Rethinking Life, Annuity and Benefits PAS Systems Part 2
Part two was originally about systems of record, but the first part raised important points about AI’s impact on future systems development and deployment. So, I decided to discuss my position on the best use of AI in policy administration systems and calculation engines.
AI and Calculation Engines
For an LA&B calculation engine designed to maintain product information, rules, and rates, AI brings limited value today. Core system calculation engines are highly deterministic, requiring precision, repeatability and stability to help ensure that policies are issued, billed, and claims are adjudicated correctly.
They need structured data, predefined rules, and can't be fluid from instance to instance without significant business issues. The calc engine becomes a point of stability, and from that solid foundation AI-based solutions including Agentic AI get some of that good data they need to do their thing.
AI and Product Dashboards
AI can be very useful in the insurance product development process if supported by a strong and open calculation engine and system of record to support what-if analysis, profitability testing and market analysis. Gen AI can be useful in setting profiles through natural languages, explaining a particular target market in detail as a default which can then be modified in the ideation process.
An example could be a target segment of Gen Zs with a pet, creating a startup and dealing with sandwich generation issues. Combining this baseline profile when instructing the AI with a large library of structured and unstructured insurance, lifestyle, and health information, your customer experience, and your calc engine, your Product Development Workbench would be incredibly flexible and efficient
My View on AI Today
I want to clarify my position on AI for insurance today. I've watched it grow from early neural nets and rules-based inference engines in the 80s and 90s to LLMs. Andrej Karpathy calls LLMs "people spirits: stochastic simulations of people" that are "brilliant interns with perfect recall but no judgment" with "jagged intelligence" and "anterograde amnesia" in his Software 3.0 model, requiring human supervision for some time to come.
I like Karpathy’s Software 3.0 model, and when considering AI for insurance systems it leads me to a few important points:
Understand which functions in your insurance ecosystems are probabilistic (grey area) and which are deterministic (black & white) and/or require detailed auditability and traceability.
Embrace a human-AI collaboration mindset over “let the Agentic AI figure it out” and assume it will evolve through machine learning alone.
To be fair to AI, when we watch the evolution of Agentic AI in insurance, we need to remember not to set a higher bar for Agentic AI than for well-trained and experienced human agents.
Insurance industry regulations can be arcane, inconsistent, and hard to reconcile even for an AI. The legal and financial cost of your AI breaking rules is too high to not directly address as part of your AI development process, especially for Agentic AI.
AI should be thoughtfully and selectively applied in insurance systems today, and we need to update that thinking as AI becomes more capable in the next few years.
Next week, systems of record.
