SFDR, CSRD, and AI: How ESG Reporting Requirements Govern AI Agents in Sustainable Finance

ESG investing has generated its own regulatory stack: SFDR (Sustainable Finance Disclosure Regulation) requires fund managers to classify products under Article 6, 8, or 9 and disclose how sustainability factors are integrated. CSRD (Corporate Sustainability Reporting Directive) requires large EU companies to report on sustainability using ESRS (European Sustainability Reporting Standards).

Both regulations increasingly involve AI: ESG scoring models, portfolio screening algorithms, automated ESRS data collection, and natural language processing of sustainability disclosures. Where AI is involved, the governance and audit requirements of these regulations apply to the AI layer.

SFDR's AI-Relevant Requirements

Principal Adverse Impact (PAI) Indicators SFDR requires fund managers to disclose how they consider principal adverse impacts of investment decisions on sustainability factors. When an AI model is doing the ESG scoring or PAI analysis, the model's methodology must be disclosed and the analysis must be auditable.

An AI agent producing ESG scores for SFDR compliance needs:

• Documented methodology (the model's decision logic)
• Audit trail of each scoring decision (what data, what output, when)
• Human review for Article 8/9 classification decisions

Article 8/9 Classification Obligations If a fund is classified as Article 8 (promoting environmental/social characteristics) or Article 9 (sustainable investment objective), the AI systems supporting that classification must produce reliable, consistent outputs — and those outputs must be defensible to ESMA.

Greenwashing enforcement under SFDR is active. ESMA and national NCAs have investigated funds that classified themselves as Article 8/9 without adequate underlying documentation. AI-generated ESG scores that aren't auditable are a greenwashing risk.

CSRD's AI-Relevant Requirements

CSRD requires large EU companies to report on:

• Environmental matters (climate change, biodiversity, water, pollution)
• Social matters (workforce, value chain workers, affected communities)
• Governance (business conduct, anti-corruption)

Using ESRS (European Sustainability Reporting Standards), the reporting must be consistent, comparable, and subject to limited assurance (phasing to reasonable assurance).

AI systems are increasingly used for:

• Automated data collection from supply chains
• Carbon footprint calculation
• Natural language extraction from ESG questionnaires
• Gap analysis against ESRS standards

Each of these creates an audit trail requirement: the CSRD report must be traceable to source data. When AI is in the chain from source data to reported figure, the AI's processing step must be auditable.

The ESG Agent Manifest

For an AI agent involved in ESG analysis or CSRD reporting:

{
  "capabilities": {
    "requested": ["data:read", "report:generate", "report:submit"]
  },
  "constraints": {
    "require_human_approval": ["report:submit", "report:generate"],
    "audit_retention_days": 2555
  },
  "policy": {
    "tool_allowlist": [
      "esg_data_provider",
      "carbon_calculator",
      "esrs_validator",
      "audit_evidence_store"
    ],
    "jurisdiction_allowlist": ["EU"],
    "require_capability_token": true
  }
}

require_human_approval on report:generate and report:submit means the sustainability reporting agent cannot produce or file CSRD reports without human review. This is the assurance requirement implemented at the infrastructure layer.

audit_retention_days: 2555 (7 years) covers both CSRD's 10-year record retention expectation and SFDR's ongoing disclosure requirements.

The Greenwashing Risk

ESMA's greenwashing supervisory report (2024) identified specific AI-related greenwashing risks:

1. Opaque ESG scoring — AI models that produce ESG scores without explainable methodology
2. Inconsistent application — AI that applies different criteria to different assets without documentation
3. Data quality gaps — AI trained on unreliable ESG data that propagates errors into disclosures

For all three risks, the mitigation is the same: an audit trail that lets you explain, for any specific ESG score or CSRD data point, what data was used, how it was processed, and who reviewed the result.

UAPK's interaction records capture the input data, the processing steps (via tool execution records), and the output for every agent action. That audit trail is the greenwashing defense.

Fund Manager SFDR + MiFID II Stack

For EU fund managers, SFDR and MiFID II operate together. MiFID II requires suitability assessments that incorporate sustainability preferences (since 2022). SFDR requires disclosures on how sustainability is integrated. Both apply to the same AI-driven portfolio management functions.

The UAPK qualification funnel recommends both SFDR and MiFID II for fund managers in the EU. The manifest questions for each framework are distinct — MiFID II focuses on trade execution and best execution; SFDR focuses on disclosure and impact analysis. The policy fields they generate are compatible.

Double Materiality and AI

CSRD's double materiality principle requires companies to assess both:

1. How sustainability issues affect the company (financial materiality)
2. How the company affects the environment and society (impact materiality)

AI models conducting double materiality assessments face a specific challenge: materiality is a judgment, not a calculation. An AI model that mechanically applies materiality thresholds without capturing the judgment context creates a defensibility gap.

UAPK's HITL requirement for report:generate ensures that double materiality assessments get human review before they're finalized. The audit log captures both the AI's analysis and the human reviewer's sign-off.