Algorithms in the service of the tax office
At the end of 2017, we wrote about the possibility of using artificial intelligence in the financial services sector. We pointed out that AI algorithms can be used by the financial industry to automate customer contacts and issue initial credit decisions. The use of algorithms by government bodies seemed to be less important at the time. However, this issue ignited much controversy at the end of 2018 due to a ruling by the Province Administrative Court in Warsaw on the freezing of a bank account under a recently introduced section of the Tax Ordinance, which also introduced the digital clearinghouse STIR into the Polish legal system.
What is STIR, and does an algorithm freeze a bank account?
These changes to the Tax Ordinance are intended to combat VAT fraud. STIR is a system enabling the exchange of information between banks and the National Revenue Administration, among other entities. A “risk indicator” is established in STIR (i.e. an indicator of the use of banks’ and credit unions’ services for purposes related to fiscal fraud).
A risk indicator is assigned to “qualified entities” (basically these are businesses, both natural persons conducting business activity and companies) on the basis of a developed algorithm, which, in the words of the Tax Ordinance, takes into account “best practices of the banking sector and credit unions for preventing the use of their operations for fiscal and other offences.” The algorithm also takes into account economic, geographical, physical, behavioural and affiliation criteria (i.e. links between an entity and other entities at risk of participating in activities related to fiscal fraud).
The best practices mentioned above may also be submitted by banks, among other entities. The risk indicator and the criteria for establishing it for a given entity are not disclosed to that entity.
The determination of the risk indicator is crucial, as it seems it is the main grounds for the head of the National Revenue Administration to request a freeze of the qualified entity’s bank account. The act does not provide for an automatic freeze of an account if a certain high level of risk is established for the given entity, and from 30 April 2018 (when the regulation entered into force) through mid-December 2018 only 40 freezes were executed under this procedure. Additionally, the Ministry of Finance claims that a freeze will occur only in exceptional cases. However, it cannot be ruled out that a practice will develop of routinely requesting a freeze in every case where the risk indicator is set at a high enough level.
Importantly, details on the algorithm for determining the risk indicator and individual freezes are not available to the public, so it is difficult to verify how the algorithm works, its reliability, and the real possibility for a human being to dispute the findings of the algorithm. It is possible that before requesting a freeze, decision-makers thoroughly verify the findings of the algorithm and reject erroneous indications, but the opposite situation could also take place. In an extreme scenario, the decision to freeze an account would de facto be taken by the algorithm, and the human role would be reduced to making the actual request for a freeze. For this reason, the practice applied by the tax authorities is the key.
Ruling of the administrative court and its practical consequences
The head of the National Revenue Administration may extend an account freeze described above for a fixed period not longer than three months. According to the judgment of the Province Administrative Court in Warsaw from September 2018 (Case III SA/Wa 2057/18), in which the court dismissed a complaint against the extension of the time limit to freeze a bank account, no strict evidentiary proceedings are conducted in the proceedings for freezing an account; it is sufficient that an analysis of cash flows shows a risk of use of a financial institution for tax fraud.
If the administration relies mainly on the algorithm to determine the risk indicator, in extreme situations this could lead to a freeze of an account of a qualified entity for a period of three months on the basis of an AI “decision.” A freeze on its bank account could have dire consequences for a business and in an extreme case could even drive it into insolvency. Therefore, it is crucial that, first, the risk indicator be established in the STIR system in a transparent manner, to ensure accountability for its indications, and second, there be a realistic opportunity for human verification of the algorithm’s findings. This could be achieved by describing how the algorithm functions so that anyone can understand it.
For the time being, it seems that account freezes are being used prudently, but it is important to observe the trend in the algorithm’s operation as more and more data is fed into it, as well as the related practice of the tax authorities.
Katarzyna Szczudlik
