From Conceptual Multi-Agent Claims to Regulated Enterprise Reality: A Privacy-Preserving, Version-Safe, and Auditable Architecture on Microsoft Azure
Recent proposals for AI agent-driven insurance platforms, particularly More and Zhang (2026), show how multi-agent workflows paired with a deterministic rules engine can replace rigid legacy claims systems. However, putting such a design into production for a European property and casualty (P&C) insurer exposes practical engineering gaps around data privacy, historical contract versions, GDPR limits on automated rejections, database load during rule simulations, prompt injection in uploaded documents, and network isolation. This paper extends the conceptual model of More and Zhang into a concrete, private-network reference architecture built on Microsoft Azure, Microsoft Foundry, and Azure Databricks, introducing early PII tokenization with an isolated Cosmos DB vault, contract-inception policy-version binding (AvbVersionId), a one-way GDPR Art. 22 circuit breaker with dual HITL paths, Z3 verification with 500,000-claim Databricks backtesting, and privacy-filtered Langfuse evaluations.
Authors
- Jannis Ebling
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-18
- DOI
- https://doi.org/10.5281/zenodo.24343974
- Primary Topic
- Artificial Intelligence Applications
- Type
- preprint