Life insurance actuarial data needs scrubbing to apply AI

Takeaways:

• Old, inaccurate policy information and death records blur risks
• Training AI on incorrect information can produce confident but wrong answers
• Paper death records persisted into the 2000s

Just as disorganized or inaccessible data can make it difficult to apply AI to claims, a lack of available and accurate data makes it tough to update actuarial tables used for life insurance.

Life underwriting data typically has been collected between 10 to 20 years ago, explained Amanda Turcotte, senior vice president and chief actuary at Amalgamated Life Insurance Company. To apply AI to accurately determine risk, a life insurer has to go back and examine underwriting factors including medical information (often supplied by Medical Information Bureau) and causes of deaths.

"If you don't know what the cause of death was, or if it's inaccurate, then you have a real challenge," she said. "That is a way bigger problem than the industry realizes."

She pointed to a 2017 study by Vermont's chief medical examiner and health department, with the University of Vermont, which compared six months of then-recent death certificates to medical information about the deceased. The study found that out of 600 death certificates, 305 had errors about the causes of the deaths.

"Insurance carriers depend on these death certificates to code cause of death into their systems," Turcotte said. "If you train an AI agent to say, based on this underwriting evidence, these are the causes of death, what you're training on is information that's half incorrect. An AI system is only as good as the data that comes in. You're going to have a system that hallucinates, provides very confident incorrect answers."

Data for causes of death is imperfect because it was collected for decades from written paper documents. Electronic death registration started in U.S. states in the mid-2000s, with North Carolina being the last to do so in 2021. This improvement is necessary to apply AI tools, including statistical learning tools, for life insurance, according to Turcotte. 

However, even with the Medical Information Bureau providing health information, in Turcotte's view, it will take a decade or more for life insurers to catch up with property and casualty insurers on the completeness and accuracy of data for AI to evaluate.

Data availability is key, according to Adnan Haque, vice president of integrated analytics at Munich Re North America Life. Some life insurers are still using 1970s-era technology to administer policies from that time, he said.

"That leads to data being extremely siloed in some cases," Haque said. "That leads to data being incomplete in some cases, and there's some investment in the baseline underlying data infrastructure that you need to be able to leverage the technology." 

Striving for better prediction of life insurance risk would make underwriting more consistent, according to Katie Kahl, chief product officer at iPipeline, a life insurance software provider.

"More predictive analytics and the ability to incorporate machine learning models and different data sources – that's going to predict the ability to assess risk more accurately, but also more quickly than could be done with more traditional methods," she said. "The other aspect is better prediction around being able to match customers with the right products."