Skattestyrelsen Transaction Network Analysis — Denmark’s machine-learning VAT carousel-fraud detector

Denmark's Skattestyrelsen (Tax Agency) has deployed machine-learning systems for taxpayer risk assessment since at least 2014, spanning VAT, personal income, corporate income tax and customs. Its most documented AI tool, Transaction Network Analysis (TNA), is a real-time data-matching and network-graph engine that flags suspicious VAT refund claims by detecting patterns typical of 'carousel fraud' — chains of fictitious intra-EU transactions used to claim fraudulent refunds.

The system uses supervised learning to segment taxpayers by compliance risk and assign audit-priority scores. It has operated continuously since 2014 and has been updated and integrated into the EU-level TNA platform, enabling cross-border fraud detection across Member States.

Skattestyrelsen's targeted compliance campaigns for money transfers report uncovering errors or fraud in approximately 8 of 10 checks, and the taxadmin.ai ISORA country report records around 70% accuracy for the TNA system.

Governance gaps are significant: the use of ML by the Danish Tax Administration is not governed by any specific ad hoc norms (only the property-valuation system falls under a dedicated statute), training-data specifications are not publicly disclosed, and no independent audit of the TNA system or its accuracy claims has been published — a contrast with analogous systems in Ireland (Revenue REAP) and Austria (PACC). The earlier EFI automated-collection system (2005-2015) was shelved after operational failures, a cautionary precedent on over-automation, even as EU TNA integration demonstrates genuine cross-border transferability.