Artificial Intelligence

Techno-efficiency helps the IRS, but must guarantee rights

The Ai Act does not consider anti-avoidance automation 'dangerous'. Predictive tools at risk of discrimination should be eliminated

by Valeria Mastroiacovo

8 August 2025

6' min read

Key points

OECD analysis
European rules
The Italian situation

6' min read

Last monththe OECD released a report on "Tax Administration Digitisation and Digital Transformation Initiatives" that summarises inventory data on the technology initiatives in taxation from the 54 jurisdictions (OECD and non-OECD economies) that are members of the Tax Administration Forum. The data were collected through a global survey on digitisation launched in April 2024. It is information that helps imagine the future of tax systems globally. It is clear that those tax administrations, which, to date, have already carried out a digital transformation for the acquisition of functional data for the implementation of taxes, are proceeding apace in the interconnection of databases and in the use of artificial intelligence to improve the efficiency of office operations and, at the same time, services to taxpayers.

The OECD analysis

The starting point is a previous OECD report from 2020 (Tax Administration 3.0) that had called for a digital transformation of administrations to implement, where possible, easily accessible and fast taxation procedures that tend to be burden-free (through the automatic acquisition of data from intermediaries or directly from third parties). A vision that found its strong point in a progressive integration of compliance in the so-called 'natural systems', i.e. those systems that taxpayers already use (e.g. those for carrying out transactions or business activities in daily life, even through personal devices), eliminating data transfer procedures for tax purposes.

In the past five years, it is undeniable that the systems and technologies used by citizens, businesses and authorities have become increasingly interconnected, mainly due to the development of application programming interfaces, as a set of rules and protocols that enable different software to communicate and interact with each other (e.g. for record keeping or the creation and conversion of electronic files and forms) and to dialogue with those of tax administrations. This technology for the implementation of data, channelled in real time to each individual's digital identity and the multiplication of contact points between the taxpayer and the administration in the case of online services (provided directly or by other institutions or companies), is today being further developed and constitutes the decisive part of the process of transformation of tax implementation forms that has already led in many countries to the pre-filled tax return, to electronic invoicing and that will continue towards further landfalls also in the business landscape. It is basically the automatic acquisition of a significant amount of qualified data: data from which other data can be extracted or processed together with others. Data that constitute the fuel for the efficient performance of the algorithms of artificial intelligence systems.

The OECD 2025 report tells us that72% of tax administrations said they use artificial intelligence: the most common use is to detect tax evasion and fraud, others include using it as part ofrisk assessment processes (64%), asvirtual assistants (59%), as support to officials in making administrative decisions (44%) and in making recommendations for action to be taken (41%). Less common uses include ensuring the integrity of IRS systems/processes (15%), automated response of customised information to stakeholders (13%) and dispute resolution (3%). Reassuringly, none of the administrations stated that they use Ia to make fully automated final decisions, as for the rest, the belief emerges that - given the potential of technology - increasing automation of services and enhancing the use of Ia reduces administrative costs and increases the efficiency of the system for the taxpayer, the business, and the IRS.

Yet it is precisely when technological efficiency - made up of simplification of procedures and immediacy of automatisms - begins to resonate like the song of the Sirens, that the legislator must bind himself to the mast of the constitutional principles enshrined in the relevant legal system and keep his bow firmly set so that the economic and financial interests of the State do not override the rights, freedoms and guarantees of taxpayers. The risk profiles arising from the implementation of technology are manifold, from security to privacy, to equality (both in the procedure and in the identification of the taxable premise) and others still if one extends one's gaze to the procedural sphere: it is by no means taken for granted that all countries intend to deal with them in the same way.

European Rules

If with the launch of the Ai Act (EU Regulation 2024/1689 of 13 June 2024) Europe has counted on the Brussels effect for a global diffusion of the protections agreed upon therein, the current international situation presents us with a still unpredictable and in any case not very reassuring scenario. In addition to this, point 59 of the recitals of the EU Regulation (stating that 'Ia systems specifically intended for use in administrative proceedings by tax and customs authorities should not be classified as high-risk Ia systems used by law enforcement authorities for the prevention, detection, investigation and prosecution of criminal offences') seems to depoliticise the protections in these areas, emphasising the functionality of Ia for the proper enforcement of taxes.

These are arguments similar to those already expounded for the derogable profiles of the regulation of the Gdpr (with the only limitation being respect for the principle of proportionality with regard to the possibility of imposing obligations and restrictions on the rights and freedoms of private individuals enshrined in the European discipline that are not strictly necessary for the achievement of the goals set by the Member State). This, however, not without contradiction since Article 5 of the Regulation prohibits the use of Ia systems that allow public or private actors to attribute a social score to individuals, which may lead to discriminatory results and the Guidelines of 5 February 2025, among the examples of impermissible uses, mention the case of tax authorities resorting topredictive control systems based on artificial intelligence using data from the tax returns of all taxpayers, in relation also to other data of various kinds (social habits, etc.).

The Italian situation

As far as Italy is concerned, we must note that the process of digitisation of tax implementation processes is now at an advanced stage, and on this reality the legislator of Delegated Law 111/2023 has structured the guiding principles and criteria (Articles 2, 4 and 17) for the tax reform by promoting the interconnection of databases and the use of artificial intelligence systems. The choices made through the implementing decrees show the construction of a tax system structured taking into account technological evolution, aimed at inducing taxpayers to spontaneous compliance, according to a logic of facilitation of procedures within a broader preventive strategy of anticipating control and assessment.

Where the data and their cross-referencing allow it (partial assessments, checks on declarations, payments, credits, etc.), automated inspections are strengthened, for which the legislator has chosen to exclude cross-examination; where, on the other hand, it is essential to carry out assessments, procedural modules inspired by the cooperation between the tax authorities and taxpayers are used, which often start, however, from findings of non-compliance with sector standards, the outcome of the use of Ia. Thealgorithms now optimise inspection procedures through more sophisticated models for managing the risk of tax avoidance or evasion, undoubtedly facilitate information exchange procedures for mutual assistance and cooperation in the international sphere, and more.

All these activities are aimed at the adoption of acts that - also on the basis of the jurisprudence of the Council of State on the use of algorithms for the purpose of adopting administrative measures (judgments 2270/2019 and 881/2020) - must be based on a clear and comprehensible logical-legal path and with respect to which the Ia must remain a mere operational tool at the service of human intelligence. Instead, from this point of view, possible uses aimed at the predictiveness of the tax assumption, as would appear in the case of the two-year prior agreement, are worrying, given that the structure and operation of the algorithm is not easy to understand from a reading of the methodological note, which, as things stand, does not allow discriminatory and prejudicial profiles to be ruled out. On the contrary, the relative discipline seems to be a valid example of how technology can constitute the silent and obscure vehicle of legislative choicesdiscriminatory in the name of an alleged efficiencyism functional, at the same time, for overcoming evasion and for the saving of costs from assessment.

It is precisely in these contexts that it is necessary to be vigilant as to the reasonableness of the choices made, because, especially in the absence of proven benefits for the generality of the citizens - that is, if, upon verification of the facts, the institution proves not to achieve the aforementioned purposes - the balance falters. Our Constitution does not allow the legislator to undermine the resilience of the tax system (Articles 2, 3 and 53 of the Constitution) through the adoption of disproportionate measures even if they are technologically advanced.