Etsy has described its approach to the environmental impact of internal AI use. For other teams, the practical starting point is a transparent inventory that separates measured data from estimates.
What Etsy’s engineers describe
In their September 29 article, Etsy engineers Emily Sommer and Sam Brundrett discuss sustainable computing and internal AI use. They explain that provider data remains uneven and describe work on resource optimization and impact accounting. The article’s stated scope excludes sellers’ AI use outside Etsy-provided tools. This is Etsy’s account of its approach, not an independent assessment of every supplier. Etsy engineering article.
An estimate can support a decision without becoming a direct measurement. The report needs to say which it is. It should also state what falls outside its boundary.
Start with a usable inventory.
For each workflow, record the provider, model, time period, request volume, and available usage data. Keep the methodology version with the result. If a provider supplies environmental data, retain the stated scope and units rather than combining unlike figures without explanation.
Do not translate tokens into a universal emissions number. Any conversion depends on assumptions that must be visible. Missing data should appear as missing data, not as zero impact.
A reporting checklist
- Define which tools and activities the inventory covers.
- Separate supplier measurements from calculated estimates.
- Keep energy and greenhouse-gas emissions as distinct quantities.
- Record the assumptions behind each conversion.
- Label unavailable provider data clearly.
- Compare equivalent tasks over equivalent periods.
- Report total impact separately from impact per task.
- Recalculate when the method or underlying assumptions change.
There is also a workload question. Count retries and discarded outputs when comparing two ways to complete a task. A smaller nominal request can be a poor choice if it repeatedly fails the acceptance test. Any proposed efficiency change still needs a quality check.
For a broader evaluation workflow, see our AI-content audit guide and AI tools directory.
Editorial takeaway: publish the boundary, method, and uncertainty alongside the estimate. A precise-looking number is not useful if readers cannot tell what it represents.
