Understanding Sweep’s AI footprint: transparency through life cycle assessment

Artificial intelligence is transforming how businesses tackle sustainability challenges, but it comes with its own environmental cost. As AI becomes essential for sustainability data management and emissions reporting, sustainability leaders face a real tension: the tools helping them track carbon footprints have carbon footprints of their own.

That’s why Sweep conducted an Life Cycle Assessment (LCA) of our digital platform, with a particular focus on our AI features. The goal was straightforward: measure the environmental impact of running Sweep’s platform with credible, methodology-backed data that both our team and customers could trust.

Why we measured our digital footprint

This LCA was driven by two factors. Externally, there’s growing scrutiny around AI’s carbon footprint and legitimate questions about the environmental cost of large language models. Internally, it aligned with our Climate Compass 2025 pledge and our commitment to transparency.

The study followed standard LCA frameworks adapted for digital services, using ADEME’s Product Category Rule for cloud services as our methodological foundation. This rule provides a recognized industry standard, ensuring our results are consistent, comparable, and credible with auditors and sustainability teams.

What we measured and how

We followed a cradle-to-gate approach, focusing on our cloud infrastructure: Amazon Web Services for computing and AI workloads, and Snowflake for data storage and analytics. These represent where digital emissions actually occur – servers running continuously in energy-intensive data centers.

The process involved defining clear boundaries, collecting emissions and energy data from our infrastructure providers, converting that data into CO₂ equivalent using relevant emission factors, and calculating impact per functional unit. The entire study underwent structured internal review covering scope, data quality, and interpretation.

Defining the boundaries

User devices and data transmission were excluded, not arbitrarily, but because reliable data wasn’t available at the granularity we needed. Our guiding principle was clear: if we can’t measure it properly, we don’t estimate it. Including poor-quality data would undermine the credibility of the results we do have confidence in. This is standard LCA practice: focus on what’s significant and measurable.

We concentrated on AWS and Snowflake because they represent the largest share of operational spend, and cost is a reliable proxy for environmental significance. Other tools like GitHub, Dust, Datadog, or Slack were excluded because their emissions data isn’t publicly available at the level of detail we required.

Two metrics for two purposes

Because Sweep does two distinct things, we selected two functional units. The first – kg CO₂e per measurement – reflects core platform usage: uploading or submitting a sustainability data entry. The second – kg CO₂e per Sweepy credit – isolates the AI layer and lets us benchmark against other large language models.

This dual approach enables different conversations with different audiences: one about platform efficiency, one about AI sustainability. For customers, this means they can now account for Sweep’s platform in their own value chain emissions reporting, specifically within purchased software. The Sweepy credit metric lets them understand the environmental cost of their AI-assisted workflows.

Almost all our emissions come from infrastructure we don’t directly own. The majority of impact comes from services running in AWS, meaning infrastructure configuration decisions are the primary lever for reduction. This is an important framing: Sweep has influence over these emissions – through choices like cloud region, service selection, and query optimization – but not direct control.

One key technical detail involves our Production Valohai AWS account, which runs two types of workloads: AI (LLM-based features like Sweepy) and machine learning (traditional predictive models). Both share the same compute infrastructure, and we currently use a 30/70 split as our best estimate of how resources divide between them. This assumption matters significantly because cloud computing consumption is the primary driver of environmental impact in our model. Replacing this estimate with directly measured compute attribution is our top improvement priority.

Being honest about limitations

Every LCA has limitations, and transparency about them is essential. Ours are threefold: a partial scope excluding user devices and network emissions, a proprietary AWS model we can’t fully audit, and a 30% AI versus ML allocation based on estimation rather than direct measurement.

That said, the results remain valid as a directional baseline. We present them with underlying assumptions clearly stated – never as standalone figures. Conducting these analyses is a priority for future iterations, giving us and our customers increasing confidence over time.

The path forward

We’ve identified four priorities for the next version: adding data transmission emissions, disaggregating embedded hardware costs within AWS, replacing the 30% AI/ML assumption with measured data, and expanding Snowflake’s scope. This study is our first concrete step – a documented and repeatable baseline. Our internal review process formalizes structured evaluation, meaning every future iteration starts from a higher baseline of rigor.

We’re committed to engaging a third-party reviewer to assess the next iteration, further strengthening the credibility of our sustainability reporting.

What this means for the industry

Digital emissions can be measured – just like travel, energy use, or supply chain emissions. AI is becoming essential to modern sustainability work, but ignoring its environmental footprint would undermine the very goals companies are trying to achieve.

This isn’t about avoiding AI. It’s about using it purposefully, measuring its impact transparently, and making informed decisions about where and how to deploy it. When designed this way, AI can drive sustainability forward without becoming part of the problem.