Research Summary: Toward Sustainable AI Deployment: A Carbon-Aware Decision Framework for Enterprise Supply Chain Systems

Enterprises frequently deploy large language models (LLMs) for supply chain decisions based on their size rather than empirical performance or environmental impact. Research indicates that larger models do not consistently yield higher decision quality, while their operational carbon footprint is often overlooked. A new Carbon-Aware AI Procurement Framework (CAAPF), based on the Technology-Organization-Environment (TOE) framework, has been developed to guide sustainable AI governance and procurement, aiming to optimize both decision quality and environmental cost.

Why it matters

This analysis highlights a critical misalignment between AI procurement heuristics and strategic objectives related to performance, sustainability, and resource efficiency. Adopting frameworks like CAAPF can enable organizations to make more informed AI deployment decisions, optimizing both operational effectiveness and environmental responsibility. This directly impacts long-term organizational resilience and competitive positioning in an increasingly climate-conscious regulatory and market environment.

Key insights

• Enterprises commonly default to the largest available language models for AI deployment in supply chain decisions, often neglecting empirical performance and environmental costs.
• Benchmarking of six large language models across 520 supply chain tasks revealed a decision quality range of 0.497 to 0.723, indicating significant performance variation.
• Models with the largest disclosed parameter totals did not achieve the highest decision quality scores within the sampled tasks.
• A Carbon-Aware AI Procurement Framework (CAAPF) has been developed as a Green Information Systems (IS) design artifact, operationalizing sustainable AI governance for enterprise procurement.
• The CAAPF draws upon the Technology-Organization-Environment (TOE) framework to integrate considerations of decision quality and estimated generation-related operational carbon.

Source

arXiv — Computers and Society — https://arxiv.org/abs/2609.14881