Join top executives in San Francisco on July 11-12, to hear how leaders are integrating and optimizing AI investments for success. Learn More
Online life insurance agency Haven Life announced an upgrade to its service earlier this week that could let you buy a life insurance policy in just a few minutes after answering about 30 questions — possibly with no medical checkup. That speed of underwriting is made possible by artificial intelligence the two-year-old company has built into its offering.
The company is one example of a new breed of fast-moving businesses built atop AI. It’s difficult to get a good idea of the types of machine learning these companies are building into their products, as they’re giving very little away at this stage. But they’re constantly iterating: testing out new machine-run decision-making processes and then migrating the next part of their business over to those systems once they prove secure.
Haven Life’s upgrade this week is a good example of that kind of iteration. A young, healthy applicant can now get a final decision within about five minutes, no medical or labs required, and buy a policy on the spot. But the company’s underwriting process still isn’t fully automated for all applicants. More complicated cases still need to be looked at by a human underwriter behind the scenes and may require a medical exam. “We’re continuing to lower that percentage [of humans in the loop]” by reassessing the data and rules the service’s AI relies on, CTO Todd Rodgers told VentureBeat.
While fully owned by traditonal life insurance company MassMutual, Haven Life is very much living on the frontier of machine learning. It’s a life of constant rethinking and adaptation, and it requires a new way of conducting business, where the speed of AI advancements dictate the speed of business decisions. “You have to have an appetite for iteration” to succeed with this type of business, Rodgers said. “Going into development with the understanding you’ll never be done is critical.”
Join us in San Francisco on July 11-12, where top executives will share how they have integrated and optimized AI investments for success and avoided common pitfalls.
And it takes a certain kind of team. “To pull this off, you need people with deep, deep subject matter expertise, you need developers who are interested in the business problem they’re solving, and you need analysts who can intermediate between the two, who can explain an objective in a way that developers know how to codify it,” said Rodgers. He said Haven Life has been lucky to have subject matter experts who are close enough to the data and code that they don’t need the intermediaries. “You also need to be colocated,” Rodgers said, “since you’ll be working together day in day out” on a project that is never finished.
When the company first launched in 2015, it started to pull in third-party data, such as prescription histories, motor vehicle records, and information from the Medical Information Bureau, to assess appropriate life insurance quotes. That didn’t make the company unique. This is data that’s commonly used across the industry, Rodgers said. Where Haven Life innovated was in gradually restructuring that data so that it could be sorted in a highly structured way, allowing the company to run increasingly sophisticated rules and algorithms against it. The more patterns the company has been able to surface in the data, the more rules it has been able to feed to its AI.
With the new update the company announced this week, it is finally cutting over to using the AI output as the primary determining factor for the policy rate. “Before now we’ve always run the algorithm in parallel with a human looking at this too,” said Rodgers. He underscored, though, that human underwriters are still involved in some percentage of cases.
There are seven prescription drugs the company has been able to develop advanced rules for to date. “Each drug takes a lot of analysis. We’d like to expand that out to a larger number of drugs, since each one we add means one more case that we can immediately underwrite [without human involvement],” said Rodgers.
“We’ve gotten to the point where we can ask follow-up questions while the client is still online in the application process based on prescription drug data we’re pulling in in real time,” said Rodgers. “As clients are answering questions, we realize what new information we need and we refine our questions accordingly.” The goal is to optimize on the amount of information the company can glean with just 30 questions.
Haven Life has built its AI in-house over a two-year process of looking at data, consulting with MassMutual’s doctors and its own team of actuaries, and building out rules. It also had the benefit of MassMutual’s data science team and the parent company’s historical data — some 1 million policies going back about 15 years.
When asked why the company hasn’t leveraged any third-party AI tools, Rodgers said, “There are tools you can use. What we’re doing, though, is pretty innovative, so there’s nothing on the market that does exactly what we’re doing. We’re very focused on the ability to be agile and flexible, and a lot of that ability is due to the fact that we’ve built from scratch.”
The reasons ready-built products are something not a good fit? “If you work with a third-party product, I think you’d probably be guided by that product’s view of the world, and it could narrow down your focus,” he said.
Rodgers offered the following advice to other companies deploying advanced machine learning: “If there’s a tool out there that works for you, great. But if not, don’t contort a tool that wasn’t meant for you. And don’t be afraid to build from scratch.”
Haven Life’s dev team, by the way, makes up about 60 of the company’s 110 employees.
We’ve asked Todd Rodgers to facilitate a panel at our VBSummit later this month to dig into the technical challenges of rolling out AI.
VentureBeat's mission is to be a digital town square for technical decision-makers to gain knowledge about transformative enterprise technology and transact. Discover our Briefings.