AI-based system for automated ESG reporting
The CSRD requires over 50,000 EU companies to publish reports in accordance with the ESRS (European Sustainability Reporting Standards) from 2024–2026. The volume of disclosures has increased three to five times compared to the voluntary GRI standards. A team of three sustainability specialists is physically unable to cope with the quarterly data collection, verification, and narrative generation for the 200-page report.
LLM pipeline for narrative generation
Architecture: data → text without hallucinations
The main risk of LLM in ESG reporting is hallucinated numbers. The regulator and auditor require verifiability of every figure. Solution: RAG architecture with a strict citation policy.
ESG Data Warehouse (Snowflake)
↓
dbt mart: предрассчитанные disclosure metrics
↓
Vector store (pgvector): описания ESRS требований
↓
LLM (GPT-4o / Claude 3.5 Sonnet)
↓
Нарратив с inline citations [data_point_id]
↓
Верификационный слой: каждая цифра → lookup в БД
If the LLM includes a number that isn't in the retrieval context, the verification layer throws an exception and doesn't publish the paragraph. In practice, 94% of narrative paragraphs are generated correctly without manual editing, according to testing data on historical reports.
Mapping data to standards
ESRS, GRI, TCFD, and SASB—different standards require the same data in different formats and contexts. The ML component, a fine-tuned text classifier (BERT), determines which disclosure requirements each data point falls under. One indicator (e.g., energy consumption by source) is automatically mapped to ESRS E1-4, GRI 302-1, and SASB energy metrics—without manual cross-referencing.
Double Materiality Assessment
The CSRD requires an assessment of: (1) how ESG factors impact a company's finances (financial materiality), (2) how a company impacts society and the environment (impact materiality). This is a matrix of 40–80 topics.
Automation of stakeholder survey
Stakeholder surveys are a mandatory element of DMA. NLP pipeline: - Response collection via a survey platform (SurveyMonkey, Typeform) - Topic modeling (BERTopic) based on open-ended responses → ESG topic clusters - Sentiment analysis for each topic - Automatic topic ranking by frequency and intensity score
In a manufacturing company case study, processing 450 open-ended questionnaires took two hours, compared to three weeks manually. Twenty-three themes were identified, ranked by materiality score.
Industry benchmarking
Peer comparison: scraping public ESG reports of competitors + LLM extraction of key KPIs → comparative tables. Allows you to determine which topics industry players consider material to calibrate your own assessment.
Automation of data collection
Supplier data collection
CSRD Scope 3 requires data from suppliers. An LLM-based email agent generates personalized data requests, tracks responses, sends reminders, and parses reply emails and documents. The response rate increased from 23% (manual) to 41% (AI-assisted follow-up) in a pilot project with 120 suppliers.
Internal reporting
ERP integration (SAP, Oracle): automatic pull of energy data, waste data, and HSE (Health, Safety, Environment) incidents. HRIS (Workday, SAP SuccessFactors): gender pay gap, training hours, diversity metrics – no manual export required.
Verification and audit
External assurance (limited/reasonable) requires an audit trail for each digit. The system stores the audit trail: data_point → source_system → raw_record_id → transformation_logic. The auditor obtains drill-down links from the report to the original counter or document.
Automated consistency checks: cross-check data between report sections (Scope 1 in the environmental section must match Scope 1 in the risk section), year-over-year variance alerts (>30% change without explanation = flag for verification).
Stack and output formats
Storage: Snowflake + dbt. LLM: GPT-4o via Azure OpenAI, Claude 3.5 Sonnet via Anthropic API. Vector store: pgvector (PostgreSQL) or Weaviate. PDF generation: WeasyPrint or Puppeteer. Output: XBRL/iXBRL for regulatory submission (ESEF format for ESRS).
Development timeline: 4–8 months for a complete pipeline from data ingestion to report generation. A basic automated data collector without LLM narratives: 2–3 months.