Scope 3 emissions typically account for the largest share of an organisation's emissions portfolio, covering indirect emissions across the upstream and downstream value chain activities. These emissions have become a central concern for organisations as emerging regulatory and government disclosure standards started moving from voluntary to mandatory reporting. However, measuring and reporting these emissions is difficult because it requires collecting and harmonising data across thousands of upstream and downstream suppliers. Consequently, many organisations use spend-based methods by linking general ledger expenditure records to emission factors as a proxy. This approach presents its own challenges, as mapping financial transactions to the appropriate Scope 3 categories and emission factors are often manual, time-consuming, and labour-intensive. This study introduces a machine learning-based approach to overcome these challenges by empirically evaluating four models, including both traditional and deep learning models, for classifying financial transaction data into EEIO categories. The authors employed a real-world dataset from an Australian higher education institution. Using a real-world dataset, the study identified complexities in realistic operational datasets, including class imbalance, semantic interdependence, and high dimensionality. Selected traditional SVM and XGBoost, together with DistilRoBERTa-base and BiLSTM deep-learning models, were accordingly adapted and implemented to address these real-world dataset issues. The results indicate that both conventional and deep learning models perform well under different hyperparameters when the identified dataset complexities are taken into account. Deep learning models outperform conventional classifiers, achieving up to 95% accuracy and offering greater accuracy and reproducibility when organisations have sufficient technical capacity, budget, and other resources. By contrast, traditional models deliver commendable performance of up to 84%, underscoring their suitability for organisations with lower technical readiness and fewer resources. Hence, this study advances automated Scope 3 emissions accounting through the application of machine learning, thereby streamlining disclosure workflows and supporting organisational progress toward net-zero targets.