State of use of AI tax systems

According to official reports from IOTA, the Hungarian tax administration (NCTA) has been using machine-learning algorithms since 1st of July 2016 onwards. In particular to detect risks of non-compliance and fraud, and to perform centralised algorithmic-based segmentation and selection of taxpayers for audit.

What functions are performed with AI?

According to C. Tamas Czinege, Director General for Taxation at the NCTA, the NCTA leverages machine-learning algorithms to perform three functions:

1. Risk detection: These types of algorithms identify risks within tax returns and tax documentation submitted to the NCTA, based on risk-factors derived from taxpayer historical data and existing data sets available in the data warehouse of the Hungarian government.

2. External risk-management (risk-scoring): Posterior to the analysis and detection of risks of tax non-compliance or fraud, algorithms scores and segments taxpayers into risk categories to perform centralised algorithmic-based selection of taxpayers for audit.

These two types of algorithms have been reportedly developed for a wide range of taxation and supportive processes, including VAT, customs, corporate and personal taxation as well as the automatic exchange of information. According to the IOTA report the following risk-factors can primarily be evaluated: risk and tax audits aspects included in the yearly tax audit areas; taxpayers’ performance established according to centrally unified approaches; qualification of taxpayers; experiences of tax audits and supportive procedures carried out earlier, information on serious infringements of taxation rules; correlations amongst tax returns and data disclosures; information disclosed by third parties (other authorities and other taxpayers); data disclosures; databases; data from international data exchange system; taxpayers’ taxation history; irregularities and contradictions emerged earlier at owners, representatives business partners and companies founded by them; publicly available data or other information that is available in databases of NCTA.

3. The IOTA report also indicate that the NCTA experiments with nudging tools, to adapt the language of standard communication to taxpayers in order to nudge them towards compliance.

What data can be processed by these systems?

The data which can be used for these algorithms is not specified, but reportedly uses a wide range of data derived from existing datasets of the Hungarian government but also data from VAT Information Exchange System (VIES), Common Reporting Standards (CRS) and Foreign Account Tax Compliance Act (FATCA), DAC 2, DAC6, data from the onlice invoice system and data from the Hungarian Electronic Public Road Trade System (EKAER).

Are these systems regulated by specific norms?

These models are not regulated by specific legal norms, but have been implemented on the basis of managerial decision of the Hungarian tax administration.