Hybrid orchestration of models for automated insurance correspondence processing in information systems

Abstract

The paper presents a hybrid architecture for automated insurance correspondence processing. The approach combines deterministic extraction, machine learning routing, local large language models, optional external language refinement, and a mandatory verification layer. Deterministic methods identify stable entities such as policy numbers, dates, amounts, addresses, and insurer reference numbers, while machine learning supports primary intent routing. A local large language model is activated only for low-confidence or attachment-heavy cases. An external model is restricted to stylistic rewriting of a minimized case summary and is not allowed to introduce new facts. The final reply is generated from templates and verified before release.