Speaking your language
If you want to understand the importance of AI, just ask ChatGPT, advised Pranav Pasricha, chief executive officer of BluePond.AI. Speaking at Intelligent Insurer’s Re/insurance Outlook USA 2024 conference in New York today (May 2), Pasricha began his presentation by summarising ChatGPT’s response.
As he explained, though, the real answer may be in the name – ChatGPT: GPT for “Generative Pre-trained Transformer”.
Pasricha detailed how he had worked with machine learning for commercial underwriting since 2013. In the past, however, it was a laborious process.
“It took us ages to train models line by line, class by class, underwriting question by underwriting question, claims value by claims value. It was a pain – a death by a thousand cuts,” he explained.
GPT changes that. “It takes 80 to 90 per cent of that workload off us. We do not have to train language models, because it understands English.” This is combined with the ability to understand information in its broad context, and synthesise and analyse language to spot different expressions of the same ideas.
“That’s a massive thing.”
Of course, challenges remain. “The one thing it doesn’t do is understand insurance,” said Pasricha.
To be useful, AI must be able to identify documents, find relevant parts, extract the insurance-specific information, and compare coverage. That is what BluePond set out to tackle and what GPT technology has unlocked.
“We could not have done this two years ago. Now we can.”
Such technology can achieve what was always hoped for from data standardisation but never happened. “It isn’t happening because a thousand people sending stuff to a thousand people are not going to change their IT systems,” he said.
AI, however, can bridge the gap.
“We have trained a huge model by ingesting lots of insurance documents to understand all the values, fields, dates, clauses, conditions, exclusions and endorsements,” he explained. It can do so even with complex policies with little standardisation, such as cyber coverage.
Crucially, with generative AI, it is possible to compare not just terms but language. “We've never been able to do this. Without gen AI, it was simply not possible to compare how the policy forms of two carriers differed on the underwriting conditions, exclusions, and endorsements,” he explained.
“You could extract values painfully and with a lot of errors, but you could not do language. Now you can.”
Did you get value from this story? Sign up to our free daily newsletters and get stories like this sent straight to your inbox.
