Evaluating AI for Environmental Sustainability – OECD Case Study
Case Study · M&E Practice
Evaluating AI for
Environmental Sustainability
An OECD-Inspired Case Study for Monitoring and Evaluation Professionals
Overview
Artificial Intelligence (AI) is transforming sectors from healthcare and agriculture to humanitarian response. While AI offers significant efficiency gains, it also demands substantial computational resources with measurable environmental consequences.
This case study draws on the OECD report Measuring the Environmental Impacts of AI Compute and Applications: The AI Footprint (2022), which examines how policy makers can better understand and react to both the positive and negative environmental impacts of AI. Rather than focusing solely on technical aspects, this case study explores how Monitoring and Evaluation (M&E) professionals can assess the environmental sustainability of AI-enabled initiatives.
Learning Objectives
- Explain why AI has an environmental footprint
- Distinguish between direct and indirect environmental impacts of AI
- Develop evaluation questions related to AI sustainability
- Identify indicators for assessing AI-powered interventions
- Apply systems thinking when evaluating digital transformation projects
Background
An international development organization plans to deploy an AI-powered platform supporting climate adaptation in agriculture across East Africa.
The platform will:
- Analyze satellite imagery
- Predict drought conditions
- Recommend irrigation schedules
- Generate crop recommendations using machine learning
The organization expects AI to improve agricultural productivity while reducing water consumption and increasing climate resilience.
Before scaling nationally, donors request an independent evaluation of environmental sustainability.
The evaluation team quickly discovers a critical challenge:
“Although AI may help reduce environmental impacts in agriculture, the AI system itself consumes electricity, requires cloud computing infrastructure, stores large datasets, and depends on specialized computing hardware.”
The evaluation must examine both environmental benefits created by the application and environmental costs associated with operating the AI system.
The Evaluation Challenge
Positive Impacts to Assess
- Is AI reducing water consumption?
- Are greenhouse gas emissions from farming decreasing?
- Has fertilizer use declined?
- Are crop yields improving?
Negative Impacts to Assess
- How much electricity does the AI platform consume?
- What cloud infrastructure supports the system?
- What carbon emissions result from model training?
- How often must hardware be replaced?
Applying the OECD Framework
The OECD distinguishes between two categories of environmental impacts:
1. Direct Impacts
These arise from AI computing infrastructure itself and occur throughout the AI lifecycle: production, transportation, operation, and end-of-life disposal.
Examples include:
- Electricity consumption
- Data centre cooling
- Greenhouse gas emissions
- Water usage
- Hardware manufacturing
- Electronic waste
2. Indirect Impacts
These occur because AI changes decisions and human behaviour.
The AI Compute Resources Lifecycle
The OECD framework examines environmental impacts across four stages of the AI compute resources lifecycle:
Production
- Raw material extraction
- Assembly
- Manufacturing
Transport
- Distribution
- Freight transportation
- Handling & storage
Operations
- Energy consumption
- Water consumption
- Carbon footprint
End-of-Life
- Collection & shipping
- Dismantling & recycling
- Waste disposal
Environmental Impact Categories (German Environmental Agency)
The Green Cloud Computing methodology recommends assessing four key impact categories:
- Abiotic depletion potential – use of minerals and fossil fuels
- Cumulative energy demand – use of renewable and non-renewable energy
- Global warming potential – impact on climate change
- Water consumption – particularly important where data centres use evaporative cooling in water-scarce regions
Key Data Points from the OECD Report
Energy Consumption
Global data centres account for approximately 1% of global electricity demand (200–250 TWh in 2020).
AI-Specific Energy Use
Google reports machine learning workloads represented ~15% of total energy use (2019–2021).
GHG Emissions
ICT industry (excluding televisions) accounts for ~1.4% of global GHG emissions.
Water Consumption
Only 33–50% of data centre operators compile and report water-use metrics.
E-Waste
ICT infrastructure accounts for ~12 million tons (~25%) of global electronic waste.
Corporate Renewable Energy
ICT companies account for nearly half of global corporate renewable energy procurement.
Evaluation Questions
Relevance
Does the AI solution address the country’s climate adaptation priorities?
Effectiveness
Has AI improved farming decisions?
Efficiency
Do environmental benefits outweigh the computational resources required?
Sustainability
Can the system continue operating without excessive environmental costs?
Impact
What long-term environmental changes can be attributed to the AI intervention?
Possible Indicators
Environmental Outcomes
- Water consumption per hectare
- Crop yield improvements
- Fertilizer reduction
- Greenhouse gas emissions avoided
AI Operational Indicators
- Electricity consumed during model training
- Electricity consumed during inference
- Cloud computing hours
- Data storage requirements
Sustainability Indicators
- Renewable energy share used by hosting providers
- Hardware replacement cycle
- Electronic waste generated
- Carbon intensity of computing infrastructure
Stakeholders
The evaluation includes interviews with:
Ministry of Agriculture
AI Developers
Cloud Service Providers
Environmental Agencies
Donors
Local Researchers
Each stakeholder provides different evidence regarding environmental performance.
Key Findings (Hypothetical)
- Water consumption decreased by 22%
- Average crop yields increased by 18%
- Farmers reported better drought preparedness
- Cloud infrastructure relied on renewable electricity for 70% of operations
- Training the AI model represented most of the system’s lifetime carbon emissions
- Reusing pre-trained models substantially reduced future computational requirements
Overall finding:
Environmental benefits outweighed operational environmental costs, although continuous monitoring of AI infrastructure remained essential.
Emerging Measurement Standards
Recent developments are making it easier to measure AI’s environmental impact:
UNE 0086 (2025)
A Spanish technical specification establishing a common framework for measuring energy consumption, carbon footprint, water usage, and AI system performance, developed in collaboration with the Secretariat of State for Digitization and Artificial Intelligence.
IEEE P7100
A proposed IEEE standard defining a measurement framework for environmental indicators in AI systems, including methodologies to separate AI-specific compute from general-purpose compute.
AI-CARE
A carbon-aware reporting evaluation tool that standardizes reporting of energy consumption and carbon emissions alongside task performance, introducing the carbon-performance tradeoff curve to visualize the Pareto frontier between accuracy and environmental cost.
Lessons for Monitoring and Evaluation Professionals
Evaluating AI requires looking beyond algorithm accuracy.
Evaluators should assess:
- Environmental sustainability
- Resource consumption
- Equity implications
- Transparency
- Long-term operational costs
- Unintended consequences
Traditional evaluation frameworks remain highly relevant but should incorporate indicators that capture the environmental footprint of AI systems.
A study on societal implications of AI in Earth observation emphasizes the importance of examining not only ecological impacts but also the societal consequences of AI applications, including potential neo-colonial implications of satellite data-driven approaches.
Reflection Questions
- What environmental indicators would you include in an evaluation of an AI-powered health information system?
- How can evaluators balance AI’s environmental costs against its social benefits?
- What additional data would strengthen confidence in the evaluation findings?
- Should organizations report the carbon footprint of AI projects alongside traditional project indicators?
- How can evaluation teams encourage responsible AI adoption without discouraging innovation?
Practical Exercise
Design an evaluation framework for an AI-enabled project in your own sector:
(Education, Health, Agriculture, Humanitarian Response, or Governance)
Identify:
- Five evaluation questions
- Five environmental indicators
- Five data sources
- Three potential risks
- Three recommendations for improving AI sustainability
Compare your framework with the OECD approach and discuss how it could support evidence-based decision-making.
Key Takeaways
Further Reading
- OECD (2022), Measuring the Environmental Impacts of AI Compute and Applications: The AI Footprint
- Rolnick et al. (2022), Tackling Climate Change with Machine Learning
- Kaack et al. (2022), Aligning Artificial Intelligence with Climate Change Mitigation
- UNE 0086 (2025), Specification for Measuring AI Sustainability
- Rehak et al. (2025), Convivial AI? Developing a Societal Impact Analysis Grid for Assessing Artificial Intelligence in Earth Observation
This case study was developed for EvalCommunity Academy to support M&E professionals in evaluating AI-enabled interventions for environmental sustainability.
Case Study · M&E Practice · AI & Sustainability
“Evidence-based evaluation for a sustainable digital future.”
