Trustworthy Agentic AI Pipelines: Human-in-the-Loop Oversight Architectures for Secure Enterprise Deployment

Abstract

Autonomous agentic AI systems introduce pipeline-specific risks—including prompt injection, privilege escalation, hallucinated reasoning, and audit evasion—that exceed the governance capabilities of current orchestration frameworks. We introduce HITL-AP (Human-in-the-Loop Agentic Pipeline), a seven-layer reference architecture that embeds governance checkpoints across planning, retrieval, and execution stages, enforced by the Agentic Oversight Control Plane (AOCP)—a deployable runtime enforcement component. Our contributions include a pipeline-stage-anchored risk taxonomy and a taxonomy of human-in-the-loop oversight models. To standardize evaluation, we propose three novel metrics: Reasoning Trace Completeness (RTC), Intervention Latency (IL), and Audit Coverage Score (ACS). We evaluate HITL-AP against AutoGen, LangGraph, and CrewAI (n=50 executions per system), demonstrating measurable improvements in auditability (RTC: 98.5% vs. ≤48.6%; ACS: 100% vs. 0%), with RTC differences statistically significant (p < 0.01), and demonstrating alignment with key principles of NIST AI RMF and EU AI Act requirements through architectural mapping. A reference implementation is publicly available at https://github.com/sarmishra/HITL-AP-Project.