State of use of AI tax systems
According to incidental reports from IOTA, the Estonian tax administration (ECTB) has been experimenting with fraud analytics and machine-learning since 2017. Based on publicly available sources, the Estonian tax administration uses at least two different models.
One machine-learning algorithm, developed by MindTitan is a risk detection tool designed to detect under-reported income and so-called ‘under the table wages’ which replaces the former rule-based algorithms with a machine-learning algorithm which, unlike the former system, would update itself based on historical taxpayer data but also the subsequent inputs of tax officials auditing companies.
The second machine-learning algorithm is ‘Tax Behaviour Rating’ (TBR): TBR analyses data provided by a taxpayer and compares it to information submitted by other taxpayers as well as information obtained from public registers. The results of the analyses are displayed to taxpayers according to a colour code (green to red), which provides taxpayers with an idea of their level of risks/deficiencies compared to other taxpayers. The use of TBR is not mandatory for taxpayers but functions on a voluntary basis to ultimately provide them with two ratings: 1. tax compliance rating: provides information on whether a company fulfils its tax obligations in a timely manner; wages declared by the company are compared with the average Estonian wages in similar positions. 2. tax behaviour adequacy rating: draws attention to potential inconsistencies in the declaration of taxpayer data to the ECTB.
What functions are performed with AI?
Machine-learning algorithms leveraged by the ECTB carry out two functions:
- Risk detection: the model analyses tax returns to identify under-reported income, the data mining technique used is not specified in publicly available sources.
- External risk-management (risk-scoring): the machine-learning model segments taxpayers into categories of risks, based on historical taxpayer data and information in public registers, to ultimately provide legal persons with a score which provides them with an indication of their individual likelihood to be audited.
What data can be processed by these systems?
The data used in the first model is not specified by the ECTB.
TBR is reportedly based on six data categories: unpaid taxes, tax declarations, tax offences committed by the company, tax proceedings, background of managing directors and average wages. TBR uses general data (name, legal address, principal activity, etc.) data declared to the ETCB (turnover, paid taxes, and number of employees), information about arrears and declarations not submitted, business licenses, licenses for warehouses, and business life cycle.
Are these systems regulated by specific norms?
There is no specific legal basis for the use of AI, yet §11(2), §59 and following of the Estonian Tax Act provides a general legal basis for the use of risk analyses.
References:
Legislation :
- Estonian Taxation Act1 Passed 20.02.2002, RT I 2002, 26, 150, available at: https://www.riigiteataja.ee/en/eli/ee/523012015008/consolide/current;
Doctrine:
- A. Soom for Kluwer Tax Blog (2020), see: http://kluwertaxblog.com/2020/11/05/could-estonian-tax-behavior-rating-facilitate-abusive-tax-practice/ - last accessed June 2022.
- MindTitan in Digital Diplomacy (May 2020), available at: https://medium.com/digital-diplomacy/mindtitan-develops-an-artificial-intelligence-model-for-the-estonian-tax-and-customs-board-to-24289df87ae2 - last accessed June 2022; see also: https://mindtitan.com/ – where the ECTB is listed as a trusted customer - last accessed June 2022.
- On TBR, see: Republic of Estonia Tax & Customs Authorities, available at: https://www.emta.ee/eng/tax-behaviour-ratings - last accessed June 2022.