Tax and customs administrations in emerging digital economies face simultaneous pressure to accelerate trade clearance, protect revenue, and maintain procedural fairness under rapidly digitising commercial activity. Traditional rule-based targeting, while operationally important, is increasingly outpaced by cross-border e-commerce growth, valuation manipulation, synthetic invoicing, shell entity proliferation, and adaptive fraud behaviour. This review synthesises peer-reviewed literature and high-value institutional publications from 2020 to early 2026 on artificial intelligence (AI), machine learning (ML), graph analytics, anomaly detection, explainable AI (XAI), and human-in-the-loop (HITL) risk management applied to tax and customs declarations. Following PRISMA 2020 reporting principles, 38 principal sources were selected from 518 initially identified records across Scopus, Web of Science, IEEE Xplore, ACM Digital Library, SSRN, Google Scholar, and institutional repositories of the OECD, WCO, APEC, CIAT, UNDP, and the International Growth Centre. The synthesis reveals three structural findings. First, high-performing anomaly detection systems consistently employ hybrid architectures that integrate supervised fraud scoring, unsupervised or semi-supervised anomaly detection, graph-based relational modelling, and active learning — rather than any single model family. Second, performance claims are highly sensitive to label scarcity, inspection bias, concept drift, and operational constraints including audit budgets, explainability mandates, and workflow integration requirements. Third, governance, legal explainability, human oversight, and continuous model monitoring are necessary design requirements — not post-implementation considerations — for public-sector deployment. Based on this evidence, the paper proposes a seven-layer AI-driven anomaly detection framework specifically designed for revenue administrations in emerging digital economies. The framework integrates data identity and interoperability foundations, risk-feature engineering, hybrid model stacking, explainable risk presentation, human-in-the-loop case management, continuous learning and drift control, and institutional governance and recourse mechanisms. The framework is directly applicable to ZATCA (Zakat, Tax and Customs Authority) in Saudi Arabia, GCC customs administrations, and comparable digital revenue administrations across emerging economies. The review concludes that technically credible AI-enabled anomaly detection is now achievable, but that scaled and sustainable deployment depends less on model novelty than on data quality, interoperable digital infrastructure, inspection feedback loops, legal explainability standards, and governance maturity.
Obaid Ur Rehman Qureshi (Fri,) studied this question.