SARS AI compliance engine — South Africa's machine-learning tax risk platform

Since 2019 South Africa's Revenue Service (SARS) has used gradient-boosting and neural-network models across more than 1.2 billion annual records — from banks, retirement funds, medical insurers, the deeds registry, the Companies and Intellectual Property Commission, the vehicle register and international partners — to generate risk scores for every taxpayer and customs entry.

Risk scores flag cases for review by human revenue officers rather than triggering automatic action; by FY2023 more than 60% of audits were AI-selected. The system operates under the SARS Act with quarterly parliamentary reporting to the Standing Committee on Finance and annual reports under the Public Finance Management Act.

In FY2024/25 SARS collected R304 billion through identifiable compliance activities, a 16.7% year-on-year increase attributed largely to AI-driven risk profiling and fraud detection. A parliamentary submission by the SARS Commissioner reported that AI detection models had cumulatively prevented more than R444 billion in fraudulent outflows, roughly 32% of compliance revenue protected.

SARS has presented the model at the G20 and to the African Tax Administration Forum as a transferable approach for other African revenue authorities; an independent 2024 academic study documented the system's architecture and legal framework.