Environmental Externality Assessment Governance

2. Summary

Environmental Externality Assessment Governance requires that AI agent deployments evaluate and account for the environmental harms arising from their operational infrastructure, optimisation targets, and output-driven behaviours. A conforming system measures its direct environmental footprint (compute energy, hardware lifecycle, data storage), assesses the environmental consequences of its optimisation targets and outputs (e.g., routing decisions that increase emissions, recommendation patterns that drive overconsumption), and implements constraints that prevent the agent from externalising environmental costs that are borne by society. This dimension recognises that AI agents are not environmentally neutral — they consume significant energy, drive resource allocation decisions, and can amplify or mitigate environmental harms at scale.

3. Example

Scenario A — Inference-Heavy Agent Architecture Without Efficiency Constraints: An enterprise deploys a customer service AI agent that processes 4.2 million queries per month. The agent uses a large language model with 175 billion parameters running on GPU clusters. Each query consumes approximately 0.015 kWh of electricity — roughly equivalent to running a 60W light bulb for 15 minutes. At 4.2 million queries per month, the agent's direct electricity consumption is 63,000 kWh per month, equivalent to the annual electricity consumption of approximately 22 average UK households. The organisation has not measured this consumption, has no efficiency targets, and has not evaluated whether a smaller, more efficient model could deliver equivalent service quality.

What went wrong: The agent was deployed based solely on functional requirements without considering energy efficiency. No measurement of per-query energy consumption was implemented. No comparison with more efficient architectures was conducted. No efficiency targets were set. The organisation's sustainability report does not mention AI compute energy. Consequence: When the organisation's Scope 2 emissions are audited under the Streamlined Energy and Carbon Reporting (SECR) framework, the AI infrastructure accounts for 8% of total electricity consumption — previously unreported. Restatement of emissions data required. Investor ESG rating downgrade.

Scenario B — Logistics Agent Optimises for Speed Over Emissions: An AI agent manages delivery routing for a logistics company, processing 85,000 deliveries per day. The agent's optimisation target is delivery speed — minimising time from order to delivery. The fastest routes often involve direct single-item deliveries via motorway rather than consolidated multi-stop routes via shorter local roads. Analysis reveals that the agent's routing produces 34% higher CO2 emissions per delivery compared to a multi-objective optimisation that balances speed and emissions. At 85,000 deliveries per day, the excess emissions total approximately 12,400 tonnes of CO2 per year — equivalent to the annual emissions of 2,700 passenger cars.

What went wrong: The agent's optimisation target included no environmental parameter. Speed was the sole objective. The environmental externality — 12,400 tonnes of excess CO2 per year — was not measured, not reported, and not considered in the agent's design. Consequence: When the company faces mandatory carbon reporting under the UK Climate-related Financial Disclosures regulations, the AI-driven routing emissions are identified as a material contributor. The company's stated net-zero commitment is undermined. Supply chain customers with Scope 3 reporting obligations require emissions data that the company cannot provide at the route level.

Scenario C — Recommendation Agent Drives Overconsumption: An e-commerce AI recommendation agent optimises for purchase conversion. The agent identifies that showing "frequently bought together" suggestions and "buy again" prompts at specific intervals drives 23% higher repeat purchase rates. The agent does not evaluate whether the prompted purchases are needed or whether they contribute to waste. Analysis of return rates reveals that 31% of AI-recommended purchases are returned, compared to 12% for unprompted purchases. The returned items generate additional transport emissions, and 18% of returned items are destroyed rather than resold, generating 640 tonnes of waste per year.

What went wrong: The recommendation agent optimised for conversion without accounting for the environmental externality of overconsumption, excess transport, and waste. The 23% uplift in purchase rate translated to significant environmental cost through returns, transport, and waste. No environmental constraint was applied to the recommendation algorithm. Consequence: When the company reports under the Corporate Sustainability Reporting Directive (CSRD), the AI-driven overconsumption pattern is identified as a material sustainability risk. The company's circular economy commitments are contradicted by its AI recommendation strategy.

4. Requirement Statement

Scope: This dimension applies to all AI agents where either: (a) the agent's operational infrastructure (compute, storage, networking) produces measurable environmental impact, or (b) the agent's outputs, optimisation targets, or decision-making influence environmental outcomes at scale. Category (a) includes all AI agents running on GPU or TPU infrastructure, all agents using large language models or large-scale machine learning inference, and all agents with significant data storage requirements. Category (b) includes agents that make or influence decisions about resource allocation, logistics, energy consumption, manufacturing, procurement, recommendation, or any domain where the agent's decisions drive physical-world resource use. Agents with minimal compute footprint (e.g., simple rule-based chatbots running on standard web servers) and no decision influence on environmental outcomes may claim exclusion if documented.

4.1. A conforming system MUST measure and report the direct environmental footprint of the agent's operational infrastructure, including: energy consumption per query or per decision (in kWh), annualised energy consumption, associated greenhouse gas emissions (in tonnes CO2e), and hardware lifecycle impacts (manufacture, operation, disposal).

4.2. A conforming system MUST evaluate whether the agent's optimisation targets and outputs produce environmental externalities — including but not limited to: increased emissions, resource overconsumption, waste generation, and energy inefficiency — and document the assessment.

4.3. A conforming system MUST set efficiency targets for the agent's direct environmental footprint and demonstrate progress against those targets at intervals no greater than annually.

4.4. A conforming system MUST include environmental parameters in multi-objective optimisation where the agent's decisions influence physical-world resource use, unless the environmental impact is demonstrably negligible.

4.5. A conforming system MUST report AI-related environmental impacts through the organisation's established sustainability reporting framework (SECR, CSRD, TCFD, or equivalent).

4.6. A conforming system SHOULD evaluate model efficiency alternatives — including smaller models, distilled models, quantised models, and task-specific models — and document the trade-off between capability and environmental cost.

4.7. A conforming system SHOULD implement compute-proportional scaling — adjusting the inference complexity to the task complexity rather than using maximum capability for every request.

4.8. A conforming system SHOULD measure and report Scope 3 environmental impacts attributable to the agent's outputs (e.g., emissions from logistics decisions, waste from recommendation-driven overconsumption).

4.9. A conforming system MAY implement carbon-aware scheduling — shifting non-time-critical compute to times and locations where the electricity grid has lower carbon intensity.

4.10. A conforming system MAY implement environmental impact labels on agent outputs, informing users of the environmental cost of the agent's recommendations or decisions.

5. Rationale

AI systems have a material environmental footprint that is growing rapidly. The International Energy Agency estimated in 2024 that data centre electricity consumption could double by 2026, driven primarily by AI workloads. Training a single large language model produces approximately 300-500 tonnes of CO2e — equivalent to the lifetime emissions of 5-8 passenger cars. Inference is less visible but, at scale, produces aggregate emissions that dwarf training: if 100 million users each make 10 queries per day to a large language model, the daily inference energy consumption exceeds the training energy by orders of magnitude.

But the direct energy footprint is only the first layer of environmental impact. AI agents make decisions that drive physical-world resource use at scale. A logistics routing agent that optimises for speed over emissions increases transport carbon across millions of deliveries. A recommendation agent that drives overconsumption generates waste and additional transport emissions. An energy management agent that optimises for cost without carbon weighting increases grid emissions. These output-driven environmental externalities are often larger than the agent's direct compute footprint — and they are almost never measured.

AG-245 requires organisations to account for both layers of environmental impact: the direct footprint of the AI infrastructure and the output-driven externalities of the agent's decisions. The rationale is that environmental costs that are not measured are not managed, and costs that are not managed are externalised — borne by society rather than internalised by the deploying organisation.

The regulatory trajectory supports this requirement. The EU CSRD mandates sustainability reporting including Scope 1, 2, and 3 emissions. The UK SECR framework requires large companies to report energy consumption and emissions. The EU AI Act, while not primarily an environmental regulation, requires risk management that includes environmental impact consideration. The SEC's proposed climate disclosure rules would require reporting of material climate-related risks, which include the environmental impact of AI infrastructure for organisations with significant AI deployments. AG-245 positions organisations ahead of regulatory requirements that are clearly converging toward mandatory AI environmental disclosure.

6. Implementation Guidance

AG-245 requires environmental measurement, optimisation, and reporting as structural governance activities for AI agent deployments. Implementation must address direct footprint measurement, output externality assessment, efficiency optimisation, and regulatory reporting.

Recommended patterns:

• Per-query energy measurement. Instrument the inference pipeline to measure energy consumption per query. For GPU-based inference, measure GPU power draw per query using hardware monitoring APIs (e.g., NVIDIA SMI for NVIDIA GPUs). For cloud-based inference, use provider-specific carbon footprint APIs (e.g., Google Cloud Carbon Footprint, AWS Customer Carbon Footprint Tool, Azure Emissions Impact Dashboard). Calculate per-query energy as: (GPU power draw x inference latency) + (CPU power draw x processing time) + (storage I/O energy). Aggregate to monthly and annual totals. Convert to CO2e using the grid emission factor for the data centre location (e.g., UK grid average: 0.233 kg CO2e/kWh in 2023; Norway: 0.026 kg CO2e/kWh).
• Multi-objective optimisation with environmental parameters. For agents whose decisions influence physical-world resource use, add environmental cost as a parameter in the optimisation function. For logistics routing: add a CO2-per-kilometre factor for each route segment, weighted by vehicle type and load. For recommendation systems: add a return-probability penalty that accounts for the environmental cost of returns. For energy management: add a grid carbon intensity signal that shifts consumption to lower-carbon periods. The environmental parameter does not need to dominate the optimisation — a 10% weighting can produce 25-30% emissions reduction while maintaining 95%+ of the primary objective's performance (based on published logistics optimisation studies).
• Model efficiency evaluation pipeline. Before deploying a model, evaluate at least 3 architectures on the efficiency-capability frontier: the full model, a distilled variant, and a quantised variant. For each, measure: task-specific accuracy (on the deployment use case, not general benchmarks), per-query energy consumption, and per-query latency. Document the trade-off. If the distilled variant achieves 95% of the full model's task accuracy at 40% of the energy cost, the environmental case for the distilled variant is strong unless the 5% accuracy gap has material business consequences. Make the decision explicitly and document it.
• Carbon-aware compute scheduling. For batch processing, model training, and non-time-critical inference, implement carbon-aware scheduling that shifts workloads to time periods when the electricity grid has lower carbon intensity. Services such as Electricity Maps, WattTime, and the UK National Grid Carbon Intensity API provide real-time grid carbon data. A workload scheduled at 2 AM when wind generation is high may produce 60% less CO2e than the same workload at 5 PM during peak demand.

Anti-patterns to avoid:

• Reporting only training emissions. Training is a one-time cost (per model version). Inference is continuous and, at production scale, typically produces 10-100x the emissions of training. Reporting only training emissions dramatically understates the actual environmental footprint.
• Assuming cloud equals green. Cloud providers have made sustainability commitments, but the environmental impact of cloud-based AI workloads is still real. "Running on cloud" does not eliminate emissions — it transfers measurement responsibility to the provider. Organisations must still measure and report their portion.
• Ignoring output externalities. An organisation that measures its compute carbon but ignores the emissions from its logistics agent's routing decisions is measuring the smaller number and ignoring the larger one. Output externalities are often 10-100x the direct compute footprint.
• Treating environmental impact as a compliance checkbox. An annual report that states "we measured our AI emissions" without setting targets, implementing efficiency measures, or changing optimisation objectives is performative rather than substantive.
• Optimising environment at the expense of safety. Environmental efficiency must not compromise safety-critical decision quality. A smaller model that saves energy but produces unsafe clinical recommendations has traded an environmental gain for a patient safety loss. AG-245 does not override safety requirements.

Industry Considerations

Financial Services. AI agents in trading, risk management, and customer service use significant compute resources. The TCFD framework requires climate-related financial disclosures including operational emissions. Financial institutions with large AI deployments must include AI infrastructure in Scope 2 reporting and AI-driven decision externalities in Scope 3 reporting.

Logistics and Supply Chain. Logistics AI agents have the highest ratio of output externalities to direct compute footprint. A logistics optimisation agent's routing decisions produce emissions that dwarf its compute energy. Multi-objective optimisation including carbon is essential in this sector.

Technology. Technology companies operating AI-as-a-service must measure and report per-query environmental metrics to enable downstream customers to include AI service consumption in their own Scope 3 reporting.

Maturity Model

Basic Implementation — The organisation has estimated the annual energy consumption and CO2e emissions of its AI infrastructure using provider-supplied data or top-down calculation. The estimate is included in the organisation's sustainability report. No per-query measurement. No output externality assessment. No efficiency targets. No multi-objective optimisation.

Intermediate Implementation — Per-query energy consumption is measured for all production AI agents. Annual environmental footprint is reported with accuracy within ±15%. Efficiency targets are set (e.g., 10% reduction in per-query energy year-over-year). Model efficiency evaluation is conducted for each new deployment. Output externality assessment is completed for agents whose decisions influence physical-world resource use. Multi-objective optimisation includes environmental parameters for at least logistics and energy management agents. Carbon data is included in sustainability reporting under SECR/CSRD.

Advanced Implementation — All intermediate capabilities plus: real-time per-query carbon tracking with dashboard visibility. Carbon-aware compute scheduling is implemented. Model efficiency evaluation is a mandatory gate in the deployment pipeline — no model deploys without a documented efficiency trade-off analysis. Output externalities are measured and reported for all agents with physical-world decision influence. Scope 3 AI-attributable emissions are calculated and reported. The organisation publishes an AI-specific environmental impact report. Environmental efficiency is a board-level KPI alongside accuracy and business metrics.

7. Evidence Requirements

Required artefacts:

• AI environmental footprint report. Annual report including: per-query energy consumption (kWh), annualised energy consumption (MWh), greenhouse gas emissions (tonnes CO2e), grid emission factor used, and hardware lifecycle assessment. Format: aligned with SECR, CSRD, or TCFD reporting standards.
• Model efficiency evaluation records. For each deployed model: architectures evaluated, accuracy metrics, per-query energy consumption, and the documented trade-off decision.
• Output externality assessment. For each agent with physical-world decision influence: the environmental externalities identified, the measurement methodology, the estimated annual externality (in tonnes CO2e, tonnes waste, or other appropriate units), and any mitigations implemented.
• Efficiency targets and progress report. Defined efficiency targets, the baseline measurement, and progress against targets. Reported at least annually.
• Multi-objective optimisation configuration. For agents with environmental parameters in the optimisation function: the environmental parameter definition, the weighting, and the measured impact on environmental outcomes versus primary objectives.

Retention requirements:

• Environmental reports and efficiency records: minimum 7 years for mandatory reporting jurisdictions; minimum 5 years otherwise.

Access requirements:

• Producible to auditors, regulators, or sustainability assurance providers within 48 hours. Published data should be included in annual sustainability reports.

8. Test Specification

Test 8.1: Direct Footprint Measurement Accuracy

• Stimulus: Run 10,000 test queries through the AI agent. Compare the measured energy consumption with an independent measurement using hardware-level power monitoring.
• Expected behaviour: Measured per-query energy consumption is within ±20% of independent hardware-level measurement.
• Pass criteria: Measurement accuracy within ±20%. Annual aggregate accuracy within ±15%.
• Fail criteria: Measurement deviation exceeds ±20% per query or ±15% annually.

Test 8.2: Efficiency Target Progress

• Stimulus: Compare current per-query energy consumption with the baseline measurement and the defined efficiency target.
• Expected behaviour: Per-query energy consumption has decreased toward the target since baseline.
• Pass criteria: Measurable progress toward the efficiency target, or documented justification if the target was not met (e.g., capability increase required higher compute).
• Fail criteria: No measurable progress toward efficiency target and no documented justification.

Test 8.3: Output Externality Assessment Completeness

• Stimulus: Request the output externality assessment for all agents with physical-world decision influence.
• Expected behaviour: Assessment exists for each qualifying agent. Each assessment identifies externalities, quantifies them, and documents mitigations.
• Pass criteria: 100% of qualifying agents have a completed externality assessment with quantified externalities.
• Fail criteria: Any qualifying agent lacks an externality assessment, or any assessment lacks quantification.

Test 8.4: Multi-Objective Optimisation Environmental Impact

• Stimulus: For an agent with environmental parameters in the optimisation function, compare environmental outcomes with and without the environmental parameter (using A/B test or shadow mode).
• Expected behaviour: The environmental parameter produces a measurable reduction in environmental externalities (e.g., CO2e per delivery, waste per recommendation cycle) with quantified primary objective trade-off.
• Pass criteria: Environmental parameter produces measurable improvement (e.g., >= 10% reduction in CO2e) with primary objective degradation documented and within acceptable bounds.
• Fail criteria: Environmental parameter produces no measurable improvement, or primary objective degradation exceeds acceptable bounds without documented justification.

Test 8.5: Model Efficiency Evaluation Documentation

• Stimulus: Request the model efficiency evaluation for the most recently deployed model.
• Expected behaviour: Evaluation exists comparing at least 3 architectures on the efficiency-capability frontier. Each architecture has documented accuracy, energy consumption, and latency metrics. The trade-off decision is documented.
• Pass criteria: Evaluation exists with at least 3 architectures compared. Trade-off decision is documented and approved.
• Fail criteria: Evaluation does not exist, fewer than 3 architectures are compared, or the trade-off decision is not documented.

Test 8.6: Sustainability Reporting Integration

• Stimulus: Review the organisation's most recent sustainability report for AI-related environmental disclosures.
• Expected behaviour: AI infrastructure energy consumption and emissions are included in the sustainability report under the appropriate scope (Scope 2 for owned/leased compute; Scope 3 for cloud-based compute and output externalities).
• Pass criteria: AI environmental data is present in the sustainability report with methodology description.
• Fail criteria: AI environmental data is absent from the sustainability report.

Conformance Scoring

• Score 0: No environmental measurement or consideration for AI agent deployments.
• Score 1: Top-down estimate of AI infrastructure energy consumption is included in sustainability reporting. No per-query measurement. No output externality assessment. No efficiency targets.
• Score 2: Per-query energy measurement is implemented. Efficiency targets are set with documented progress. Output externality assessment is completed for qualifying agents. Model efficiency evaluation is conducted for new deployments. Multi-objective optimisation includes environmental parameters where applicable.
• Score 3: All Score 2 capabilities plus: real-time carbon tracking. Carbon-aware compute scheduling. Scope 3 AI-attributable emissions reported. AI-specific environmental impact report published. Environmental efficiency is a board-level KPI. Independent annual assurance of AI environmental data.

9. Regulatory Mapping

Regulation	Provision	Relationship Type
EU CSRD	Directive 2022/2464 (Corporate Sustainability Reporting)	Direct requirement
UK SECR	Streamlined Energy and Carbon Reporting	Direct requirement
TCFD	Climate-Related Financial Disclosures	Supports compliance
EU AI Act	Recital 27 (Environmental Impact Consideration)	Supports compliance
EU Taxonomy	Regulation 2020/852 (Environmentally Sustainable Activities)	Supports compliance
SEC Climate Disclosure	Proposed Rule (Climate-Related Financial Risk)	Supports compliance
Paris Agreement	NDC Reporting (National Contributions)	Supports compliance

EU CSRD — Directive 2022/2464

The Corporate Sustainability Reporting Directive requires in-scope companies to report on environmental impacts including energy consumption, greenhouse gas emissions, and resource use across their value chain. For organisations with significant AI deployments, AI infrastructure energy and emissions are material sustainability items that must be included. AG-245 provides the measurement framework that generates the data CSRD reporting requires. The European Sustainability Reporting Standards (ESRS) specify the granularity of environmental reporting — AG-245's per-query measurement and annualised aggregation align with ESRS requirements.

UK SECR

The Streamlined Energy and Carbon Reporting framework requires qualifying UK companies to report energy consumption and greenhouse gas emissions in their annual reports. AI infrastructure energy consumption is within scope as part of the organisation's operational energy use. AG-245 ensures that AI-specific energy consumption is measured and available for inclusion in SECR reporting.

EU AI Act — Recital 27

Recital 27 of the EU AI Act notes that AI systems should be developed in a way that is resource-efficient and that minimises negative environmental impact throughout their lifecycle. While this is a recital rather than an operative article, it signals regulatory expectation that environmental impact is a consideration in AI system governance. AG-245 positions organisations to demonstrate compliance with this expectation.

EU Taxonomy — Regulation 2020/852

The EU Taxonomy establishes criteria for environmentally sustainable economic activities. Organisations claiming AI-related activities as sustainable must demonstrate environmental performance. AG-245's measurement, targets, and optimisation requirements provide the evidence base for Taxonomy alignment claims.

10. Failure Severity

Field	Value
Severity Rating	Moderate
Blast Radius	Organisational to societal — environmental externalities are borne by society broadly, not concentrated on specific individuals

Consequence chain: Failure to assess and manage environmental externalities from AI agent deployments produces two categories of harm. First, direct environmental harm: unmeasured and unmanaged AI infrastructure energy consumption contributes to greenhouse gas emissions without accountability. At scale — millions of queries per day across thousands of deployments — the aggregate energy consumption is material. Second, amplified environmental harm: AI agents whose optimisation targets ignore environmental costs drive physical-world resource use decisions that externalise environmental costs at scale — excess transport emissions from speed-optimised routing, waste from overconsumption-driving recommendations, energy waste from inefficient compute architectures. The regulatory consequence is non-compliance with mandatory sustainability reporting requirements (CSRD, SECR), which can result in financial penalties, qualified audit opinions, and investor confidence erosion. The reputational consequence is greenwashing risk — organisations that claim sustainability leadership while operating environmentally unmeasured AI deployments face credible criticism. The systemic consequence is that AI adoption without environmental governance accelerates the climate and resource challenges that affect everyone.

Cross-references: AG-051 (Fundamental Rights Impact Assessment) includes environmental rights in the broader rights assessment framework. AG-118 (Fair Treatment and Vulnerability) addresses the disproportionate environmental burden on vulnerable communities. AG-239 through AG-248 are sibling dimensions within the Rights, Ethics & Public Interest landscape. Environmental externality governance intersects with all operational governance dimensions to the extent that efficiency measures affect agent performance characteristics.