Trusted AI in the insurance industry

Minimizing internal and external risks with digital ethics

AI has emerged as the key enabler for automated decision-making. Much is already technically possible, but there has been little regulation to date. Now that the EU has formulated guidelines for trustworthy AI, BaFin and EIOPA are following suit in the insurance environment and formulating their own approaches to digital ethics and regulatory requirements. At the same time, the pressure of expectations on the part of the insured is increasing – they expect decisions to be transparent and comprehensible, especially if they are automated. Insurers themselves also want to understand and comprehend their internal AI applications better, for instance to safeguard themselves against liability issues. The importance of digital ethics and compliance is increasing, which raises the issue of manageable solutions.

A framework for trustworthy AI in the insurance environment

Insurers are advised to follow EIOPA’s framework to make their own AI systems trustworthy, traceable and verifiable. This has adapted the EU’s general guidelines to the insurance industry and is guided by the following principles:

Proportionality
Fairness & non-discrimination
Transparency & explainability
Human oversight
Data governance
Robustness & performance

Using these principles, existing AI applications can be transitioned to the new framework of trustworthy AI. This takes into account the requirements of BaFin and DAV, meets current expectations of policyholders and, last but not least, reduces internal risks by making AI decisions transparent.

Technical and procedural approaches to harmonization

In a specially developed workshop format, the AI Use Case Impact Assessment, which Comma Soft has already conducted jointly with numerous clients, the AI model is assessed with a view to internal compliance requirements and technical guidelines that must be adhered to. The key is to maintain a sense of proportion and develop tailored and appropriate measures for the AI application. This may include technical adjustments in the operational AI environment (keyword algorithms, explainable AI and MLOps) on the one hand and procedural approaches to harmonize the internal control system with the new framework on the other.