One of the great promises of technology innovation in the last decade has been in the field of artificial intelligence and machine learning, with countless startups attempting to create “the next big thing” in all sectors of society. However, two limitations have created roadblocks for its wide application in the life and other insurance sectors.
The first challenge is accuracy. Modern machine learning models have been developed to tackle highly complex data sets – but for low stakes purposes. Accordingly, they might yield 99% precision. This is often good enough for applications such as image recognition and web search, but for finance and health this is an unacceptable level of error.
The second issue is explainability. Machine learning can be transparent for humans, but often it takes place in a black box. In black box models, humans are generally not able to understand or explain how the machine learning application arrives at its results. In the insurance industry, this lack of transparency creates a serious dilemma for regulatory and legal compliance. This is especially the case when these systems are intended to be used in insurance premium, underwriting and claims assessment settings.
Despite these challenges, there are of course teams around the world trying to find ways to use artificial intelligence to automate tedious and expensive tasks in the world of insurance. In this article we will look at one of these implementations.
Modernising legacy core systems in the life insurance business has been an ongoing activity for the last decades. One of the major hurdles in the process has always been transferring millions of policies and the reimplementation of the associated actuarial functions. Typically, the project requires a so-called migration from one or more source systems to the new target system. This is because the legacy policies come from multiple systems built in previous decades. Such a migration for even a mid-sized portfolio will take 2–3 years and cost 20–30 million euros. It also ties up valuable actuarial resources that could be used for more forward-thinking activities. It is a perfect case for automation.
To understand how migration might be automated, it is first necessary to look more closely at the conventional process. When migrating old policies to a new system, the data sets need to be transferred of course – but also the actuarial functions that are used to manage the policies over time, aligned with the underlying business plans and original agreements with the customer. These functions can be understood as mathematical equations that define a relationship between input variables and an output variable, for example the insurance premium, reserves, benefits and surrender values.
The complex and costly part in migration is transferring these functions, because these are highly individualised. Traditionally actuaries would have to find and analyse the functions from the old source system and try to match them up to functions in the new or target system. If there is no good match, a function from the new system would either need to be adapted, or otherwise a new function must be created. With millions of policies to be migrated, it is easy to see how this can take years of valuable and expensive actuarial time to complete.
msg life has been working on the ambitious idea of automating this process for several years and can now offer a viable commercial implementation for partial automation. It starts again with the actuary looking to match functions. When this is successful, his or her work is done. If there is no match, several automation techniques can be used.
One of the most powerful methods is using deep neural networks, which is a gold standard machine learning model used today. Its strength is based on its ability to work with very large data sets and that it dynamically learns and optimises its results. However, this is a black box process as explained above and there is a problem with explainability.
So instead, we start with some other, less common methods, for example symbolic regression. Symbolic regression is based on the premise of searching through the space of all known mathematical formulas to find the one that best fits the input and output data. As such it is explicitly explainable. While it is not efficient with very complex equations, it can often be suitable for actuarial functions.
Another technique is called neural trees. Neural trees are related to neural networks. However neural trees are deterministic. They are essentially using the structure of logic or decision trees, which most of us are familiar with. Just like symbolic regression, they are not well-suited for the most complex equations, but are often powerful enough to match actuarial functions. Since they are deterministic, they too are explicitly explainable.
This hybrid mix of methodologies addresses the critical problems of explainability, accuracy and the occasional outliers in predictive AI models. msg life has thus achieved a workable automation process for policy migration and is already using it in real-world applications. This is an ongoing project under the working title TRAIL.X (TRustworthy Artificial Intelligence in Life insurance). We are continuously expanding the range of functionality and optimising results. We expect it to become a true game changer for the insurance industry.
Today’s insurers experience increasing pressure to upgrade their IT systems. Yet the daunting financial costs – but even more significantly the investment in time and expert resources – can seem unsurmountable. The msg life automated policy migration system offers a new way to make the big move faster and economically feasible.
The Artificial Intelligence Act – update
The Artificial Intelligence Act – system and regulations
The Artificial Intelligence Act – an overview
Deep learning – between hype and disillusionment
How do artificial neural networks learn?
Machine learning – why does it (sometimes not) work?
AI in the insurance industry – terms and definitions
Table of contents
IF YOU NEED MORE INFORMATION
let us know.
We are happy to help!
Tel: +49 711/9 49 58-0