos.” Those who currently have a clean data infrastructure and clearly structured processes will be able to benefit more from it than companies that still have homework to do.
In his opinion, this will only work if the workforce is on board: “If AI is perceived as an efficiency program with the aim of job cuts, silent resistance is to be expected,” said Schnitzler.
