Proposal
This policy proposal establishes environmental sustainability requirements for organisations developing, deploying, or operating AI systems under AIMSS certification. AI systems can consume significant computational resources during training, fine-tuning, testing, and inference. These activities may create environmental impacts through energy consumption, greenhouse gas emissions, water usage, hardware demand, and electronic waste. Organisations seeking certification must therefore demonstrate that environmental impact is measured, managed, and reduced where practical.
The organisation shall maintain an AI Environmental Impact Policy approved by senior management. The policy shall define when environmental assessment is required, who owns environmental reporting, which metrics must be tracked, and how environmental considerations are included in AI design and deployment decisions. The policy shall apply to material AI systems, including large-scale model training, high-volume inference systems, repeated experimentation, and AI services that require significant cloud or hardware resources.
Before approving a material AI project, the organisation shall assess expected compute needs and environmental impact. The assessment shall include estimated training or fine-tuning resources, expected inference volume, infrastructure location where known, hardware requirements, and anticipated energy usage. For high-impact systems, the organisation shall also estimate greenhouse gas emissions using a documented methodology. Where exact data is unavailable, reasonable estimates may be used, but assumptions, data sources, and calculation boundaries must be recorded.
The organisation shall adopt a responsible compute approach. Project teams must consider whether the same business or social objective can be achieved through a smaller model, more efficient architecture, lower-frequency inference, model reuse, data optimisation, or non-AI alternative. For large-scale AI systems, environmental review shall be included in the release approval process. Approval records must explain why the selected model size, compute intensity, and deployment method are proportionate to the expected benefit.
Ongoing monitoring shall be required for material deployed systems. Organisations shall track energy usage or proxy indicators, such as cloud compute consumption, accelerator hours, inference volume, or model utilisation. These metrics shall be reviewed periodically to identify waste, inefficiency, or unnecessary scaling. Where material environmental impact is identified, the organisation shall establish reduction measures, which may include model compression, caching, efficient hardware selection, renewable-energy procurement, workload scheduling, or retirement of underused systems.
The policy shall also address hardware lifecycle impacts. Organisations using dedicated AI hardware shall document procurement, utilisation, expected lifespan, disposal method, and recycling or reuse practices. Suppliers providing AI infrastructure shall be assessed for relevant environmental information where available.
Public transparency shall be proportionate. Organisations operating high-impact AI systems shall publish an environmental summary describing the categories of AI activity measured, the methodology used, key reduction initiatives, and progress against targets. The summary should avoid misleading claims and clearly distinguish measured data from estimates.
Certification evidence shall include environmental impact assessments, compute logs, cloud usage reports, emissions calculations, methodology documents, reduction plans, procurement records, supplier information, and management review minutes. Failure to measure material AI environmental impact should be treated as a major nonconformity. Deliberate misstatement of environmental claims, concealment of material impact, or repeated failure to act on significant waste should be treated as a critical issue for certification review.