A Separation-of-Concerns Architecture for Regulated Document Generation: Model-Agnostic Compliance Enforcement in Life Insurance Workflows
The operations of life insurance are among the most document-heavy and heavily regulated financial services. Such conditions have motivated growing interest in generative AI for document automation. The probabilistic nature of large language models brings unacceptable risks into regulated workflows. Hallucinations, omitted mandatory disclosures, and non-compliant phrasing could expose insurers to regulatory scrutiny, legal liability, and consumer harm. The paper presents a compliance-aware reference architecture that reconfigures the design problem by decoupling probabilistic generation and deterministic compliance enforcement. The architecture employs retrieval-augmented grounding over a continuously versioned corpus of authoritative sources, along with automated validators that check for the completeness of disclosures, traceability of evidence, detection of prohibited phrases and a human-in-the-loop review gate which captures audit trails for regulatory scrutiny. Our architecture is evaluated against four widely used LLMs: Claude 3.5 Sonnet, Llama 3.170 Billion, Amazon Nova Pro, and Pixtral Large. For this, we create 50 test cases, which span five document types, yielding 200 baseline and 200 compliance-aware drafts. After excluding 3 failed generations from Pixtral Large, a total of 197 baseline-compliance matched comparisons were used in the final analysis. The compliance-aware pipeline reduces unsupported claims by $\mathbf{8 0} \boldsymbol{\%}$, from $\mathbf{2 6. 5 \%}$ in prompt-only outputs to 5.3%. After human review, all models achieve 100% final approval, with pre-review ready-to-send rates ranging from 78% to 90.7% across models. Results provide statistically significant evidence that the model-agnostic architecture substantially reduces unsupported claims while remaining operationally feasible, providing insurers with a promising proof-of-concept architecture for deploying generative AI in regulated document workflows, motivating further validation at production scale.