Georgia e-Treasury AI — 42 machine-learning models for anomaly detection in 7 million annual public payment transactions

Georgia's Ministry of Finance, through the LEPL Financial-Analytical Service, integrated machine-learning models into its Public Financial Management System (PFMS) to reduce payment errors and anomalies in State Treasury disbursements, processing more than 7 million annual payment transactions through the e-Treasury module (one of four integrated PFMS components). Development and validation were supported by the IMF Fiscal Affairs Department and funded by the Swiss State Secretariat for Economic Affairs (SECO).

Reduce payment errors and anomalies in Treasury disbursements using machine learning.

Two channels operate within the system: a Green Channel where 42 supervised models learn patterns from historical data to detect anomalies in payment orders automatically, and a Red Channel where a model trained on Treasury officers' past rejection decisions flags problematic orders for human review.

Testing showed the system correctly identified 6 of 8 previously known erroneous payment orders and achieved more than 80% accuracy in predicting which orders officers would reject; full operational deployment was still being finalized at the time of reporting, so these are testing rather than post-deployment audit results.

The IMF published a detailed technical account (September 2024) presenting the approach as a transferable model for Treasury Single Account consolidation in other emerging economies; the system is designed to scale to a 33% increase in transaction volume from ongoing TSA expansion.