Getting a Grasp on Your Company’s AI Emissions

As companies integrate AI into everyday business activities, understanding and managing the environmental impacts associated with its use is becoming increasingly important. Several factors are driving this shift, including evolving disclosure requirements, growing scrutiny of how AI-related impacts are reflected in sustainability reporting, and broader expectations around responsible AI use.

Regulatory requirements are one reason to start evaluating AI-related emissions. California’s SB 253 will require in-scope companies to report Scope 3 emissions, where AI-related emissions would be categorized. The revised European Sustainability Reporting Standards, published on July 3, 2026, go a step further by specifically addressing emissions associated with digital infrastructure. Under ESRS E1, companies, where applicable, are required to disclose GHG emissions from purchased cloud computing and data center services as a subset of Scope 3, Category 1 (Purchased Goods and Services).

For companies with Scope 3 reduction targets or transition plans, growing AI-related emissions could affect progress toward targets and require companies to reassess existing reduction strategies. The challenge is not only that AI use may expand, but that its impact may remain obscured within broader purchased services calculations until it becomes more material. As provider disclosures improve, companies will be able to quantify AI-related emissions more accurately, potentially revealing a larger contribution to Scope 3 than anticipated. If AI-related emissions become meaningful, they may require adjustments to existing transition plans and planned emissions reduction measures. Understanding the potential scale of AI-related emissions early can help companies assess whether AI could become a significant driver of Scope 3 emissions and account for that possibility.

One of the immediate challenges, however, is determining how to calculate AI-related emissions.

The challenge in determining the total size of AI emissions

To understand how corporations should start accounting for their AI emissions, we need to understand the landscape and information that exists. Different providers, models, and use cases (generating an image versus writing code) all consume energy and therefore produce emissions at wildly different rates. To understand the size of the ‘AI emission’ problem, we need to understand those nuances, but structural gaps in the data make this impossible. While we can't yet publish a single "AI emits X tonnes per year" number for the whole market, we can say that the measurable footprint is already material and is only going to grow.

A useful way to think about the whole problem: an annual emissions estimate equals to per-unit intensity × total volume × grid intensity.

As a simple illustrative example, let’s say a chatbot query uses approximately 0.0005 kWh (per-unit intensity, note actual per-unit intensity varies widely), the service handles approximately 500 million queries per year (total volume), and it runs on a grid averaging approximately 400 gCO2e/kWh (grid intensity). Annual emissions = 0.0005 kWh × 500,000,000 × 400 gCO2e/kWh = approximately 100,000 kg CO2e (100 metric tons) per year.

The challenge however is that for most AI modalities, we can generally approximate an intensity factor per-query, but total query volume is missing.

Alternatively, tokens are a universal unit of volume. Providers meter and bill in tokens, so a headline figure like Google's ~3.2 quadrillion tokens per month is an accumulation of chat, coding, image, video, and Search together. However, they have radically different emissions implications based on usage and type. In this case, we have the total volume, but not the per-unit intensity. Using transportation emissions as an analogy, we are reporting the equivalent of total miles and omitting the breakdown by vehicle type.

Finally, note this only covers emissions from AI usage, the energy used to answer a chat prompt, generate an image, etc. and not the emissions associated with training models because training is not part of the per-use footprint, and how training emissions should be allocated to an end user is still unsettled. See here for a deep dive on the data that exists, including the variance across per-unit intensity.

Getting a grasp on your AI-related emissions

Despite rapid growth in corporate AI adoption, disclosure of AI-related emissions by end-user companies remains limited, due to the challenges discussed. On the supply side, a few providers are starting to disclose their measurements: See Google's measurement of Gemini (Elsworth et al. 2025) and Salesforce is beginning to publish energy and carbon figures in its AI model cards in 2026 (using the AI Energy Score method it co-developed), the nearest thing to standing per-model disclosure. These developments demonstrate that more detailed measurement is possible, but this level of information is not yet consistently available across AI and models.

However, starting to account for your AI usage emissions is possible.

CNaught was able to measure their own AI usage emissions with the best available data (see full methodology here), focusing on their largest provider by far, Anthropic’s suite of Claude tools. This measurement was not without limitations and necessary assumptions, however. The largest data gaps included the lack of:

• Time-based usage data (at hourly or daily resolution): Monthly summaries can’t tell you when compute happened, which means they can’t be matched to how clean the grid was at the time. This will only grow more important as the GHG Protocol considers a shift to hourly matching for Scope 2 accounting.
• Grid routing & location data: Without knowing where the compute ran, every emissions estimate has to assume an average that may be far from the actual number, based on where the queries are processed. This data doesn’t need to be precise data center locations - just precise enough to match grid intensity maps.
• Model specific energy consumption information: This is the single biggest gap - if providers disclose their energy per query or per token by model, customers could replace assumptions with actual factors. Model providers already release model cards with AI performance benchmarks. They should include energy efficiency alongside these metrics.

In the end, CNaught’s calculated total 2025 organizational emissions came to 111 tonnes CO₂e (which were addressed using high-integrity carbon credits and they are Climate Label-certified), of which AI usage accounts for roughly 2.8 tonnes a year, about 2.5% of that footprint. While AI isn't a material driver of CNaught’s emissions today, it can scale fast: Product and Engineering teams currently account for the largest share of CNaught’s AI emissions, and as tools like Claude Cowork bring AI workflows to non-technical teams, usage is climbing across the board. If every employee used AI as heavily as CNaught’s single heaviest user does today, CNaught’s AI emissions would nearly quadruple to roughly 9% of their total footprint.

So what can companies do?

For many companies, AI-related emissions may still represent a relatively small share of the overall GHG footprint. However, AI adoption is expanding rapidly across business functions, and companies should not wait for complete data or standardized methodologies before beginning to understand its impact. Starting now allows companies to establish a baseline, identify the largest sources of AI usage, and build processes that can be refined as provider disclosures and measurement methods improve. It can also help companies determine whether AI-related emissions are becoming significant enough to affect Scope 3 targets or transition plans.

Early measurement of AI-related emissions can also help improve the quality of the underlying data. The more customers request information on energy consumption, emissions, and the location of computing infrastructure, the stronger the incentive for AI providers to make this information available. Companies can therefore use their role as customers not only to improve their own inventories, but also to encourage greater transparency across the market.

Practical advice to get started:

• Start with data that is already available. Identify the major AI tools used across the organization and determine what usage information each provider discloses, such as token consumption, number of queries, model usage, or other activity metrics. Establish a regular process for collecting this data. Since some providers retain detailed usage data for only a limited time (e.g., 90 days), regularly downloading this data helps ensure that information for developing the inventory is complete.
• Develop an initial emissions estimate. Although supplier-specific factors are not yet available, existing research and benchmark data can support an indicative estimate. Companies should document key methodologies, assumptions, and limitations so calculations can be refined as better data become available.
• Find tools that help you track it. A year ago the consensus was that AI-detection software was too inaccurate to rely on. That's changed on two fronts: detection algorithms have improved sharply (tools like Panagram are now used in peer-reviewed studies) and services like GreenPT are building sustainability into AI usage from the start. CNaught’s own tool, Carbonlog, helps you track emissions from Claude Code usage, with a methodology that will continue to be refined as more data become available.
• Benchmark against your existing inventory. Compare how much AI activity contributes relative to other emissions sources in your inventory. This can help entities prioritize emission reduction efforts based on where the impact is greatest and where the entity has the most ability to influence outcomes.

Once you’re measuring and you understand the magnitude of your AI-related emissions, real opportunities open up that you can start acting on.

• Choose where and when queries run where providers allow it — routing compute to cleaner grids and at specific times of day is a direct lever on emissions.
• Prioritize the largest sources of impact. Not all AI activity has the same energy requirements. Rather than attempting to measure every use case with the same level of detail, companies can initially focus on AI providers, business functions, and applications with the greatest usage or expected environmental impact. This can help concentrate data collection and supplier engagement where they are most likely to improve the accuracy of the inventory.
• Choose the right model for the task. More computationally intensive models are not necessary for every use case. Employees may be able to use smaller or less resource-intensive models for relatively simple tasks while reserving advanced reasoning models for applications that require them. Companies can incorporate this principle into employee AI guidance and training, helping users understand that model selection can affect not only cost and processing time, but also energy consumption and associated emissions.
• Engage your suppliers. Request information from key providers on customer-level AI-related emissions and water use, as well as their renewable energy and efficiency commitments. These requests can be incorporated into supplier questionnaires, procurement processes, contract discussions, and broader vendor governance.
• Benchmark against your existing inventory. Compare how much AI activity contributes relative to other emissions sources in your inventory. This can help entities prioritize emission reduction efforts based on where the impact is greatest and where the entity has the most ability to influence outcomes.

Ready to begin quantifying your AI-related emissions? Reach out to Sodali & Co and CNaught to help you get started.

About Sodali & Co

Sodali & Co advises companies on using sustainability as a lens to power business value. When it comes to AI-related emissions, Sodali & Co guides companies from data discovery to supplier engagement, and prepares companies for its implications to sustainability reporting, target setting, and reduction strategies.

About CNaught

CNaught is the easiest way for sustainability teams of any size to purchase and manage an affordable, science-backed portfolio of high-integrity carbon credits. CNaught helps companies offset emissions from AI usage alongside unavoidable emissions from the rest of their operations, with every project backed by the CNaught Guarantee, and AI-powered reporting and marketing tools included.