How can insurers realise the true value of AI?

As artificial intelligence (AI) and digital transformation find their way into every aspect of our daily lives, we are gradually seeing changes taking place in different sectors.

Progressively, AI is permeating the insurance value chain and it is set to significantly transform the industry. In product development, AI is enabling insurers to create more profitable and effective products based on insights from past claims and product uptake in the market.

Underwriters are also using AI to assist in creating a better understanding of risk for new and underserved markets.

Also, with the integration of conversational interfaces, there are improvements to customer service and advances in fraud detection are enabling efficient claims processing. 

Yet, despite these AI transformations, the potential of AI has yet to be fully realised. For now, AI’s role in the insurance industry is largely limited to optimising existing business processes rather than developing new and disruptive business models.

There are a few key reasons for this:

Insurance companies are currently integrating applications to use machine learning/deep learning (ML/DL) models to gather insights from the enormous amount of data they generate.

However, the increasing complexity of ML/DL models requires enormous amounts of compute power. According to research from OpenAI, the computing power required over the past few years has increased 300,000 times between 2012 and 2018.

At the moment, only a few niche technology companies have the skilled data scientists required to develop complex models, enormous datasets required to train these models, infrastructures that are required to deploy these models at scale.

The problem is not the lack of data in the industry, but that insurance entities are struggling with an enormous amount of data that isn’t AI-ready.

Data needs to be cleaned, integrated, moved to appropriate infrastructure, governed and managed continuously.

It also has to be labelled correctly for accurate decision making; this labelling process is time-consuming and expensive. Further, negative data is not easily available to train ML/DL models in failure scenarios.

For example, you would never send a fleet of self-driving cars out on a mission to crash on purpose just to help an AI decide what went wrong at a crash scene.

For industries such as insurance that operate in strict regulatory environments, the opaqueness of these models is an issue.