Towards an AI Framework for Scalable Compliance Monitoring in Regulated Sectors

Abstract

Ensuring regulatory compliance at a large scale continues to be a significant challenge in sectors that are highly regulated, such as healthcare and data privacy. This study introduces an agentic AI system for automated compliance monitoring that uses large language models (LLMs), rule-based evaluation, and decision routing that takes into account how sure the AI is. We suggest two modular workflows to show how well the framework works at finding responsibilities in legal text and coordinating activities based on how sure the extraction is. The first workflow looks at the Health Insurance Portability and Accountability Act (HIPAA) and employs a set of annotated clauses to check how accurate LLM extraction is by comparing it to ground truth in an organized way. Jaccard similarity and basic information retrieval metrics including precision, recall, F1-score, and per-field accuracy are used to judge performance. The second workflow is based on the General Data Protection Regulation (GDPR) and adds a way to score confidence. Obligations with high confidence scores (>0.9) are automatically recorded for audit, whereas lower-confidence extractions are sent to people to go over. This architecture cuts down on human oversight a lot. More than 88% of situations are handled automatically, and the average time it takes to make an inference per document is less than 15 seconds. Our results show that agentic AI systems may be used in real-time compliance procedures, allowing for decision-making that is scalable, auditable, and in line with human values in all regulatory settings.

Ähnliche Arbeiten