Volume 69

Human-in-the-loop Agentic AI for energy systems: A decision-support-first framework for safe human oversight Anh Tuan Nguyen, Thi Ngoc Yen Huynh, Theodore Kabangu Nkashama, Yong Han Ahn

Abstract

Agentic AI systems are increasingly proposed for decision support in energy operations because they can retrieve information, invoke tools, coordinate multi-step reasoning and interface with external simulation or optimization tools. However, energy systems are safety-critical and operational technology adjacent environments in which erroneous, ungrounded, or manipulated outputs can create unacceptable operational risk. This study proposes a decision-support-first framework for agentic AI in energy systems, in which agents assist with analysis, planning, and justification while humans retain formal authority for all operational actions. The study first synthesizes core interaction patterns for agentic workflows in energy applications, including retrieval-augmented generation, tool-using reason–act loops, multi-agent coordination, and human approval predicates. It then analyzes major deployment barriers, including stochastic outputs, brittle tool interactions, prompt injection, automation bias, and cyber-physical security exposure. Based on these findings, a phased deployment framework is proposed to guide adoption from offline prototyping to human-approved advisory use. Finally, the study outlines evaluation metrics emphasizing safety, workload, traceability, and auditability. The study argues that agentic AI can create practical value in energy systems only when deployed as an auditable, human-governed decision-support layer rather than as an autonomous controller.

Keywords Human-in-the-loop, Agentic AI, Energy system, Decision support, Operation, OT security, prompt injection, auditability.