BgGPT at NRA — Bulgaria's Sovereign AI for Tax Administration

Bulgaria's National Revenue Agency (NRA) deployed BgGPT, a sovereign open-weight large language model developed by INSAIT (Institute for Computer Science, Artificial Intelligence and Technology) at Sofia University, in March 2026 - described as the first large-scale public administrative deployment of a national sovereign LLM in the country. BgGPT is based on Google's Gemma 3 open-weight model series and runs on NRA's own hardware rather than a cloud provider, to avoid data leakage. Separately, the NRA has operated an ML-based tax-risk scoring system since at least 2016.

BgGPT lets NRA staff search tax and social-security legislation and internal administrative documents with real-time responses; the separate ML risk-scoring system aims to prioritise audit resources by classifying tax returns as high- or low-risk.

Following a successful internal pilot, the NRA expanded BgGPT to its central office and all seven territorial directorates, with national integrator Information Services handling technical rollout; the deployment targets daily active use by more than 7,000 NRA employees.

Academic work published in 2024 documents the NRA's ML risk-scoring system auditing approximately 15,000 tax returns per month using algorithmic high-risk/low-risk classification. As of March 2026, BgGPT was still in expanded rollout across the agency, with the 7,000+ daily-user figure stated as a deployment target rather than a confirmed usage level.

The source material explicitly notes that no staff-productivity or error-reduction metrics for BgGPT, and no published accuracy or revenue-impact metrics for the ML fraud-scoring system, are available.