State of use of AI tax systems

The first publicly documented use of tax machine-learning algorithms by the tax administration of the Kingdom of Spain (AEAT) dates back to as early as 2012, according to official sources from IOTA.


What functions are performed with AI?

Based on publicly available data, tax machine-learning algorithms perform at least three types of functions for the AEAT:

  1. Social Network Analysis (SNA) ‘TESEO’: the SNA algorithm visually represents a network of individual taxpayers using graph theory, it represents a network of taxpayers as a combination of nodes for individuals or points of interests and lines which quantitatively and qualitatively measure relations between the nodes. TESEO is then used to identify and represent relationships between individual taxpayers in near real-time.

  2. Risk-detection 'INFONOR': INFONOR automatically detect and flags suspicious data, transactions and relevant results in real-time, without any active participation from tax inspectors. 

  3. Risk-detection 'DEDALO': DEDALO is used to identify and locate taxpayers, for which no precise information is available. Taxpayers are identified by DEDALO through other search parameters than the NIF-number or the name, such as the real estate, bank account, vehicles, or even disparate information in zujares.

  4. Risk-detection 'ZUJAR': ZÚJAR is a computer tool, which allows processing of information in the databases for the selection of taxpayers. It enables the filtering of taxpayers by predefined variables through Boolean algebra. The ZÚJAR program, in turn, contains zújares, i.e. units of ordered information to develop tax management, inspection and collection actions. ZÚJAR classifies and divides data according to different concepts and filters thousands of variables and millions of records. Therefore, it can draft lists of taxpayers or specific attributes for risk-scoring. ​

  5. Risk-detection 'GENIO': According to IOTA, GENIO is a supplementary tool that allows the issuing of standardized reports as a conclusion of the risk analysis performed. 

  6. Risk-detection 'PROMETEO': PROMETEO issues a detailed report of tax digital documentation, such as accounting documents, VAT book and bank accounts. PROMETEO also acts as a data-matching tool, enabling the treatment of accounting and computer records obtained from taxpayers and the comparison of these documents with information already present in the data warehouse of the tax administration. 

  7. External risk-management (risk-scoring) 'HERMES': According to IOTA HERMES is a risk and profile system for the analysis of taxpayers that defines profiles (risks groups) to allow adapting the resources to the defined profiles. The system makes also use of the already existing ZUJAR infrastructure. This risk analysis method supposedly enables immediacy and flexibility in the process.

  8. External risk-management (risk-scoring) ‘Anti-Fraud Tool HLF’: The HLF tool predicts risks of fraud or non-compliance associated with individual taxpayers for social security or wage-related fraud, to subsequently rank and (pre-)select taxpayers for further audits by officials of the State Labour and Social Security Inspectorate of the Kingdom of Spain. This tool is particularly used to combat undeclared work and bogus employment.

  9. Taxpayer assistance VCA chatbot ‘AVIVA’: AVIVA is chatbot designed to automatically answer taxpayer queries and FAQs of legal persons regarding SII (sunministro inmediato of informacion), corporate income taxation, VAT and e-invoicing. According to statements of Spanish tax officials, the tool was updated during the SARS-COV-2 pandemic to include FAQs asked as a result of measures taken to combat the spread of the virus.


What data can be processed by these systems?

The data collected and processed for TESEO is not specified, but reportedly includes individual and companies’ bank accounts, cadastral property, foreign assets and a wide range of additional data. Spanish tax officials reported that so far, TESEO has measures more than 49 different types of relations among more than 530 million connectivity arcs measured. It does not confine itself to measuring relations between taxpayers, but also includes the relations of minors and other taxpayers.

The data collected and processed for the development and use of the ‘anti-fraud HLF tool’ is not specified.

The data collected and processed for the development and use of AVIVA is not specified, but reportedly includes frequently asked questions to the AEAT and the answers provided to taxpayers by tax officials.


Are these systems regulated by specific norms?

As of February 2022, the use of these tax machine-learning algorithms is not regulated by specific legal norms.



References: