Agentic Validation of AI System Requirements: A Hybrid Architecture for Regulatory Compliance

Abstract

AI systems face growing demands driven by both technical complexity and the legal obligations introduced by the EU AI Act. Requirements Engineering (RE) currently lacks systematic tooling to validate project-level quality requirements against these legal obligations in a traceable, auditable, and scalable manner. This paper presents my PhD research, which investigates how automated tooling can support continuous validation of quality requirements against regulatory obligations in AI projects, using the EU AI Act as primary application domain. I describe the research vision, report key findings from a first empirical study that grounds the problem in practice, and introduce CARES (Compliance-Aware Requirements Engineering System), a hybrid architecture that integrates rule-based checks, LLM-based semantic reasoning, and structured human escalation to validate quality requirements using a novel 4CT quality schema. Further, I outline a four-phase evaluation roadmap and discuss contributions to the RE and AI communities.