AI experimentation has accelerated across the life insurance industry, but moving from pilots to production-ready underwriting workflows remains a greater challenge. In this Q&A, Patrick Sullivan discusses where carriers are getting stuck, what tends to break at scale, and how leading insurers are approaching AI adoption in a way that is practical, trusted, and aligned with the realities of life underwriting.

Patrick Sullivan: With the rise of agentic software and more capable large language models, it has become much easier to build and test AI applications. As a result, we’re seeing many pilots and proofs of concept in the market today, and these models are becoming quite effective at understanding life underwriting.

There are many ideas and use cases that people want to test.  The challenge is that while it has become easier to run pilots, it is just as hard, if not harder put things into production. POC code typically has to be rewritten for maintainability. Then carriers are dealing with increased cyber risk, stricter compliance requirements, and generally more constraints around what can be deployed. 

The other challenge is around collaboration. Success depends on strong partnerships between the teams building AI and the underwriting and medical teams using it. Things often break down at the workflow level. An underwriting team may tolerate extra steps during a pilot, but in production, the focus is on getting cases through quickly and efficiently. AI has to be built into existing workflows rather than added as another step in the process.

Data can also become an issue. Development efforts are often based on relatively small samples of underwriting data. When solutions move into production, we can discover categories of cases or data issues that were not represented during development, requiring the solution to be adjusted or rebuilt.

Patrick Sullivan: One challenge is the underlying technology foundation. If a company does not have a modern technology stack and modern software development practices, scaling AI can be difficult. Things like cloud expertise, automated testing, continuous integration and deployment all become essential when you are trying to move beyond a handful of isolated use cases.

The other area is controls. Carriers need guardrails around how AI is used in production. That includes handling a series of issues: hallucinations, latency, service availability, and cost. If those controls are not in place, it is more likely a solution loses the organization's confidence.

It’s as much a leadership challenge as a technology challenge. If senior leadership sees AI as a priority and believes it can help the business move faster, improve accuracy, or offer something new to the market, then many of the tactical challenges can be addressed through the right investments, staffing models, and cross-functional teams.

Patrick Sullivan: The organizations making the most progress recognize that collaboration is essential. They bring together engineering, underwriting, business analysis, and data science, and they create environments where those disciplines can work closely together. It does not necessarily have to be one team, but it does require people to work together consistently.  Adaptability is the key attribute enabling  success.

Life insurance underwriting is different from many of the problems AI has been applied to elsewhere. You're evaluating a person's health, and there is potentially unlimited information to consider, interpret, or verify. That's why underwriting continues to require both technology and human expertise.

The carriers making progress are also realistic about where automation belongs. Some parts of underwriting can be fully automated and probably should be. Other parts never will be. Some situations will always need to be a human in the loop, particularly when a case requires judgment, context or another piece of information to consider. These carriers are also designing AI to fit naturally into existing underwriting workflows, rather than asking underwriters to take extra steps or work around the technology. The key is recognizing the difference and designing workflows that support both automation and human expertise.