Authors - Sourabh Chordiya, Subhrakanta Panda, Akanksha Rathore Abstract - The rapid advancement of Artificial Intelligence (AI) and Large Language Models (LLMs) has unlocked powerful new capabilities for solving complex, multi-step problems. However, this progress has intensified concerns about the environmental sustainability of AI systems. While prior research has examined carbon emissions associated with training and inference in conventional LLM pipelines, emerging paradigms such as Agentic AI, where autonomous agents coordinate to execute multi-stage tasks, and Retrieval-Augmented Generation (RAG) introduce additional layers of computation that remain insufficiently studied from an emissions perspective. In particular, existing carbon measurement frameworks do not adequately capture the dynamic, distributed, and memory-intensive operations characteristic of these systems. This paper analyzes the limitations of current carbon accounting tools and available literature when applied to Agentic AI and RAG-based architectures. The widely used measurement frameworks capture only a fraction of the total computational footprint in such systems, largely omitting emissions arising from memory access patterns, retrieval processes, and inter-agent communication. These overlooked components become increasingly significant as AI workflows shift from single-system inference toward multi-agent orchestration and knowledge retrieval pipelines. Based on this analysis, the paper proposes directions for a comprehensive life-cycle carbon assessment framework and an Eco Rating tailored to next-generation AI systems. Such a framework must account for heterogeneous hardware usage, dynamic inference paths, retrieval infrastructure, and communication overhead across distributed agents. The findings highlight a substantial blind spot in current sustainability evaluations and underscore the urgent need for standardized methodologies that reflect the true environmental impact of emerging AI paradigms.
