What stands in the way of applying artificial intelligence to the insurance industry?

Artificial intelligence is increasingly prevalent in today's world and offers many opportunities, for a variety of industries, including the insurance industry. Machine learning models are being used to analyze data, identify patterns and make business decisions. However, while the models can help improve business processes and reduce risk, there is still little adoption in the insurance industry. Where does this difficulty come from? In this article, I will discuss four key hurdles that must be overcome in order to successfully use artificial intelligence in the insurance industry.

Understanding

The first key barrier to applying artificial intelligence is understanding it. The problem usually stems from the huge expectations of machine learning models fueled by the actions of IT giants. This creates a situation where we set unattainable requirements for the artificial intelligence models we create at the beginning of a project, and consequently we are met with great disappointment when they are not achieved.

On the other hand, few organizations realize that many business processes, despite the use of artificial intelligence, will still require the control of our employees. Besides, most often the best results are achieved through the cooperation of a human specialist in a particular field and a machine learning model supporting it. The model is able to suggest thousands or even millions of decisions in a short period of time, but the final decision, especially in non-obvious cases, should be left to a human.

Specialists

The second obstacle to machine learning adoption by companies is limited access to machine learning specialists. Seemingly, getting a programmer to build a model that solves the problem of our choice is not that difficult at all. This is the illusion that many organizations succumb to as they try to build teams specializing in data science on their own. However, such a practice is very difficult, expensive and involves a lot of risk.

A better approach is, as a first step, to work with companies that specialize in building and implementing artificial intelligence models. I emphasize the latter for two reasons. First, most models prepared on a laptop are not suitable for production launch, and building a launch environment is a topic for a separate project. Second, to build machine learning models we need a team of specialists with a variety of skills - from data preparation, business analysis, to model training, to their aforementioned production launch. Of course, as time goes on and machine learning develops in the company, it is worth investing in our own teams as well.

Environment

The third problem facing any organization wishing to deploy artificial intelligence is its runtime environment. Well, many models require a lot of computing power, and in turn, the window in which calculations can be made is short - an hour a day, for example. In other cases, the size of the environment will depend on the volume of queries, which will be larger in the middle of the day, for example, and much smaller in the evening or at night.

The above scenarios clearly favor running artificial intelligence models in the cloud, as the cost of purchasing hardware and licensing an environment ready to handle performance peaks will usually be too expensive relative to the ROI of the project. Viewed from this perspective, the problem for Insurance Companies, then, is the low adoption of the cloud, related in part to existing regulations and in part to a reluctance to take the risk of migrating our companies' environments to the cloud.

Explainability

The explainability of machine learning models is their ability to explain what factors and to what extent contributed to a particular decision. In the insurance industry, this is important because decisions based on models often involve premiums.Since decisions in some artificial intelligence models, such as neural networks, are very difficult or even impossible to explain, this limits their applicability not only for regulatory reasons, but also simply for transparency of decisions to customers.

On the other hand, problems with model explainability should not be taken as a red light for implementing such solutions in the industry. First, we can often use simpler models that are also easier to explain. Second, there is a dynamic development of tools to help explain on what basis one decision was made and not another. Add to this any areas of Insurance Companies' operations not subject to such strict regulation, such as marketing or internal process improvements. The explainability of models is therefore an additional obstacle, but not, as many suggest, preventing the development of artificial intelligence in the insurance sector.

Summary

To summarize my discussion, I would like to point out that of the mentioned limitations, the first two are completely universal and apply to all industries, while the next two are specific to the financial industry due to its regulations. Consequently, a noticeable trend in the market is the greater sophistication in the application of machine learning of companies in unregulated industries, such as retail, logistics, manufacturing or e-commerce.

On the other hand, by observing these very industries, we can see with the naked eye how much machine learning algorithms support growth and give an advantage to companies that can use them properly. So it is only our decision whether to focus on what obstacles our insurance company faces in applying artificial intelligence, or simply to overcome these adversities and gain a competitive advantage through the proper use of models.

This article was published in 'Insurance Magazine' No. 12/2023