Suggestions
Chad Hersh
Head of Worldwide Business and Market Development, Life Insurance Industry at Amazon Web Services (AWS)
Chad Hersh is a seasoned professional with a wealth of experience in the insurance and consulting industries.
Currently leading the L&A/Group Benefits practice for Capgemini in North America, Mr. Hersh has held executive positions at renowned companies such as Majesco and Amazon Web Services.
With a background in IT and management consulting, he has contributed significantly to the strategic direction and growth of various organizations.
Mr. Hersh's expertise spans areas such as e-business, core systems strategies, emerging technologies, and risk mitigation strategies.
He is a sought-after speaker at industry events and has authored numerous articles in insurance trade publications.
His educational background includes a BA in Economics and Managerial Studies, a Master's degree in Accounting and MIS, and professional development at institutions such as Rice University and Columbus Academy.
Highlights
Are we already at that point? I’ve been talking for the last few months—very hypothetically—about the future when Generative AI would be able to pass the licensing exams for insurance agents. To benchmark how far off that might be, I fed Anthropic’s Claude Sonnet 3.5 the first 60 or so questions of a practice test for the Texas life insurance agent exam. Now, full disclosure, it was a very poorly written test, and several of the questions were either simply missing words or were too ambiguous to definitively get a correct answer. Despite this, on a test that 53% fail on their first try—and an even higher percentage fail when repeating—and for which a passing grade is 70%, Claude got about 90% of the questions right (I say “about” because there was one question that was missing some words, which it understandably missed, and another that was poorly worded…but if we count those as wrong, it’s still over 90%; it might be closer to 95% if we cut it some slack).
Now, it did miss another question that I’m not convinced the “correct” answer was correct for, and maybe somebody reading this can tell me whether disability have the same two year contestability period as life insurance…the practice test says it does, but I’m not actually sure that’s right. As a result, Claude was mortified that it later answered that life insurance has a one year contestability period (actual quote from Claude: “I apologize for this error. It's particularly egregious because we had just discussed the two-year contestability period in relation to another question,” and it then went on to lecture me about the fallibility of AI, the importance of verifying these things with legal counsel, etc. which I found to be rather impressive!).
All that said, it’s very clear Claude could have easily passed the licensing exam with flying colors. It was also clear that its reasoning for the answers it gave was at least on par with a typical agent’s. And most importantly, this was not some fine-tuned, insurance-trained model. This was out-of-the-box Claude 3.5 Sonnet, with zero context provided beyond the initial prompt, and no regulations uploaded to supplement its knowledge. I believe that had I fed it the Texas insurance regulations it would have done even better.
That brings us to the point: Should a solution like Claude Sonnet 3.5 be allowed to provide the same kinds of advice an agent would give? Obviously it would need to be given proper system prompts and guardrails. It would need to be trained on carrier-specific products, and someone would need to provide it regular updates on laws and regulations. It would need to be given carrier guidelines around suitability, ethical practices, etc. But all of that is quite manageable. Will it still hallucinate? Absolutely. Will that hallucination likely be more than offset by its ability to access more-than-encyclopedic knowledge. Perhaps.
It certainly seems that we’re fast approaching a tipping point. Thoughts welcome in the commments!
I had the pleasure of speaking with nearly a dozen insurance CXOs--mostly CEOs with a few Chief Digital Officers--and I came away with a realization of sorts. While I often speak/write about the potential, risks, and rewards of Gen AI for insurers, I don't spend enough time talking about the most pragmatic bits. One of the group challenged me on "when is it the right time to build versus working on the big ideas." This was a fantastic point, and one with a fairly straightforward answer (that I'll try to articulate better here than I did without organizing my thoughts in person!).
Gen AI is in its infancy. Nobody--literally nobody--has much more than a one year headstart on anyone else, and for most use cases, if you see something that someone else builds and you want to mimic it, improve on it, etc., the build time is often weeks, maybe months, but never years. That means that today you can start building out all of the relatively simple, high value use cases (document ingestion and summarization for claims and/or underwriting, code assistant, agent chatbot, etc.), while working on developing your vision for (and organizing your data for) "big ideas."
So when do you build those big ideas? Simple...try building them today. If it's not reasonably straightforward to do now, wait a couple of months and try again. Gen AI and all of the tools being built around it are evolving incredibly rapidly, and nobody knows what things will look like 2 years or 5 years down the road. Jeff Bezos is fond of saying that you know with certainty that 10 years from now customers won't want higher prices or slower shipping, so you're always able to work on improving those things. Similarly, if you build things today that improve agent/broker and customer service, provide more accurate pricing/underwriting, reduce operating costs, etc. there's a very good chance you won't regret it. Other than Gen AI model builders, can you point to a company that has achieved a sustainable advantage with Gen AI? Of course not...it hasn't been around (practically speaking) for two years yet, so at best we can only speculate about what's sustainable. But a sustainable competitive advantage isn't inherently better about changes that benefit an entire industry through, for example, a better cost structure.
Lots of teams win championships with lots of base hits and not a ton of home runs; others win almost entirely with defense. It might not be as glamorous, but it gets the job done and done well. Similarly, if you leverage Generative AI now to solve challenges that have existed for decades, you might not win an innovation award, but you'll get bottom line results. Pragmatic use of Gen AI may not be a panacea, but it can certainly be a winning strategy. Of course, that doesn't mean you shouldn't think big...it just means you can work on the fundamentals while planning for your next big thing.
I look forward to your comments...I'm sure there will be mixed feelings about this approach!
