The Bank for International Settlements Innovation Hub has released Part 2 of its Project Symbiosis technical report, setting out applied research and a proof of concept called the Novel Emissions Optimiser (NEMO) to improve how firms collect and calculate scope 3 emissions in supply chains, identify emissions-reduction measures, and connect suppliers with potential funding sources for decarbonisation. The report describes an end-to-end architecture covering data collection, impact calculation, and reporting. It proposes reducing data-collection friction through flexible ingestion (CSV, XLSX and JSON uploads, with SFTP for files over 500 MB), a transformation pipeline into a cloud SQL store, and supply-chain data collection workflows; AI testing showed poor performance for an agent designed to automate complex data mappings, but materially better results for parsing standardised transaction certificates using GPT-4o. For calculation, it outlines generator services (product, logistics, facilities, and business operations), hierarchical product classification using OpenAI ada-002 embeddings and support vector machines, and a graph-based modelling engine (Neo4j) that uses probabilistic inference to fill data gaps and support multiple impact categories, including the 16 product environmental footprint categories. Reporting is designed to support structured analysis and exports aligned to common regulatory and standards-driven metrics, including greenhouse gas emissions by scope and GHG Protocol category and proportions of primary and extrapolated data. NEMO is presented as a reductions and transition-finance layer with two AI approaches: a generator path that creates measures and recalculates baselines via a third-party modelling engine to address methodological mismatches, and a matcher path that selects measures from an expert-curated database and applies pre-computed relative impacts to a user-provided baseline. Testing on a 200-product labelled set indicated that prompt iteration and reasoning models improved generator performance, with OpenAI o3 selected for measure generation, while the matcher achieved high correctness (up to 99%) but remained well below the targeted completeness level. The investment matcher element is described as an emulation of matching measures to financing opportunities, with future development envisaging API-based integration with green finance providers.

Original source View in tracker