AGENTIC AI-DRIVEN REGULATORY COMPLIANCE AUTOMATION: DEVELOPING A RISK-AWARE FRAMEWORK THAT INTEGRATES LARGE LANGUAGE MODELS, AUTONOMOUS AI AGENTS, KNOWLEDGE GRAPHS, POLICY-CONSTRAINED WORKFLOW EXECUTION, AUTOMATED EVIDENCE COLLECTION, AND HUMAN-IN-THE-LOO

Abstract

Modern enterprises face a convergence of complex fast‑evolving regulatory frameworks and traditional rule‑based compliance systems no longer keep up. These deterministic engines are rigid cannot read data and cannot adjust when rules change, which creates big operational risk. I have seen that Large Language Models or LLMs can understand language very well but using LLMs alone in high‑stakes situations is dangerous because LLMs can hallucinate produce outputs that are not guaranteed and do not offer a clear record of how a decision was made. This review paper proposes a risk‑aware framework for Agentic AI‑driven regulatory compliance automation. The framework puts LLMs inside a structure that has deterministic guardrails. The new design pairs specialized, autonomous AI agents that manage workflows with Knowledge Graphs or KGs and Graph‑Augmented Retrieval. These tools give context and help reduce hallucinations. A non‑bypassable Policy‑as‑Code layer or PaC stops tool calls that could cause problems blocks the Deputy Problem and keeps the system closed when something fails. To make decisions auditable the framework records evidence automatically in WORM storage using cryptographic hash chains. It also adds a Human‑in‑the‑Loop or HITL validation step that raises issues based on risk. By separating language generation from the knowledge that represents the rules the architecture keeps monitoring compliance adapts quickly when laws change and produces decisions that are fully explainable and defensible, in court.