The United Nations Standard Products and Services Code (UNSPSC) is a four-tier hierarchical taxonomy comprising segments, families, classes and commodities, designed to facilitate globally consistent categorization of goods and services. The Queensland Government quarterly processes over seven million new procurement transactions, all of which, in line with recommendations from the Queensland Audit Office, are mapped to the UNSPSC hierarchy to support timely decision-making and improve operational efficiency. In this paper, we introduce the Categorization Architecture with Predictive Reasoning and Alignment (CAPRA), a framework for rigorously evaluating real-world labeling pipelines. CAPRA systematically compares multiple candidate configurations across four key dimensions: accuracy, model size, inference speed and API cost. To maintain data quality without overburdening human experts, who cannot feasibly validate millions of transactions each quarter, we embed an AI-driven feedback mechanism and a retrieval-based review process that selectively flags only the lowest-confidence predictions for expert adjudication. This “selective confidence-based review” closes the data-service loop by focusing domain expertise where it matters most. Finally, we demonstrate CAPRA’s effectiveness through extensive experiments on the Queensland Government expenditure. Our results show that CAPRA outperforms a fine-tuned LLM by 19% in prediction accuracy while requiring only 10% of its inference time, enabling seamless deployment in a production environment.