State of use of AI tax systems
In Germany, tax assessment procedures have been fully automated with machine-learning algorithms since 2017 with the introduction of the Taxation Modernisation Act (‘Gesetz zur Modernisierung des Besteuerungsverfarhens’), which introduced several amendments to the German fiscal Code (‘Abgabenordnung’). Since the adoption of this act, the German fiscal code foresees that tax authorities may fully automate the administrative acts in relation to tax assessments procedures, including the issuance, correction or withdrawal of tax benefits.
What functions are performed with AI?
Tax machine-learning algorithms perform four different functions:
- External risk-management (risk-scoring): The machine-learning model predicts fiscal risks in tax returns and tax documentation to subsequently segment taxpayers into categories of risks for algorithmic-based selection of taxpayers to audit.
- Automated verification of tax returns (ADM): The tool automatically (without human intervention) processes tax returns to conduct, correct, withdraw, revoke, cancel or amend tax returns and credits of withheld taxes and prepayments. The system foresees the possibility to complete a free-text field to provide additional information to be checked by a human tax official.
- Taxpayer assistance: Virtual conversational assistant (VCA) ‘Steuerchatbot’ which automatically answer taxpayer queries in a range of relevant tax domains for taxpayers in the region of Baden-Württemberg. The questions posed and answers provided by the VCA are processed by IBM Deutschland for further updates.
- Nudging: according to incidental sources from IOTA members, the German tax administration has also been experimenting with nudging tools, which adapt the language of standard communication to taxpayers based on analysis of historical taxpayer data and individual profiling.
What data can be processed by these systems?
Section 29 AO provides that personal data collected by the German tax administration is only processed with regard to a specific taxation procedure. Personal data collected for another procedure or another reason must be explicitly permitted under Sec. 29(1)(4) AO. However, one of the permitted grounds for processing under said sub-section is the development and reviewing of risk models and automated tax procedures. Hence, it can be inferred that legally the tax administration may use any data that the Federal Tax Office has at its disposal.
Are these systems regulated by specific norms?
The risk-scoring and the ADM algorithms are regulated by Sec. 88(5) of the Abgabenordnung.
The VCA and the nudging tools are not regulated by specific norms.
- Gesetz zur Modernisierung des Besteuerungsverfarhens, BGBI I, p. 1679, 18 July 2016;
- Abgabenordnung, Sec. 29, 88(5), 155(4).
- D. Hadwick & S. Lan, ‘Lessons to be learned from the Dutch Childcare Allowance Scandal: A Comparative Review of Algorithmic Governance by Tax Administrations in the Netherlands, France and Germany’, World Tax Journal, Vol. 13, Issue 4 (2021);
- N. Braun Binder, ‘Artificial Intelligence and Taxation: Risk Management in Fully Automated Taxation Procedures’ in T. Wischmeyer & T. Rademacher (eds.) Regulating Artificial Intelligence (Springer, 2019), pp. 295-306;
- IOTA Papers, Nudging in the Context of Taxation (February 2019), p. 7;
- Regarding the VCA Steuerchatbot, see: https://steuerchatbot.digital-bw.de/ - last accessed June 2022.