Abstract
- Enterprises routinely lose between three and eight percent of addressable procurement spend to fragmentation, duplicate payments, maverick buying, contract price leakage, and unmanaged tail spend. This paper presents ProcureAI, a full-stack, AI-driven spend-intelligence platform that ingests multi-source procurement data, classifies it automatically using a hybrid NLP classifier combining Multinomial Naive Bayes with TF-IDF cosine k-NN re-ranking, detects leakage through a six-detector unsupervised anomaly ensemble, segments the supplier base with K-Means++ clustering augmented by Herfindahl-Hirschman Index concentration analysis, forecasts future expenditure via a four-model inverse-MAPE-weighted time-series ensemble with residual-quantile prediction intervals, and synthesises all signals into a ranked, costed, auditable cost-reduction portfolio. Applied to a 36-month synthetic enterprise dataset of ~8,000 transactions and $85 million in total spend, the platform achieves ~86% NLP classification accuracy (macro-F1 ~87%), detects ~4,000 anomaly events, achieves a forecasting ensemble MAPE of ~18%, and identifies $7.3 million (8.4%) in annual run-rate savings. All algorithms are implemented from first principles in TypeScript, ensuring full auditability without opaque model binaries or external ML services.