Business Context: Why ESG Reporting Needs Automation
A typical SME preparing its first GHG inventory spends 3-4 months manually collecting energy and activity data, mapping it to emission factors, calculating totals, and writing the narrative disclosure. This is not a knowledge problem. The methodology (GHG Protocol Corporate Standard) is publicly available and well documented. It is a data extraction and computation problem. Most of the required data already exists in supplier invoices, utility bills, fuel receipts, and logistics records. The problem is getting it out of those documents and into a consistent format. For larger companies preparing CSRD-compliant reports, the data collection challenge scales with the number of suppliers and facilities. Third-party supply chain emissions (Scope 3) require data from hundreds of suppliers, which creates a verification and aggregation challenge that cannot be solved with spreadsheets. AI changes the feasibility of both problems. Document extraction from invoices and utility bills is a well-suited AI task. Narrative generation for standard disclosure templates is where LLMs provide the most immediate time saving. TCFD-aligned language requirements and GHG Protocol categorisation rules can be encoded into the system prompt.
Architecture: Document Ingestion, Emission Factor Mapping, and Disclosure Generation
The system has three functional layers. The ingestion layer accepts uploads of utility bills, fuel receipts, supplier invoices, logistics reports, and travel expense records in PDF or CSV format. Claude extracts structured activity data from each document: energy consumption in kWh, fuel consumption in litres, distance travelled in kilometres, waste weight in tonnes. Each extracted activity record is stored in PostgreSQL with the source document reference, extraction confidence, and data category (Scope 1, Scope 2, or Scope 3 category). The calculation layer maps each activity record to the appropriate BEIS emission factors (for UK operations) or IPCC emission factors (for international activities), applies the correct conversion factor, and aggregates totals by scope and category. The disclosure generation layer accepts the aggregated emissions data and produces narrative disclosures. Claude generates TCFD-aligned climate narrative sections (governance, strategy, risk management, metrics and targets) and GHG Protocol-aligned scope summaries. The output is formatted as a structured report template that companies can edit and publish. All calculation steps are logged with the source data and emission factor references, providing the audit trail required for third-party assurance.
AI Components: Extraction, Classification, and Narrative Writing
Claude handles three distinct tasks. First, activity data extraction: given a utility bill or invoice, identify the relevant consumption figure, the unit, the time period, and the activity category. This is structurally similar to financial document extraction but with the specific taxonomy of GHG Protocol categories. Second, Scope 3 category classification: given a supplier invoice or logistics record, classify the activity into the appropriate Scope 3 category (purchased goods and services, business travel, upstream transport, waste, etc.) using the GHG Protocol Scope 3 Standard category definitions. Third, disclosure narrative generation: given the aggregated emissions data and a disclosure framework template (TCFD, GRI 305, or CSRD), generate the narrative sections in compliant language. The narrative generation task is where the time saving is most visible. Writing a TCFD-aligned strategy section that correctly describes physical and transition climate risks, the company's scenario analysis approach, and its emissions reduction targets takes an experienced sustainability consultant hours. Claude produces a first draft in seconds that the consultant then reviews and adjusts.
Challenges: Data Quality, Scope 3 Completeness, and Assurance Readiness
ESG reporting is only as good as the underlying data, and most companies discover during their first GHG inventory that their data quality is significantly worse than they expected. Invoices are missing, utility bills are aggregated across facilities, fuel purchases are recorded without vehicle assignments. The AI system cannot fix missing data, but it can identify where data gaps exist and quantify their materiality to the overall inventory. SpeedMVPs builds a data completeness dashboard that shows coverage by Scope, category, and reporting period, so companies know exactly what they are reporting on and what is excluded. Scope 3 completeness is the hardest problem. Getting primary data from every supplier is impractical. The standard approach is to use spend-based estimation for suppliers where primary data is unavailable. The system supports both primary data entry and spend-based estimation, with clear disclosure of which method was used for each category. Third-party assurance (required for TCFD disclosures by many large companies and increasingly expected for CSRD) requires that the calculation methodology, emission factors used, and source data are all traceable and reproducible. The calculation log in PostgreSQL provides this traceability.
Outcomes: Reporting Time and Quality Improvements
Companies that pilot AI ESG reporting tools consistently report the same experience: the time from data collection to draft report drops from months to weeks. More specifically, the document extraction step, which is the most labour-intensive part of the process for non-accountants, becomes a matter of uploading files and reviewing the AI-extracted data rather than manually transcribing figures. Sustainability teams report that the AI-generated narrative drafts save 3-5 hours per disclosure section, with the consultant's time shifting from writing to reviewing and adjusting rather than composing from scratch. For finance teams preparing UK TCFD disclosures, the structured calculation log with emission factor references significantly reduces the time spent on evidence gathering for external assurance.
Lessons: Start with Scope 1 and 2, Add Scope 3 Iteratively
Scope 1 and 2 emissions are typically well-suited to AI-assisted calculation because the data sources are relatively consistent (energy bills, fuel receipts) and the emission factors are clearly defined (BEIS factors published annually). Scope 3 is significantly more complex: 15 categories, each with different data source requirements, estimation methodologies, and relevance thresholds. Attempting to build a complete Scope 3 system in an MVP is a reliable way to produce something that does not work well for any category. Start with the Scope 3 categories most material to your target user's industry: for a manufacturing company, purchased goods and services and upstream transport are usually the largest categories. For a professional services firm, business travel and employee commuting are most relevant. Deliver those well, then add categories iteratively based on user demand.