第2部分：技术报告探索范围3会计和过渡财务的人工智能（英）

Project SymbiosisPart 2: Technical report Exploring AI for scope 3 accountingand transition finance October 2025 Contents Overview3 1Data collection4 Data transformation5AI in data transformation7Supply chain data collection8AI in supply chain data collection9APIs and integrations10 2Impact calculation11 Generator services13AI for data classification15Modelling engine17LCA database21Impact category flexibility22 23 Reporting and analysis 4Reductions25 Overview25NEMO26Features and functionality27Reductions analyser28Generator30Matcher36Investment matcher38NEMO LLM evaluation and results39Evaluation methodology39Evaluation results – generator45Evaluation results – matcher52 Abbreviations, Acronyms, and Definitions55 Overview As set out in Part 1, the first goal of Project Symbiosis is to explore how advanceddata techniques and artificial intelligence (AI) can be leveraged to more accuratelycollect, interpret and calculate scope 3 emissions and other impact data incorporate supply chains. For the purpose of this goal, the project performedappliedtechnology researchthat explores and explains the AI techniques that could beused to achieve this. The second goal of Project Symbiosis is to explore approachesfor identifying opportunities to reduce such scope 3 emissions. The third goal is todesign a “matching engine” to match suppliers with funding sources to decarbonisethe supply chain (referred to as financeable emissions reduction opportunities). Forthe purpose of these goals, the project developed a proof of concept (POC) referredto as theNovel Emissions Optimiser or NEMO.12 Taken togetherthe applied technology research and NEMOprovide a blueprintthat seeks to addressa number of the challenges,with a specific focus on: ̵Reducing the friction and effort involved indata collectionto calculate acorporate carbon footprint (CCF). High levels of friction consistently result inlower levels of footprint accuracy and a reduced focus on decarbonisationmeasures. ̵Increasing the accuracy of emissionscalculationsby utilising a more granularset of emissions factors and using modelling techniques for targeted proxyselection of missing data points consistent with a range ofreportingregimes. ̵Helping users identify and understand thereduction potentialof impactfuldecarbonisation measures, with the ability to forecast and align projectedemissions to a target. I.Data collection Data collection is among the most challenging aspects of carbonaccounting due to fragmented supply chains and inconsistent data.Accordingly, in Project Symbiosis we explored solutions that simplifydata collection, automate the transformation of disparate dataformats, support supplier engagement and fill data gaps with logicalassumptions. While emerging AI shows promise in further automatingthe process, its effectiveness to date has been poor, warranting furtherresearch and development. Data collectioncovers the set of features used to aid users in the collection of thebroad sets of data required to calculate a CCF. Given that the accuracy of thecalculations isdirectly related to the qualityof the data collected, reducing thefrictionof collectinghigh-quality data material,has been a focus when it comes tofeature development. Data collection is frequently stated by users to be the most painful part of thecarbon accounting journey, for a few reasons: •The breadth and variety of required data means engagement with a broad set ofstakeholders, the vast majority of whom have limited knowledge about the datarequirements for a carbon calculation.•Data are often siloed, inconsistent and incomplete due to a lack of anoverarching and established data strategy and the governance to support it.•Access to data from the supply chain is very limited. The vast size and complexnature of supply chains makes individual engagement difficult if approached viatraditional means. To address these problems, features relating to a number of key requirements wereexplored. These features enable: •Customers to upload data in any format, to reduce the friction associated withmanual manipulation.•The transformation of data in varying quality and formats into usable data forthe calculation.•The extension of the reach of users, powering data collection from their supplychain. Explored features focused on product and logistics data sets due to their relativeimportance for supply chain emissions: A.Data transformation Effectively utilising customer data sets for the emissions calculations requires thatthe data are normalised and mapped to the data format required for thecalculations. In a world in which all customer data followed a consistent format andall third-party systems exported data in the same way, transformation of the datawould be a low-effort, one-time exercise. However, this is not the case. It would beunreasonable to assume that all customer data are of the same quality, althoughthird-party systems exporting data consistently for the purposes of carb