Large language models (LLMs) such as ChatGPT are not properly helping underwriters, leading one cyber expert to suggest underwriters should look at small language models (SLMs) instead.
The insurance industry possesses massive amounts of data subsets, just by the nature of the granular information it collects. And that’s where SLMs can shine compared to their LLM counterparts, says Ed Ventham, co-founder and director of Assured Cyber Ltd., a cyber-exclusive insurance brokerage based in London, U.K.
SLMs are tailored to perform exceptionally well in specific tasks or narrow domains. Because they are trained on targetted datasets, they can achieve higher accuracy in their specialized fields compared to LLMs.
“The small language models, where you’re basically training a bot to understand subsets of data, that’s where we’re going to see a massive uptake…” Ventham said last month during the INFUSE webinar Cyber Underwriting: What’s Next? from insurtech company Send. “I think this will become more and more powerful.
“That is something that is being used and…will continually spin up as we go forward, which is really, really exciting.”
But that’s not to say LLMs can’t be useful, says Helen Rios, underwriting director at U.K.-based Ki Insurance. “We’ve actually got a team who focuses specifically on use cases for large language models, and we’ve had some real successes,” she said during the webinar.
But she warned that LLMs need to be used in a very controlled framework; otherwise, they won’t provide the needed output. “You can get a level of confidence score out of a large language model, so what’s that threshold that you’re comfortable with?”
Related: Is AI a threat to underwriting talent?
Ki Insurance, which describes itself as Lloyd’s first fully digital and algorithmically powered syndicate, uses an algorithm for the majority of its underwriting. But the algorithm needs data to be structured to ingest it and provide an output.
“The real challenge that we have within the insurance industry in general is that there’s vast amounts of data out there,” Rios says. “It’s just that it’s not in a structured, useable manner.
“We can use some of this tooling to create more structured data, and that data will then feed into systems that help augment our view of risk.”
As the industry learns and enhances its AI skills, this should also bring new talent into the narrowing pipeline of underwriting talent, adds Tammy Kocher, vice president and head of cyber underwriting at U.S.-based MGA Millennial Specialty Insurance (now branded MSI). “While you have a little bit of a skill gap, I think once we catch up to that, you might get talent that’s wider than our traditional underwriting trajectory.”
Rios agrees a skills gap exists in the insurance industry as well as a bit of a reluctance to embrace some of these technologies.
“There is a gap in the knowledge on the tech side here, but the danger is that we move too far over to that side, and we lose sight of the commercial side of insurance,” Ventham warns.
No matter what the data tell you, it won’t express the emotional decision going into a customer’s security technology procurement journey.
“That’s so important when you’re actually evaluating what a risk is,” Ventham says. “And I think that’s something that cannot be lost, because that’s the art of insurance.”
Feature image by iStock.com/GamePH
