I hear this question quite frequently because people typically have no direct interaction with a reinsurer. This presents a challenge when it comes to recruitment. People are familiar with technology companies and large primary insurance providers because they interact with their products, insurance offerings, and services. However, as a reinsurer — essentially the insurer of insurers — our clients are insurance companies, and we operate exclusively in the B2B space. This creates a different set of challenges compared to a primary insurer.

For a primary insurer, optimizing client interactions and customer journeys is a major focus. They prioritize engagement and touchpoints with policyholders. However, these elements are largely irrelevant for a reinsurance company.

You could say that while a primary insurer manages a high frequency of risk, handling a high volume of life, health, and motor insurance policies, a reinsurer deals with a smaller number of more complex risk structures. This distinction has significant implications for how we approach data, AI, and analytics.

We invest a lot of time in understanding causality rather than being exclusively triggered by correlation. This has several implications. It means our data is highly heterogeneous. When dealing with lower-frequency heterogeneous risks, as opposed to high-frequency homogeneous risks, the data landscape naturally becomes more diverse.

Also, a significant portion of our data is unstructured. Complex risk evaluations often involve lengthy, detailed textual documents, such as loss reports. This is quite different from the structured data typically associated with a simple car accident.

Since we underwrite complex risks, we have a unique need to merge data science expertise with deep subject matter knowledge. Our underwriters may be engineers or physicians, so we must ensure they are equipped with basic data science abilities. A data scientist, on the other hand, typically lacks the necessary subject matter expertise.

One of our key challenges is to intelligently integrate the domain expertise Munich Re has developed over more than a century with modern data science capabilities.

In primary insurance, the focus is often on efficiency, client interaction, and risk selection. However, in our domain, when we talk about data and AI, our primary emphasis is on the underwriting decision, ensuring the quality and accuracy of our decisions rather than solely optimizing for efficiency.

The insurance industry has a long history of creating risk models based primarily on internal structured data. Decades ago, we began incorporating structured external data, which can be purchased for attributes such as building characteristics and company data sets.

The latest advancements in AI, particularly in language models, now open up new possibilities for leveraging internal unstructured data. The insurance industry has started with simple methods like keyword extraction. For example, extracting a diagnosis code from a medical record because a diagnosis code followed by International Classification of Diseases (ICD) coding has a certain structure. This allows us to determine the diagnosis from the code itself.

These techniques were in use long before LLMs emerged. The insurance industry has had a tradition of analyzing structured internal data and has also integrated structured external data for purposes such as marketing campaigns. In recent years, it has adopted natural language processing and techniques like keyword extraction to extract insights from internal unstructured data, which was previously difficult to utilize.

Much of the information that remains unstructured and external, such as data available on the web, presents new opportunities. When considering our business model, a key focus is on leveraging these types of information sources to improve risk assessment.

I don’t believe that AI will be capable of making decisions in the very near future, at least in the reinsurance industry. At the moment, AI is positioned to support experts and underwriters. The final decision in our business is made by a human — by the underwriter. AI can provide signals and help optimize certain steps, such as wording comparisons, which could become very powerful.

AI, especially generative AI (GenAI) and large language models (LLMs), is still in its early stages and will improve over time.

We are already seeing developments like agent-based architectures, which may allow us to successfully combine the broad but unspecific knowledge of LLMs with more precise information sources in the future. LLMs know a lot, but they lack precision.

When it comes to mitigating issues like incorrect information, a key approach is to integrate smaller, more accurate information sources — such as company-specific data or rule-based systems that have been used as expert systems for decades.

One of the key challenges will be how to effectively combine these different data sources. This means that architectures for integrating structured and unstructured data, both internally and externally, need to become smarter.

We have been using data and AI for decades, particularly in medical underwriting for life and health businesses. The focus is on how we transition from structured data to making decisions or recommendations for medical underwriters. Additionally, technologies like LLMs can help extract more relevant medical information from medical records.

For non-life and health, there is a more classical focus on wording. LLMs can already understand parts of a contract, and in the future, they may even be able to process entire contracts.

In reinsurance, contract wordings are highly tailored, varying from one contract to another. However, they must still meet specific criteria typically outlined in the guidelines. Comparing contract wordings against a broad set of relevant guidelines is another key use case.