Accounting for AI Inference in Corporate GHG Inventories: A Four-Tier Methodology for Scope 3 Category 1 Reporting

The Hidden Cost of the AI Prompt

Companies using AI services need to report their carbon footprint. However, it is difficult to know exactly how much energy those services consume. While generative AI has moved from a novelty to a core business tool, the environmental math behind it remains murky. A new study by Guillermo Llopis proposes a structured way to bridge this gap. It matches the precision of an emission estimate to the specific data a company can actually access.

Matching precision to data availability

The central question the study addresses is how corporations can accurately report greenhouse gas (GHG) emissions. Specifically, it looks at emissions generated by AI inference—the process of running a trained model to generate a prediction or response. These fall under Scope 3 Category 1 emissions, which include purchased goods and services like API subscriptions and enterprise AI seats.

Under the Corporate Sustainability Reporting Directive (CSRD), these disclosures are becoming mandatory. Unlike a utility bill that shows kilowatt-hours, an AI subscription often only shows dollars spent or messages sent. The researcher asks how a sustainability manager can build a defensible inventory when the provider won't reveal raw energy data. The goal is to create a methodology that scales from "best guess" to "high precision" without requiring a data science degree.

The flaws in economic proxies

Until now, the field has largely relied on two extremes. Companies either omit AI emissions entirely or use generic economic input-output (EEIO) factors. These factors are mathematical shortcuts based on input-output models (systems that estimate emissions by linking economic spending to sectoral environmental impacts).

The paper finds that this approach is fundamentally mismatched for AI. EEIO factors are calibrated to capture the entire economic footprint of a sector. This includes labor, R&D, and profit margins. The authors report that applying these spend-based factors to AI inference can overestimate actual emissions by 10–40× compared to physical, electricity-based methods. Relying on these proxies treats a digital prompt as if it carried the weight of an entire IT department's payroll.

Building a four-tier hierarchy

To solve this, the study develops a four-tier framework that prioritizes data quality. The investigation begins by mapping the different ways companies interact with AI. Some use direct APIs where they can see exact token counts (the basic units of text processed by an LLM). Others use "seats" in software like Microsoft Copilot where usage is obscured.

The framework follows a logical descent, illustrated in .

The researchers suggest starting at Tier 3, which uses provider-certified carbon reports. However, the paper notes that as of mid-2026, no major provider offers this as a standard feature. If that is unavailable, the study moves to Tier 2a, which uses exact token counts from billing logs. Tier 2b uses proxies like message counts to estimate tokens. Finally, Tier 1 serves as the spend-based fallback for SaaS products where no usage data is visible.

To make these tiers functional, the authors derive new emission factors. They take GPU-level energy benchmarks from the ML.ENERGY Leaderboard v3. They scale them to the full facility level by accounting for Power Usage Effectiveness (PUE). PUE is a ratio describing how much energy is used by IT equipment versus cooling and lighting .

These energy figures are then multiplied by regional grid carbon intensities. This produces a final kilogram of $\text{CO}_2\text{e}$ per million tokens.

Location matters more than model size

The findings reveal that the environmental impact of AI is driven more by geography than by model complexity. The authors report a 13× spread in carbon intensity for the same workload depending on the cloud region used. For example, performing Class B inference in Singapore is roughly 13 times more carbon-intensive than doing the same task in Sweden .

This regional dominance is a critical insight. While moving from a "small" model to a "frontier" model increases emissions, the jump in carbon intensity from shifting grids is much larger.

The study also uncovers a "water-carbon paradox" that current ESG (Environmental, Social, and Governance) tools often miss. While Sweden offers one of the lowest carbon intensities due to its hydro-dominated grid, it also shows the highest water footprint in the dataset .

This occurs because hydroelectricity involves significant water loss through reservoir evaporation. A company optimizing solely for carbon might inadvertently spike its water consumption. This trade-off is vital for firms operating under strict water-stress reporting requirements.

Moving from estimation to audit

The implications of this work are twofold. First, for practitioners, the paper suggests that the most effective lever for reducing AI emissions is selecting cloud regions with lower carbon intensities. Second, for regulators and providers, the study identifies a "transparency roadmap" [Table 8]. To move from imprecise estimates to audited reality, providers must surface token counts and specific infrastructure regions in customer billing dashboards.

If this methodology gains traction, the compliance challenge for AI will shift from a question of magnitude to a question of documentation. As demonstrated in the paper's case study of a 200-person firm, the total emissions are often quite small—well below 1 $\text{tCO}_2\text{e}$. However, the requirement to prove those numbers is very real.

The immediate next step for organizations is to audit their existing AI subscriptions. Determine which ones allow for Tier 2a token-based reporting. Identify which are stuck in the inefficient Tier 1 spend-based loop.

Figures from the paper