The Casualty Actuarial Society published a CAS-funded proof-of-concept study in which a two-stage LLM pipeline extracted 36 structured actuarial variables from unstructured claims documents

Event date / reported

Evidence stage

Tasks this bears on

Where this applies

The Casualty Actuarial Society's own peer-reviewed research forum (CAS Forum, Vol. 2026), published 22 June 2026, with funding and support from the CAS Artificial Intelligence Working Group named in the paper itself. Three of the five authors are at universities (Thomas Jefferson University, William and Mary University); the other two run small consultancies (Lieberthal and Associates LLC, MDSight LLC), and a search found no commercial LLM-claims-analysis product for sale by either, so this reads as the professional body's own research rather than a vendor's pitch. The paper itself calls its system a proof-of-concept: a two-stage LLM pipeline that reads unstructured claims documents, medical records, adjuster notes, call transcripts, and extracts 36 structured actuarial variables for reserving, ratemaking and claims management. Every result the paper reports, the 36-variable extraction, the 14 core variables checked against two independent clinician reviewers scoring at least 4 out of 5 accuracy with moderate interrater agreement, and a chain-ladder backtest that lowered reserve error from 6.5% to 4.0%, was produced on synthetic claims data generated from the Synthea FHIR tool plus LLM-written notes and calls, not on any insurer's real claims. The paper's own text says validating the pipeline against an organization's actual claims portfolio is a future step before production use, not something it has done. This establishes that the actuarial profession's own body has a working prototype for the data-assembly half of this task, tested only on synthetic data; it does not establish that any insurer has used it, how much of a real experience study it could replace, or anything about how many actuaries do this work.

What this means

What it does not yet show

This is a proof-of-concept tested only on synthetic data, not on any insurer's real claims. It does not show any insurer using it, say how much of a real experience study it could replace, or say anything about actuary headcount.

What you can check

Does it change the assessment?

Source