evidoria

← Back to browse

Good practice Imported

Poreska Uprava Big Data CRM — Serbia's AI Compliance Risk Management Platform

Serbia · Belgrade · See the Serbia profile · See the Belgrade profile

Evidence: Observational / pre–post Top 50% 60/100 · Ask Evidence Copilot about this practice

Since 2018 Serbia's Tax Administration has run an AI/big data compliance risk management system with University of Novi Sad; salary-tax revenue grew ~70% nominally (2018–2022), attributed jointly to the CRM platform and broader administrative reforms.

~70 % nominal growth
Salary-tax revenue growth (2018–2022)
~50 % nominal growth
Total tax revenue growth (2018–2022)
90–98 %
Outlier-detection model accuracy (academic test datasets) (as reported in University of Novi Sad paper)
Poreska Uprava Big Data CRM — Serbia's AI Compliance Risk Management Platform

Details

Maturity
Established
Promoter
Poreska uprava (Tax Administration of Serbia) / University of Novi Sad
Period
2018–present
Keywords
tax administration, compliance, big data, fiscal

Context

Serbia's Poreska Uprava (Tax Administration) launched a compliance risk management (CRM) initiative in 2018 in partnership with the Faculty of Science at the University of Novi Sad, applying big data analytics and machine learning to salary-tax declarations.

Objectives

The system aims to detect deviations in payroll reporting, score employer compliance risk, and prioritise limited inspection resources toward likely non-compliant employers.

Activities

The CRM processes administrative tax records using machine-learning methods; a University of Novi Sad academic paper documents use of logistic regression, random forest and support vector machines. In January 2022, Serbia introduced eFiskalizacija, mandatory real-time electronic fiscal receipts for all consumer-facing transactions, extending the data pool available to the risk-scoring models.

Results

Between 2018 and 2022, Serbia's salary-tax revenue grew approximately 70% in nominal terms, with total tax revenue up around 50% over the same period; a World Bank blog (January 2024) attributed these gains partly to the CRM platform alongside macroeconomic growth and administrative reforms. The academic paper reported 90–98% accuracy on tested outlier-detection datasets.

Conclusions

The revenue growth cannot be attributed exclusively to the AI system — macroeconomic factors and broader reforms were co-drivers — and the Risk Management Unit remained understaffed as of 2022 per IMF Serbia country reports; algorithmic logic is not publicly disclosed.

Implementation
Indicative cost
Medium (€50k–€500k) — Academic-government partnership (University of Novi Sad) plus national eFiskalizacija infrastructure; no public cost figures found in source.
Time to results
Long (> 3 years) — Operating continuously since 2018, with eFiskalizacija data expansion from January 2022 and World Bank documentation through January 2024.
Staffing & skills
Poreska Uprava (Tax Administration of Serbia) Risk Management Unit, Faculty of Science, University of Novi Sad, as academic/technical partner

Conditions for success

  • Academic-government partnership providing technical machine-learning expertise
  • eFiskalizacija (2022) mandatory e-receipts expanding the data pool for risk scoring
  • World Bank engagement supporting the Risk Management Unit

Common failure modes

  • Risk Management Unit remained understaffed as of 2022 per IMF Serbia country reports
  • Algorithmic logic is not publicly disclosed
  • Revenue growth is confounded with macroeconomic growth and other reforms, not isolated to the AI system

Where it fits

Governance type
national tax authority with academic partnership
Scale
national (100,000+ VAT-registered businesses under eFiskalizacija)
Income level
upper-middle-income (Serbia)

Commonly funded by

National / regional programmes

Indicative funding routes for practices of this type — always check each programme's current calls and eligibility rules.

Do you run this practice? Claim it — verified implementers get a public contact pathway and can propose corrections.

Data sources

Where this practice's information was retrieved from, and when.

Attachments

Similar practices you may find useful