Hartford Business Journal

22 HARTFORDBUSINESS.COM | OCTOBER 16, 2023 AI In CT Dan Campany is vice president of innovation at The Hartford, which is using data from sensors to look for ways to prevent losses ranging from car accidents, worker injuries and property damage. HBJ PHOTO | STEVE LASCHEVER Risk Assessment Insurers spends billions to deploy generative AI systems, departments, while keeping humans 'in the loop' By Andrew Larson alarson@hartfordbusiness.com T he impact of generative AI on insurance companies, oper- ating in one of the most highly regulated industries, is expected to be profound, but the risk-averse industry is proceeding with caution. Property and casualty insurer The Travelers Cos., which has a large presence in Hartford, plans to spend $1.5 billion on the technology this year, much of it dedicated to research and development. Like other insurers, it's looking to incorporate AI into many aspects of its business, from customer service and claims processing, to underwriting. An overarching belief among insurance executives is that while technology can't replace humans — it requires human oversight to func- tion — generative AI is capable of performing certain tasks more quickly and effectively than people, and it can help employees become more efficient in serving customers' needs. AI can accelerate the process of risk evaluation and underwriting, and can even generate and submit compliance reports, said Dan Romuald Mbanga, director of generative AI solutions at Google, keynote speaker at the Insurance Capital Summit, held Sept. 27 at The Bushnell in Hartford. "It's been applied in many forms and ways that shape the industry. And generally, it really changes the game in the entire ecosystem," said Mbanga, who works with Google DeepMind researchers and Google Cloud engineers to design and create generative AI programs. While AI has existed for years, generative AI — which Mbanga described as using human-like capa- bilities to generate content — is still being developed, tested and refined. Creating efficiencies Insurers, which have troves of data spread across numerous departments, are using AI to help sort through and analyze that information quickly. AI can help employees perform specific queries with better speed and accuracy. Recently, Travelers developed, and is now piloting, a generative AI claims knowledge assistant tool. It's trained in proprietary, technical source material that was previously only accessible in thousands of different documents. Claims professionals use the tool to quickly access information during interactions with customers and other partners. Also, Travelers is using an AI system that "ingests" legal complaints filed against insured parties, high- lights liability and coverage issues, and even helps recommend appro- priate defense counsel. Dan Campany, vice president of innovation at The Hartford, said AI can be especially useful in gleaning insights from datasets, as long as the right guardrails are in place. "Generative AI is not deci- sion-making," Campany said. "Generative AI does not state or imply logic in the same way that a human being does. What it is useful for is helping humans become more efficient. Instead of having to create content from scratch, you get a bit of a head start." The innovation team Campany leads takes data from sensors to look for ways to prevent losses ranging from car accidents, worker injuries and property damage. "We capture a lot of data from those sensors, then we put it in the cloud and it's all anonymized …," Campany said. "And we analyze and look for trends to predict where and when (losses) will happen." He imagines a day when he can walk into his office in the morning and ask generative AI to answer a question, and then use the informa- Dan Romuald Mbanga As insurers embrace AI, they must also monitor 'hallucinations' By Andrew Larson alarson@hartfordbusiness.com W hile insurers are using generative AI to become more efficient, they're also accounting for new risks of error. AI may make more mistakes than humans in performing certain tasks. Paradoxically, as AI improves, it can become smarter in some ways, but make more mistakes in others. Christopher Sirota, product manager for emerging issues and innovation at Verisk, an analytics and risk assessment firm based in New Jersey, said "drift" occurs when AI generates different answers to the same question, including answers that are inaccurate. For example, the new version of ChatGPT 4 ostensibly "knows" more than its predecessor, ChatGPT 3.5. But the older version is better at certain tasks. "ChatGPT 3.5 is actually getting a little smarter for some things, but not so smart for others, and the same is true for ChatGPT 4," Sirota said. Insurers must be aware of AI's tendency to make mistakes, in order to account for that risk, along with the need for humans to monitor AI for hallucinations, Sirota said. Human oversight remains an important component of emerging generative AI. A combination of physics, human oversight and machine learning have resulted in generative AI that is capable of following instructions, showing human-like reasoning skills. The key to using AI effectively involves a human asking the right questions and giving it the right prompts, using language it understands. "How do you know exactly how to ask the ChatGPT or generative AI to do those things? What are the personas you have to train it upon? What is the corpus of data that you have to train upon? Those are the key areas that we're seeing orga- nizations starting to look at," said Doug Vargo, who leads the U.S. emerging technologies practice at CGI, an IT and business consulting firm with a downtown Hartford office. But keeping humans involved can also mean removing their innate biases, which can become embedded in AI language models and the data it is trained to scour. Vargo said there will be a heavy focus on transparency and ethics, along with the trustwor- thiness of data, as generative AI is implemented.

• Cover