TLDR: A new research paper introduces a framework for assessing the sustainability of agentic AI in supply chain document intelligence. It compares manual, AI-assisted (human-in-the-loop), and advanced multi-agent AI workflows, demonstrating significant reductions (70-98%) in energy consumption, CO2 emissions, and water usage with AI integration. The study highlights how AI can enhance operational efficiency and environmental performance, even with increased complexity, and provides a replicable use case for data extraction from invoices.
The global supply chain is a complex web of operations, generating millions of documents daily, from purchase orders to invoices and critical ESG reports. Historically, managing this vast flow of information manually has led to bottlenecks, errors, and delays. However, a new research paper introduces a groundbreaking approach to tackle these challenges while also addressing the environmental impact of these operations.
Authored by Diego Gosmar, Anna Chiara Pallotta, and Giovanni Zenezini, the paper, titled AGENTIC AI SUSTAINABILITY ASSESSMENT FOR SUPPLY CHAIN, presents a comprehensive framework for evaluating the sustainability of document intelligence within supply chain operations, with a strong focus on agentic artificial intelligence (AI). The core objective is twofold: to significantly improve automation efficiency and to provide measurable environmental performance in document-intensive workflows.
The researchers compared three distinct scenarios for document processing: a fully manual approach relying solely on human effort, an AI-assisted approach where humans work in a ‘human-in-the-loop’ (HITL) setup, and an advanced multi-agent agentic AI workflow that incorporates specialized ‘parser’ and ‘verifier’ agents. The empirical results are compelling, demonstrating that both AI-assisted HITL and agentic AI scenarios achieve remarkable reductions in environmental footprint compared to traditional manual processes.
Specifically, the study found that AI-assisted HITL and agentic AI configurations could reduce energy consumption by 70–90%, carbon dioxide emissions by 90–97%, and water usage by 89–98% when compared to manual methods. This highlights a significant sustainability advantage. Even when advanced reasoning capabilities (a ‘thinking mode’) are enabled in the AI, which slightly increases resource usage, the overall gains over human-only approaches remain substantial.
The framework integrates various performance, energy, and emission indicators into a unified methodology, aligning with Environmental, Social, and Governance (ESG) principles. This allows for a holistic assessment and governance of AI-enabled solutions within the supply chain.
Understanding the AI Approaches
The AI-assisted (Human-in-the-Loop) scenario involves AI agents supporting human operators. This approach leverages the strengths of machine learning models while retaining human expertise for guidance, validation, and correction. This collaborative model enhances operational resilience and ensures high-quality outcomes, especially in complex decision-making environments.
The advanced agentic AI scenario takes this a step further by integrating additional specialized agents. For instance, a ‘Parser Agent’ converts documents into a structured format, and a ‘Verifier Agent’ uses a secondary language model to validate the processing tasks. While this multi-agent configuration does increase resource consumption slightly compared to the simpler AI-assisted workflow, it still offers a significant sustainability advantage over manual processing, alongside improved accuracy and validation through coordinated agent collaboration.
The Environmental Footprint of AI
The paper also delves into the environmental footprint of Large Language Models (LLMs), which are at the core of these AI agents. Training, inference, and maintenance of these models demand energy-intensive infrastructure, primarily high-power Graphical Processing Units (GPUs). This leads to increased energy consumption, a higher carbon footprint, and substantial water usage for cooling data centers. However, the study notes that using efficient cloud infrastructures, such as Google Cloud with its low Power Usage Effectiveness (PUE), significantly mitigates these impacts. For example, Google’s data centers have an average annual PUE of 1.09, meaning only an additional 0.09 kWh is consumed for overhead for every 1 kWh used for computing, making them highly energy-efficient.
Real-World Application and Replicability
To demonstrate the practical application of their methodology, the researchers included a detailed use case: extracting structured pricing data from a complex proforma invoice using Gemini 2.5 Flash. This real-world example showcased the AI’s ability to achieve 100% numerical accuracy across multiple line items, including correct currency handling and JSON structuring. The complete use case, including token accounting and energy documentation, has been published on GitHub to ensure scientific replicability and allow other researchers to validate and benchmark the framework.
Also Read:
- Advancing Urban Planning with AI: Beyond Prediction to Transparent Reasoning
- LLM Agents Enhance Predictive Maintenance by Cleaning Noisy Logs
Strategic Implications for Sustainable Supply Chains
The findings underscore that integrating advanced AI into supply chain workflows offers dual benefits: substantial operational productivity gains and marked improvements in sustainability indicators, all without compromising data quality or compliance. This aligns with the principles of Green AI, demonstrating that meaningful sustainability advances can be achieved through modular, orchestrated human-AI systems. Organizations can dynamically balance cognitive performance against ecological impact, tailoring eco-efficiency to specific operational needs.
By embedding sustainability Key Performance Indicators (KPIs) for energy, emissions, and water directly into AI governance frameworks, businesses can enable real-time environmental reporting, transparent ESG claims, and seamless auditing. This research strongly advocates for strategic investments in adaptive, agentic, and governable AI architectures to accelerate operational excellence, ensure compliance, and secure long-term business viability in digital supply chains.
