The Future of AI in Insurance

The future of insurance came into focus at the recent Verisk Insurance Conference London, which hosted a well attended discussion on how artificial intelligence (AI) is reshaping the sector.

Moderated by Verisk Specialty Business Solutions Chief Product Officer Richard Smith, the session explored if and how AI might streamline the unpredictable art of evaluating risk. The panel included Convex Chief Technology Officer Martin Young and AWS insurance expert Arno de Wever, and gave attendees a glimpse into the potential new world of underwriting.

Acceleration

Martin Young kicked off the session by outlining several use cases where AI could provide value at Convex, like summarising documents and extracting data to accelerate underwriting. "Anything that could help us speed that up and make that easier, is going to be very welcome," he said. Of all the potential applications, Young felt that improved data extraction offered the biggest potential impact for the business.

Arno de Wever agreed on the promise of using AI for data extraction, especially for unstructured submissions. "Being able to score a submission turnaround, which now takes a couple of days, to a couple of hours with a customer is key and is the first step to extraction of core attributes from documents," he explained.

With large language models, underwriters could ask questions of submission documents and automatically retrieve key details like the insured, policy period or other relevant details.

Whilst promising, de Wever cautioned that AI still requires human validation. "At this moment, you need to have very strict guardrails of what you want your output from your large language model to do. And you also need to have a human if you want use cases outside this specialty," he advised.

Beyond data extraction, de Wever highlighted other areas where generative AI could assist underwriters, like identifying anomalous risks and prioritising submissions. This could "save so much of the unwanted stuff" currently occupying underwriters' time, he stated.

When asked about security concerns, de Wever emphasised the importance of training models on internal data separate from any foundation model. He also noted that bias remains an issue in AI systems. As an example, de Wever cited a study that analysed political bias in different large language models.

For now, de Wever said conversational AI works best alongside humans, with the AI "whispering" suggestions to human agents. This allows the human to catch any nonsense or bias from the AI before it reaches the customer.

Imperative to develop

Young acknowledged that AI will increasingly penetrate insurance technology, making it imperative for carriers to address potential risks. But he advised starting conversations about comfort levels and boundaries for AI usage, both internally and with regulators.

"Some of the results are unexplainable. With the technology we have now, it may take us a while, but we can always explain how we got to a certain point.”

With AI promising to transform underwriting and other insurance functions, the panellists agreed continued human oversight remains essential.

Companies like Convex are keen to explore AI's possibilities, but are proceeding cautiously given accuracy, security, and ethical concerns. Striking the right balance will allow insurers to realise productivity gains without compromising sound risk selection.