Scope
This guide helps your team identify and assess environmental governance gaps in AI systems, focusing on the inference phase, the continuous operation of AI models after deployment. It applies to organizations deploying AI directly or using third-party vendors. You don't need environmental science expertise; standard audit methodology applies to a risk most assurance functions haven't cataloged.
Key Concepts and Definitions
Training Phase: The resource-intensive process of building an AI model using large datasets and compute resources. Happens once or infrequently before deployment. Measurable and bounded.
Inference Phase: The continuous operation of a deployed AI model, processing queries, automating decisions, and generating outputs. Runs at scale for the system's entire operational life. Unmeasured in most regulatory frameworks.
Governance Gap: A material organizational exposure outside current regulatory requirements, creating an unowned and unmeasured risk.
Vendor AI Consumption: The use of third-party AI services where the model training and inference infrastructure is operated by an external provider, shifting measurement responsibility to contract terms rather than internal controls.
Requirements Breakdown
Current Regulatory State
No global framework mandates inference-phase environmental reporting. AI systems can operate billions of times daily without any obligation to report energy consumption or water usage.
The EU AI Act illustrates the gap: Environmental provisions in early drafts were diluted during negotiation, and binding inference-phase reporting requirements didn't survive into the final text.
This creates an assurance problem: Organizations carry an exposure that is unmeasured, unreported, and typically unowned.
What This Means for Audit Programs
Your annual audit plan should address:
- Risk ownership: Whether any function owns AI environmental impact as a risk category.
- Measurement capability: Whether the organization can quantify ongoing operational footprint, not just one-time training costs.
- Vendor risk: Whether third-party AI contracts include verifiable environmental disclosure requirements.
- Claims substantiation: Whether external statements about AI efficiency or sustainability can be independently verified.
Implementation Guidance
Step 1: Map AI Consumption Across the Organization
Start with an inventory. Identify where AI systems are deployed or consumed, including:
- Internal AI products serving customers or employees.
- Third-party AI services embedded in vendor platforms (CRM tools, customer service automation, document processing).
- Development environments where AI models are being built or tested.
For each instance, document whether it's internally hosted or vendor-provided. This determines whether you're auditing internal controls or third-party contracts.
Step 2: Test for Risk Ownership
Ask directly: Who owns the environmental footprint of your AI systems?
Review risk registers, role descriptions, and committee charters. In most organizations, the answer will be "no one," which is itself the finding. An unowned risk defaults to ungoverned.
If ownership exists on paper but not in practice, test whether the assigned function has:
- Defined metrics for accountability.
- Access to the data needed to measure those metrics.
- Reporting lines that create consequence for non-performance.
Step 3: Assess Measurement Capability
For internally deployed AI:
- Can your organization measure energy consumption attributable to inference, separate from general infrastructure load?
- Are water usage metrics (for cooling) tracked and allocated to AI workloads?
- Is there a mechanism to track consumption per query, per user session, or per operational hour?
For vendor-provided AI:
- Do contracts require the vendor to disclose environmental impact?
- Are disclosed figures standardized (aligned to a framework like the GHG Protocol) or vendor-defined?
- Is there a contractual right to audit vendor-provided environmental data?
A measurement capability that stops at training has captured the bounded problem and missed the one that scales.
Step 4: Review External Claims
If your organization makes public statements about sustainability, efficiency gains, or environmental targets that touch AI:
- Identify the specific claim (e.g., "Our AI-powered platform reduces carbon footprint by X%").
- Trace the claim to its source data.
- Test whether that data is independently verifiable or derived from unaudited vendor assertions.
An unverifiable claim is a greenwashing exposure. It sits with your organization, not the vendor that supplied the number.
Step 5: Evaluate Vendor Contracts
For each third-party AI service, review contract terms for:
- Disclosure obligations: Does the vendor commit to reporting energy or water consumption?
- Measurement standards: Are metrics defined using a recognized framework, or are they vendor-specific?
- Audit rights: Can your organization independently verify vendor-provided environmental data?
- Update frequency: Are disclosures one-time or refreshed as usage scales?
Most contracts will have none of these provisions. That's the current baseline, and it's a vendor risk gap your audit function should document.
Common Pitfalls
Treating this as a sustainability initiative rather than a governance exposure. Sustainability teams may own environmental strategy, but ungoverned risk is an audit matter. Don't defer to another function when the risk is unmeasured and unowned.
Waiting for regulation to define the requirement. The EU AI Act demonstrates that environmental provisions can be diluted or excluded during negotiation. Relying on future regulation means accepting the exposure until someone else forces the issue.
Accepting vendor self-reporting without audit rights. A vendor's claimed efficiency metric is not verifiable data unless your contract includes the right to substantiate it.
Confusing training metrics with operational risk. Training happens once; inference runs continuously. If your organization can only describe the footprint of building a model, you've measured the smaller problem.
Filing this under "emerging risk" and deferring action. AI systems are already deployed at scale. The risk isn't emerging, it's live and unmeasured.
Quick Reference Table
| Audit Question | What You're Testing | Red Flag |
|---|---|---|
| Who owns AI environmental impact? | Risk ownership and accountability | No named owner or risk register entry |
| Can we measure inference separately from training? | Measurement capability for scaled operations | Only training-phase metrics exist |
| Do vendor contracts require environmental disclosure? | Third-party risk controls | No disclosure clause or audit rights |
| Are external AI sustainability claims verifiable? | Greenwashing exposure | Claims trace to unaudited vendor data |
| Is environmental data standardized or vendor-defined? | Comparability across vendors | Each vendor uses proprietary metrics |
| Does anyone report AI environmental metrics to leadership? | Governance visibility | No reporting line exists |
When to escalate: If your testing reveals that AI environmental impact is unowned, unmeasured, and the organization has made external claims about its sustainability posture, that's a material governance gap requiring executive and board visibility. Frame it as a risk the organization is carrying without assurance, not as an environmental compliance failure. The latter invites delay; the former demands a control response.
