AI's environmental footprint has become one of the most talked-about sustainability topics in the last two years, but much of that conversation runs on assumptions that don't hold up once the emissions math is examined. Two misconceptions in particular tend to shape how companies think about this, often incorrectly.
While this piece focuses specifically on carbon emissions, it's worth noting that AI also carries other significant environmental and social impacts, such as water consumption for cooling and local environmental justice concerns, that deserve attention in a broader sustainability strategy.
Misconception one: Everyday use carries the same weight as training
The most significant driver of AI-related emissions isn't day-to-day usage. It's the training process that builds the model in the first place. Training a large model requires enormous compute run continuously over an extended period, and that process accounts for the overwhelming majority of a model's total emissions footprint. By comparison, the emissions tied to an individual prompt or query, once a model is already trained and deployed, are comparatively small.
That creates a genuinely uncomfortable dynamic. The heaviest emissions have often already happened by the time an everyday user interacts with a tool. In that sense, the sustainability question shifts. It's less about whether using an already-built tool for a simple task is inherently harmful, and more about avoiding unnecessary additional training and scaling, and using existing models as efficiently as possible rather than defaulting to the largest, most resource-intensive option for every task.
What that means in practice: right-sizing the model to the task
This is where model choice becomes a real, actionable sustainability lever rather than an abstract concern. Not every task requires the largest, most capable model available. A quick back-and-forth question doesn't carry the same computational load as generating a long report or a complex creative asset, and using a lighter model for lighter tasks reduces the energy and emissions associated with that specific request, without requiring anyone to stop using AI altogether.
Some companies have already started building internal guidance around this. One example is an internal AI usage guide built to help employees understand which model fits which task, steering routine requests toward lighter models and reserving larger, more resource-intensive models for work that actually requires that level of capability. It's a practical, low-friction way to reduce a company's AI-related footprint without asking anyone to change their workflow in any meaningful way.
Misconception two: AI is the only emissions story worth focusing on
The intensity of public conversation around AI's environmental impact can create the impression that it's the primary sustainability issue companies need to address, when in reality it's one factor among many. Every part of running a business, virtual or not, carries emissions. Software beyond AI tools, business travel, facilities, and countless other day-to-day operational choices all contribute meaningfully to a company's total footprint, and none of that goes away simply because AI is getting more attention.
Letting AI dominate a sustainability strategy risks losing sight of larger, more established sources of emissions that deserve just as much, if not more, ongoing attention. A company that spends significant energy scrutinizing its AI usage while ignoring something like unnecessary business travel or inefficient facilities isn't necessarily making meaningful progress. It's just responding to whichever topic is generating the most public conversation at the moment.
How to actually measure it
For companies trying to move past assumptions and toward real numbers, measurement comes down to emissions factors: a quantified relationship between a unit of activity and the emissions associated with it. In the context of AI usage, the most common current approach uses spend as that unit, meaning emissions are estimated based on dollars spent on a given AI service, similar to how many broader software and services costs are treated.
Organizations such as the EPA and equivalent international bodies publish these factors annually, along with other publicly available and subscription-based sources. Because these factors are recalculated each year based on the prior year's data, and AI usage has grown faster than these factors have been updated, current figures likely understate the actual emissions tied to AI usage today. That's expected to improve as more recent data gets incorporated into future factor updates.
For companies that need more precision than a spend-based estimate provides, some organizations are developing methodologies that incorporate actual token or API usage data where vendors make it available. This moves the estimate closer to actual usage rather than relying purely on dollars spent, though it depends on having a vendor willing and able to share that level of detail.
The larger point holds regardless of which measurement approach a company uses: more data, even imperfect or partial data, produces a better estimate than no data at all. Waiting for perfect emissions factors or full vendor transparency isn't a reason to delay taking a first, reasonable pass at understanding where AI fits into a company's overall footprint.
FAQ
-
For most companies, AI tool usage falls under Scope 3, specifically the purchased goods and services category, since it's treated as a purchased service rather than a directly operated resource.
-
Most do not, particularly larger providers. Smaller AI vendors that build on top of existing models, sometimes called AI wrapper companies, often don't have emissions data to share either, since the underlying model's emissions are technically part of their own Scope 3.
-
Often, yes. Asking a vendor for usage details like token volume, underlying models used, or electricity consumption can provide enough information to build a reasonable estimate, even when the vendor has no formal emissions figure to share.
