· Updated September 18, 2026

Your AI Stack Is a Scope 3 Problem. Your ESG Team Doesn't Know It Yet.

The EU cut 61% of mandatory ESRS datapoints in July 2026. Scope 3 survived. Purchased AI services still belong in your Category 1 inventory, and most sustainability teams can't see the spend.

Salish Sea Consulting

The short version: Purchased AI services are Scope 3 Category 1 emissions under ESRS E1. That includes OpenAI, Microsoft Copilot, Azure OpenAI, Anthropic, and the Google Gemini API. The July 2026 ESRS rewrite cut mandatory datapoints from about 1,073 to 320. It kept E1's full Scope 1, 2, and 3 structure. The new value-chain data cap doesn't cover gross GHG emissions either. The spend sits in finance. The reporting obligation sits with ESG. Nothing connects them.

What Category 1 measures: Scope 3 Category 1 (purchased goods and services) covers the upstream emissions from producing everything a company buys. That includes the compute behind cloud and AI services.

Where the requirement lives:

  • FY2025 and FY2026 reports (2023 ESRS): disclosure requirement E1-6
  • FY2027 onward (revised ESRS): disclosure requirement E1-8

Why AI services belong in Category 1

When you pay OpenAI for API access or Microsoft for Copilot seats, you're buying a service that runs on energy-intensive compute. That compute has an emissions footprint. Under the GHG Protocol, which ESRS E1 is built on, that footprint is your Category 1 emissions.

This isn't an edge case or an interpretation question. Purchased services with material energy intensity are in scope. AI services are purchased services with material energy intensity.

Why the blind spot exists

AI spend gets tracked where spend gets tracked, which is finance. The OpenAI enterprise agreement lives in procurement. Azure OpenAI charges are itemized in accounts payable. Copilot licensing is bundled into an M365 contract.

None of this flows to ESG by default. ESG teams get emissions data through purpose-built channels. Utility bills cover Scope 2. Expense systems cover travel. Supplier questionnaires cover upstream Scope 3. AI services fit none of these pipelines. They arrived faster than reporting infrastructure adapted.

Emissions factors are a second problem. OpenAI doesn't publish per-token emissions data. Microsoft publishes sustainability commitments, not compute-to-emissions factors. Most teams end up applying industry-average factors to cloud spend. That's a defensible "best available data" approach, but only if you have the spend data to apply it to.

What "no data" looks like in an audit

CSRD reports go through limited assurance. Auditors don't just check that a Category 1 number exists. They check whether the boundary is appropriate.

If your Category 1 disclosure covers materials and professional services but skips AI services, expect questions. The revised ESRS does let you exclude activities that aren't significant drivers of emissions. But you have to disclose that you used the relief. You also can't show AI spend is insignificant without knowing what the spend is.

So "no data" isn't a safe position. It has to be defended. And defending it takes the same data you'd need to report it.

Who reports when, after the Omnibus

The Omnibus I Directive (EU) 2026/470 narrowed CSRD scope in March 2026.

  • Wave 1 companies above the new thresholds keep reporting for FY2025 and FY2026. Member states may exempt those below the thresholds.
  • EU companies with more than 1,000 employees and more than €450M turnover apply the revised ESRS from FY2027, with first reports in 2028.
  • Non-EU parents with more than €450M EU turnover and an EU subsidiary or branch above €200M report from FY2028.
  • Listed SMEs are out of mandatory scope.

Fewer companies are in scope. The ones still in scope face the same Scope 3 requirement. FY2027 data starts generating on January 1, 2027. Every API call from that date is in the reporting period.

What comes next

CSRD isn't asking too much. AI services entered companies through procurement and IT without triggering ESG data capture. The spend is real. The emissions are real. The infrastructure to connect them hasn't been built.

My next piece covers the method: how to apply the spend-based approach to purchased AI services under the GHG Protocol.

If your CSRD Scope 3 inventory doesn't account for AI services yet, get in touch.

Sources