Reserve Acceleration with Agentic AI
his author saw a lot of sessions, pitches, and deep dives on reserving, technical acumen, AI approaches, and workflow efficiencies in Las Vegas at the 2026 Casualty Loss Reserve Seminar.
They might all be summed up in a best-of-breed session titled “Reserve Acceleration with Agentic AI” by a KPMG team and an active client at Root Insurance.
Jeremy Smith, FCAS, told the audience that many steps of progress and pilots are ongoing at Root Insurance with KPMG, with several illustrations of advancing machine learning models into agentic AI-accelerated workflows across the Root claims organization.
Some examples were text analytic-supported feature extraction and predictive modeling of high-value claims problems like litigation, reinsurance recovery, attorney involvement, suspicion of fraud, and payment integrity.
These agentic workflow capabilities have also been turned into key supporting activities in claims, including loss reserving.
He introduced the technical environment as Copilot along with Databricks and then foreshadowed that much of the current best progress in generating value for clients is where teams are using agentic AI that is enabling machine learning (ML), that is, ML masquerading as AI; they now see a ten-times faster time to accelerate data science in claims into reserves, yet caution that there is still an actuary-in-the-loop requirement.
While general-purpose AI capabilities for extracting data, making structured from unstructured data, and ingesting loss run data are largely table stakes for many steps now, the ability to support a full spectrum of end-to-end reserving is still more of an ambition than ready for production in this engagement. The current progress in the engagement for reserving was reported as mostly expanding the amount of data, toolkits, and processes that can be combined into the tight timeframes surrounding making a reserving call, but AI is not ready to be a reserving actuary here yet.
The speakers that followed filled in all the pieces to make this one of the best sessions at CLRS, showcasing that going a lot faster with agentic AI is feasible, yet the need to be making correct decisions and using actuarial judgment is still ultimately very essential and still a step that requires people with actuarial training.
Micah Ruth from Root Insurance discussed the progression of AI capabilities and how changes in technology are creating innovation opportunities that are being applied in claims processes.
Ruth shared that many steps in customer interaction can be supported with classical AI –and it fits in its own lane supported by scripts. Generative AI needs more authority on what is returned to a user, so it’s better for it to support functions like summary statements and abstracting bullet points from pages of text and mounds of unstructured data.
Ruth pointed to use cases where agentic AI can be more autonomous and deal with complexity itself. She concluded with an example of payment integrity where a multi-agent system has been created targeting identity fraud challenges and two AI agents communicate to understand the risk of fraud before issuing payments. The speed and scale of the multi-agent task exceeds anything a person could do.
Nate Loughin, ACAS, of KPMG discussed how large language models and AI agents can accelerate knowledge-based tasks, particularly by retaining and summarizing lengthy conversations and complex information. However, he cautioned that AI can still generate inaccurate or fabricated information, especially when tackling the technical complexities and professional judgment required in reserving.
Smith observed that there was a bit of humor in the fact that in the real world, the senior and junior staff find the technology intriguing and satisfying, while those actually doing the reserving work in the middle are getting stuck dealing with identifying and resolving any AI-generated hallucinations. So, for reserving itself, they are still skeptical of using AI for decision-making, but they definitely see an acceleration of the work process that gives them more time to think the hard thoughts on making accurate and defensible reserve judgments. It was also noted that AI makes it much easier to try to incorporate more data and even outside tools and trends.
Ruth emphasized that proper governance leads to trust, and she noted that each company needs to bring its own values into incorporating AI into its processes. She noted that at her company there is a culture of being “values led,” which means fairness first over bias potential of an AI that may try to be cheaper or faster. At any juncture, her challenge for AI is to make being trustworthy a priority, or WWAHD — what would a human do?
Sophia Song, FCAS, summarized the state of the industry for AI. She noted that personal lines are ahead of commercial lines, but there is a larger opportunity in commercial for the strongest parts of current AI capabilities with the hundreds of pages of facts and forms typical in commercial insurance. She laid out the case for humans getting better faster using visualization powered by AI, as well as change-detection alerting for everything from filings to data trends to emerging claims risks.
To illustrate how agentic AI is reshaping reserving, Song introduced the “Accelerate, Elevate, and Scale” framework, demonstrating how AI can complete initial analyses in a fraction of the time required by humans, freeing actuaries to focus on professional judgment, deeper insights, and creating value for the business.
Loughin did a demo of the various applications in predictive analytic workflows before showing the current reserving uses. A big feature was how the AI process generated code that can be applied on a go forward basis and shared as knowledge across the organization. Additionally, the demonstration showed that AI can accelerate most efforts getting data in shape and generating analyses. Nonetheless, the study concluded that it favors humans over robots when it comes to picking a peg loss ratio.
Loughin showed a multi-agent opportunity by breaking agentic responsibility apart — orchestrated agent work can be task-specific, like in reinsurance recovery, to pull and read bordereaux documents, to read claims, review demand packages, and plan strategies for claims outcomes.
A brief Q&A session hit on the importance of governance, the need for better data, the concerns over reproducibility issues of generative technology, and concluded with a large round of applause.