Artificial Intelligence in European Union Tax Administrations: A Comparative Assessment

Abstract

The study aims to examine trends in the integration of artificial intelligence within the operational processes of tax administrations across the Member States of the European Union. It explores both the functional domains in which AI can be deployed and the institutional, ethical, regulatory and technological constraints that shape its deeper integration. The analysis relies on publicly available data from the Organisation for Economic Co-operation and Development (OECD), complemented by information from other open sources. Based on this dataset, the study develops a Tax AI Index (TAI) to provide a comparative quantitative assessment of the extent to which AI systems have been operationally integrated into EU tax administrations. The index is constructed from four subindices capturing (1) the use of artificial intelligence in communication between tax administrations and economic agents (TAIIS); (2) the integration of artificial intelligence in data management systems (TAIDS); (3) the application of algorithmic systems in tax enforcement, compliance control and administrative decisions (TAIRES); and (4) mechanisms for accountability, transparency and ethical oversight in the use of artificial intelligence (TAIGS). The empirical results indicate significant heterogeneity in the levels of digital transformation among the EU-27 Member States. In most countries, the adoption of artificial intelligence remains at an experimental or pilot stage, suggesting that its broader operational application is still evolving. To place these findings in a broader context, the analysis is complemented by an external measure of digital government development, allowing for a comparative assessment between AI adoption in tax administrations and overall public sector digital maturity.

1. Introduction

Contemporary tax systems are characterised by increasing institutional complexity, persistent information asymmetries, structural constraints in administrative capacity and the diminishing effectiveness of traditional instruments for administrative and operational governance. These dynamics necessitate the systematic adoption of innovative policy and administrative approaches through which tax administrations can respond more effectively to emerging risks and operational challenges. At the same time, artificial intelligence (AI) has emerged as a transformative technology with expanding applications across economic and non-economic sectors. Although its diffusion is associated with efficiency gains and broader welfare improvement, it also generates significant regulatory, ethical and governance-related risks, particularly in relation to algorithmic bias, limited transparency, lack of explainability and challenges in assigning accountability for automated decisions (Veale & Brass, 2019; Kroll et al., 2017). These developments prompted the European Union (EU) in 2024 to adopt a harmonised regulatory framework governing artificial intelligence. Regulation (EU) 2024/1689 (European Union, 2024) establishes a formal and legally binding definition of an AI system, characterizing it as a “machine-based system” designed to operate with varying degrees of autonomy and capable of exhibiting adaptiveness after deployment. Importantly, such systems are defined by their ability to “generate outputs such as predictions, content, recommendations, or decisions that can influence physical or virtual environments” (European Union, 2024), thereby highlighting their capacity to shape economic behaviour and institutional processes. In taxation, these characteristics are particularly consequential, as the use of AI in areas such as risk profiling, audit selection and compliance monitoring may directly affect taxpayers’ rights, procedural fairness and the legitimacy of administrative decision-making.

Within this broader regulatory and technological context, taxation represents a particularly suitable sphere for AI application. Tax systems are characterised by complex procedures, extensive data and recurring compliance risks, which increase the potential benefits of algorithmic assistance. At the same time, the reliance on large-scale data processing and automated decision-making may amplify existing informational asymmetries and introduce new forms of systemic risk if not properly governed. At the European level, AI is conceptualised as a horizontal enabler capable of generating economic, social and environmental benefits across virtually all sectors of activity (European Union, 2024). In the field of taxation, AI may support both the strategic design and optimization of tax policy (Zheng et al., 2022; Mökander & Schroeder, 2024) and the operational function of tax administration, including risk assessment, compliance management, fraud detection and resource allocation. Ballas and Kolovou (2024) emphasise that, in the contemporary global context, the use of artificial intelligence in tax administrations should be understood as a necessity rather than a technological luxury. Against this background, a systematic and multidimensional assessment of the implications of AI integration in public finance becomes analytically indispensable.

The object of this study is the institutional activity of tax administrations within the EU Member States, whereas its subject concerns the degree and modalities of integration of AI-based instruments into their operational and strategic practices. The selection of the EU Member States as the territorial scope of the study is not arbitrary but reflects the interplay of both homogeneous and heterogeneous characteristics. On one hand, the European Union has established a common regulatory framework governing the application of artificial intelligence (Maragno et al., 2023), while also facilitating enhanced coordination among Member States. On the other hand, the tax administrations across the EU differ in terms of organizational structure, administrative capacity, functional responsibilities, and, as will be demonstrated in this study, their level of digitalization. This combination allows the study to capture both shared regulatory features and variations in AI adoption and implementation across Member States.

The primary objective is to examine the evolving role of AI in EU tax administrations. The study applies a structured comparative assessment across the EU-27 Member States. This assessment is based on a novel, author-developed framework built around the TAI and its components (TAIDS, TAIGS, TAIIS and TAIRES). Its aims to identify prevailing implementation patterns, to evaluate the associated advantages and to critically assess the risks inherent in excessive, opaque or inadequately governed reliance on algorithmic systems. This objective is particularly important given that AI has the potential not only to improve efficiency and compliance but also to introduce significant risks if deployed excessively or without appropriate governance (Gupta, 2019; Wirtz et al., 2019; Ulaşan, 2023; Vatamanu & Tofan, 2025). The study is relevant because it combines a review of the advantages and disadvantages of AI with a structured comparative assessment of its adoption across EU Member States, providing a balanced perspective. Its significance is further reinforced by the growing attention to AI in the public sector, which has historically received less scholarly focus compared to the private sector (Sun & Medaglia, 2019; Zuiderwijk, 2021; Munjeyi & Schutte, 2024), highlighting the need for systematic analysis in governmental contexts. In this regard, there is an evident need for comparative research examining how different tax administrations implement AI in practice, as existing evidence in this area remains limited (Islam et al., 2025), particularly through multi-dimensional cross-country benchmarking using composite AI adoption indicators.

In contrast to existing scholarship, which typically focuses on isolated dimensions such as legal compliance (Peeters, 2024, Boguski, 2025; Tilaganboev, 2026), ethical considerations (Jobin et al., 2019; Alarie & McCreight, 2023; Yordanova, 2024; Guglyuvatyy, 2025) or efficiency gains (Martinez, 2025), the present study develops an integrated analytical framework. This framework combines theoretical and descriptive analysis with a structured methodological approach for cross-country comparative assessment of AI integration within tax administrations in the European Union. It contributes to the literature by conceptualizing AI adoption as a broader institutional process and by offering a clear framework for systematic comparative evaluation. It also provides actionable insights for policymakers on both the opportunities and the risks associated with AI use in public administration, emphasizing that benefits and hazards must be considered together. The analysis focuses on the operational and strategic transformation of administrative processes rather than on the technical architecture of AI systems themselves. The main hypothesis posits that AI integration within EU tax administrations possesses substantial potential to mitigate operational frictions, reduce non-compliant practices, enhance fiscal capacity and improve allocative efficiency. Nonetheless, such digital transformation simultaneously entails significant systemic and governance risks, including risks that transcend the immediate boundaries of tax administration, particularly for institutions operating at an early stage of technological maturity.

The paper is structured as follows: Section 2 provides a comprehensive review of the dominant strands in the academic and policy literature on AI applications in tax administration. This section further identifies key institutional and operational challenges associated with AI adoption. Section 3 combines a descriptive analysis of AI implementation practices across EU Member States with the development of a quantitative comparative framework designed to assess relative levels of AI integration. Section 4 applies the proposed methodological framework and presents the empirical findings, which are discussed both at an aggregate level of integration and to a more detailed examination of individual subindices capturing specific functional domains of AI applicability. Section 5 presents the conclusion.

2. Literature Review

From a theoretical perspective, the integration of artificial intelligence into tax administration can be interpreted within the broader framework of information economics and public sector governance. Tax systems are inherently characterized by information asymmetries and principal–agent problems, which constrain the effectiveness of traditional enforcement and compliance mechanisms (Akerlof, 1970; Laffont & Martimort, 2002). While taxpayers often possess private information about their real financial situation (Akerlof, 1970), tax authorities must rely on observable reports and designed incentive structures to enforce compliance (Laffont & Martimort, 2002). These features constrain the effectiveness of traditional enforcement and compliance mechanisms. In this context, AI-based systems can be conceptualized as tools that reduce informational frictions, enhance predictive capacity and support more efficient allocation of administrative resources (Wirtz et al., 2019). At the same time, however, their deployment introduces new governance challenges related to opacity, accountability and the potential amplification of existing biases (Veale & Brass, 2019), thereby requiring careful institutional design and oversight.

Although increasing attention is currently being devoted to the integration of AI into operational and managerial processes, a historical perspective reveals that various forms of artificial intelligence can be traced considerably further back. According to data from the OECD (2025b), approximately a decade ago nearly 1/10 of public administrations had already incorporated AI tools into their operational processes. Nevertheless, in recent years tax administrations at the global level have increasingly sought to take advantage of the opportunities afforded by advanced digital and technological instruments. OECD (2025b) further reports that by 2023 this proportion had risen steadily to nearly 70%, providing compelling evidence of the accelerated diffusion of AI technologies within the public sector. Projections suggest a continued and potentially exponential expansion in the deployment of AI for taxation purposes in the coming years.

De La Feria and Grau Ruiz (2022) identify several structural drivers underlying the growing momentum toward AI integration in tax administration. These include the fiscal pressures stemming from the global financial crisis at the end of the first decade of the twenty-first century and its implications for public finance sustainability, the pursuit of greater efficiency and effectiveness in tax enforcement; the mitigation of systematic biases inherent in human (subjective) decision-making in the pursuit of administrative fairness and the reduction in exposure to corruption-related risks.

2.1. Principal Functional Domains of AI Deployment Within Tax Administrations

A substantial body of academic literature focuses on the implications and consequences of artificial intelligence applications in tax processes. It should be noted that research practice is characterised by its dynamism and growing interest, especially over the past few years. Among the main objects of analysis in most of these studies are the following:

The deployment of AI for the optimization of operational processes within tax administrations

Tax administrations may deploy AI-based instruments to automate a wide range of operational activities, ranging from the routine processing of tax returns and related financial documentation to the advanced processing and analysis of extensive unstructured data (Janssen et al., 2020). In practical terms, such automation supports a more efficient allocation of administrative resources and reduces reliance on human labour in routine and standardised activities (Salah & Awwad, 2024). On the other hand, artificial intelligence systems may support the digitalization and automation of audit, verification and control procedures, thereby minimizing the likelihood of operational errors and reducing the level of bureaucratic burden for economic agents (Onet, 2025). These systems are further used in strategic planning related to tax collection activities and in efforts to reduce transaction and administrative costs (Munjeyi & Schutte, 2024; Rekunenko et al., 2025). Gabriel et al. (2025) emphasise that the adoption of artificial intelligence in tax administration may transcend technological modernisation and function as a structural driver of economic development and institutional capacity building. Accordingly, one of the core principles for the responsible governance of trustworthy AI, articulated by the OECD (2019), explicitly emphasises the role of AI in promoting inclusive growth, sustainable development and societal well-being.

The use of AI in tax fraud detection and prevention

Faúndez-Ugalde et al. (2020) emphasise that modern tax administrations primarily focus on ensuring compliance with tax legislation, particularly by detecting tax fraud. Consequently, existing academic and institutional studies indicate that one of the principal advantages of AI deployment in tax administration lies in its capacity to detect potential instances of unlawful tax practices through advanced process modelling and the identification of anomalous patterns in financial transactions (Lin et al., 2021). Such models also make it possible to detect hidden relational patterns and network connections among taxpayers involved in coordinated or extensive illegal schemes, particularly in the field of indirect taxation, including cases of tax misreporting or systemic compliance failures (Nyok, 2025). In this context, Yalamati (2023) argues that the continuous learning capacity of AI systems enables them to adapt dynamically to evolving fraud patterns, thereby providing tax administrations with a proactive mechanism for addressing emerging compliance risks.

López et al. (2019) investigate the application of neural networks, as a class of machine learning models, for the detection of personal income tax fraud in Spain and conclude that such tools, on one hand, do not require significant fiscal resources and, on the other hand, demonstrate a high level of effectiveness (approximately 84%).

Belahouaoui and Alm (2025) systematise several categories of determinants influencing the effective use of AI in detecting tax fraud, including technological (availability, quality and integration of data infrastructures), organizational (primarily related to administrative and managerial capacity), institutional and regulatory (linked to the legal framework and ethical standards), economic and financial (related to investment costs of digital transformation) and socio-political factors (with primary emphasis on public trust and societal acceptance of AI-based governance mechanisms).

However, Anjarwi (2026) cautions that the expanding use of AI may also generate countervailing effects, insofar as similar technologies can be leveraged by taxpayers to develop increasingly sophisticated and less detectable strategies for unlawful tax minimisation. Chen et al. (2017) emphasise that the deployment of digital technologies in tax administration requires thorough ex ante evaluation. Without such assessment, unintended behavioural responses by taxpayers may emerge, potentially exerting a detrimental effect on overall compliance.

The use of AI in forecasting and predictive analytics

Beyond operational activities, AI may also be used as a strategic instrument supporting the forecasting process of expected taxpayer liabilities and, consequently, revenue collection performance (Bankole et al., 2025). In this context, AI facilitates the construction of models that systematically link tax dynamics with broader macroeconomic and fiscal development trends. In this function of tax administrations as well, machine learning algorithms constitute the principal analytical instruments. The strength of these models lies in their ability to capture not only linear relationships between macroeconomic and fiscal developments but also to identify and empirically validate complex nonlinear dependencies derived from extensive empirical evidence. Moreover, these approaches address several limitations of traditional forecasting techniques, particularly their vulnerability to model specification errors and delayed adjustment in the presence of sudden exogenous or endogenous economic shocks. Additional advantages of machine learning include the integration of diverse data sources for forecasting purposes and the early detection of structural shifts in the tax base driven by behavioural changes among economic agents. Van Duc et al. (2024) also highlight that AI-based forecasting models can provide early warnings of potential revenue shortfalls, thereby enabling governments to implement timely corrective measures to mitigate emerging fiscal risks. It also enables advanced quantitative simulations to assess the potential effects of macroeconomic and policy variables on tax revenue performance.

The use of AI in risk assessment of specific taxpayers (profiling and targeting) in audit activities

Another significant advantage of AI tools, particularly in the control functions of tax administrations, is their capacity to enhance the accuracy of analysing and evaluating taxpayers’ economic activities (Zhou, 2019). Shakil and Tasnia (2022) and Bensaci and Knenouf (2026) emphasise that such analytical capabilities are especially relevant for monitoring multinational enterprises. To this end and for the purpose of clustering taxpayer risk profiles, AI enables the analysis of historical and behavioural information obtained from prior audits, tax returns and related documentation. It also allows the integration of extensive datasets and multiple parameters within the assessment process (Battaglini et al., 2025). In this context, AI operates as a real-time tax risk simulation mechanism, generating predictive risk scores that estimate the likelihood of non-compliance and adjust classifications as new data emerge (Min & Yanting, 2018). This approach facilitates the identification of distinct risk groups and enables the strategic allocation of administrative resources toward high-risk categories, while reducing resource deployment in areas characterised by routine activity and minimal risk exposure. Alexopoulos et al. (2023) apply an AI model and conclude that, during 2016–2017, the Bulgarian National Revenue Agency audited approximately 15,000 tax returns per month classified as high risk. Their model, however, suggests that the actual number of genuinely high-risk cases may have been substantially lower, potentially around 500 or, in some instances, as few as 100, thereby considerably reducing the resources required for such audits. Typically, predictive models and machine learning pipelines are employed to collect and preprocess data, select appropriate modelling algorithms, train models on historical datasets and validate their predictive performance against prospective outcomes. It should be noted that AI tools such as machine learning systems also enable periodic data updates and dynamic reclassification of taxpayer risk profiles in response to evolving information (Rahman et al., 2024).

The use of AI for enhancing taxpayer services and administrative efficiency

Certain types of AI-based instruments, such as chatbots or virtual assistants, can help to improve the quality of services provided by tax administrations to economic agents. These tools are especially valuable for addressing routine taxpayer inquiries, facilitating the submission and processing of tax returns and supporting the delivery of administrative services. A key advantage of such AI tools is their continuous and unrestricted availability to taxpayers (Aldemir & Uysal, 2025), in contrast to the temporal limitations associated with standard administrative working hours. Empirical evidence generally suggests that AI adoption in service delivery can enhance user trust and satisfaction among economic agents. Rahman et al. (2024) emphasise that the effectiveness of such initiatives in tax administration is contingent upon the digital literacy of taxpayers (service users) and their capacity to engage with technological interfaces. In recent years, numerous tax administrations have deployed AI tools to expedite the processing of tax returns and improve administrative efficiency.

The effective application of artificial intelligence in tax administration depends fundamentally on coordinated interaction between AI systems and human expertise within administrative units. It is misguided to assume that AI adoption will necessarily result in the displacement of or reduction in expert personnel within tax administrations. On the contrary, final decision-making authority should remain with human experts within revenue authorities, informed by AI-driven data processing and analytical recommendations.

2.2. Challenges in the Implementation and Use of Artificial Intelligence Within Tax Administrations

Undoubtedly, the use of artificial intelligence may significantly enhance the functioning of tax administrations across multiple dimensions. At the same time, the implementation of these technologies entails significant challenges whose importance is comparable to the benefits of their adoption. The use of AI raises a range of questions extending beyond tax administration and relating to the broader evolution of global economic and institutional processes (Campion et al., 2022). This is why, in recent scholarship, growing emphasis has been placed on the potential adverse consequences of AI adoption and its characterization as a significant long-term technological and systemic global risk (Nikolova, 2024).

One of the greatest challenges in the potential use of AI in tax processes is related to legislative decisions in individual countries to introduce frequent changes in the regulatory and tax framework (Mpofu, 2024). As a result, administrations may encounter technical difficulties arising from the need to reconfigure AI-based instruments, especially when legislative changes are significant. These challenges arise not only in tax administrations with advanced AI integration but also in institutions that are in the early stages of AI adoption or implementing pilot initiatives. Furthermore, the maintenance and enhancement of digital infrastructure is usually associated with innovative solutions requiring substantial investments and the availability of highly qualified experts within tax administrations capable of operating AI technologies and interpreting analytical outputs (Shakil & Tasnia 2022; Djellaba et al., 2024; Hristov, 2025; Nyok, 2025; Pamisetty, 2025; Bensaci & Knenouf, 2026). Butt (2024) also notes that the higher costs associated with the implementation of AI may pose significant constraints for administrations with more limited institutional and financial capacity. When AI implementation in tax administrations is considered within the broader e-government development process, technical competencies of public sector employees emerge as a key enabling factor (Ziemba et al., 2016). However, skill gaps may constrain the pace and effectiveness of digital transformation.

A further substantive critique of artificial intelligence is that many AI-based applications operate as a “black box”, characterised by limited transparency and constrained oversight of decision-making processes—an issue relevant beyond tax administration (Richmond et al., 2023; Han et al., 2025). Through machine learning and related analytical techniques, AI systems process and model information. This creates risks, including the possibility that processed information may lack full reliability or completeness. On the other hand, algorithmic modelling may generate risks of bias and disproportionate targeting of taxpayers classified as high-risk in the absence of adequate evidentiary foundations (Nieto Olvera, 2025; Islam et al., 2025), potentially contributing to forms of social inequality (Martinez, 2025). Janssen et al. (2020) and Ulaşan (2023) argue that the behaviour of AI systems stems from the fact that they are trained on historical data, which often contains elements of discrimination reflecting existing social biases. A similar case occurred in an EU Member State, the Netherlands, where for several years thousands of foreign residents were accused by the tax administration of fraudulent claims related to child benefit payments. The problems originated from the deployment of an algorithmic tax fraud risk assessment system used in the administration of child benefit payments. The system relied on criteria later regarded as discriminatory, including the volume of submitted claims and nationality, which influenced the classification of certain applicants as high risk. Consequently, numerous affected individuals were required to reimburse funds, thereby undermining their financial stability and economic security. The scandal ultimately led to the resignation of the national government. Although the case in the Netherlands is not fully analogous to contemporary autonomous AI systems, it illustrates that automated decision-making under certain conditions can generate significant socio-economic consequences. This necessitates a decision at the European level to provide mechanisms for effective human oversight whenever high-risk systems are deployed (European Union, 2024). Although the European regulatory framework does not classify AI systems used by tax authorities as high-risk in administrative procedures, similar preventive measures should nevertheless be applied within tax administrations to ensure transparency, accuracy and security.

The lack of transparency in the processes of modelling and taxpayer classification renders tax administrations institutionally vulnerable in the event of disputes initiated by affected economic agents (Sun & Medaglia, 2019). Guglyuvatyy (2025) highlights the so-called right to explanation, which seeks to safeguard taxpayers’ procedural rights by enabling them to understand how AI systems generate specific assessments (e.g., risk scores, tax liabilities, inconsistencies). Gupta (2019) further highlights an additional concern, namely citizens’ diminished trust in public institutions when they are aware that decisions are being made not by human agents but by digital assistants. In this regard, the Finnish Tax Administration has developed a set of ethical principles governing the deployment of AI, with particular emphasis on ensuring meaningful human oversight and accountability in AI-assisted decision-making, including the authority to override algorithmic outputs. Within the broader debate on transparency in the application of AI in tax administrations, scholars argue that the trajectory of AI integration may follow two structurally divergent logics. One is oriented towards enhanced procedural transparency and the constraint of discretionary power, while the other suggests that algorithmic infrastructures, if insufficiently transparent, may inadvertently or deliberately facilitate the obfuscation of corrupt practices, thereby weakening mechanisms of institutional accountability (AllahRakha, 2025).

Among the principal challenges associated with the integration of AI into tax processes is the fact that substantial volumes of sensitive personal data become subject to algorithmic processing and control, which may at times conflict with the constitutional rights of citizens (Pica, 2021). In addition, the rapid technological expansion accompanying AI deployment increases exposure to cyber risks and broader systemic digital vulnerabilities (Stoyanov, 2024). It should also be noted that the processing of large amounts of data generates risks related to confidentiality and data protection, which require appropriate legal regulation, particularly with regard to AI-driven profiling and automated decision-making.

Peeters (2024) observes that under Regulation (EU) 2016/679 (European Union, 2016), automated decision-making, including profiling, is subject to restrictions where it produces legal effects or similarly significant consequences for data subjects. However, as the author notes, the Regulation provides for specific derogations that may extend to tax administrations under certain conditions.

Boguski (2025) argues that, particularly at the European level, the pronounced heterogeneity of national tax legislation and domestic rules governing AI deployment constitutes a major impediment to deeper AI integration in taxation. For this reason, efforts are underway at the EU level to establish a common legal framework aimed at harmonising the principles and regulatory standards applicable to artificial intelligence (European Union, 2024). Conversely, Hossain et al. (2025) report that, according to a survey of multinational enterprises, a majority of respondents consider that AI may facilitate greater standardisation across tax systems, thereby reducing administrative complexity and, notably, limiting opportunities for aggressive transfer pricing practices.

It should also be emphasised that the quality of data obtained from diverse sources may generate additional challenges, particularly with respect to reliability and evidentiary validity. This is especially relevant where information is aggregated from social media platforms, interoperable public registers, or commercial databases. In the context of heterogeneous tax legislation and multi-source data processing, any prospective exchange of information between tax administrations within international cooperation frameworks requires careful legal assessment. This includes questions of jurisdiction, data protection compliance and institutional accountability.

3. Data and Methodology

The methodological framework of the present study is based on a comparative quantitative assessment of the level of AI tool implementation within EU tax administrations. The study covers the 27 Member States of the European Union.

The empirical foundation of the analysis draws upon data from TaxAdmin.AI (n.d., developed by David Hadwick), the Inventory of Tax Technology Initiatives (OECD, n.d.) and the International Survey on Revenue Administration—ISORA (International Monetary Fund, n.d.). ISORA is a recurring joint publication of the OECD, the IMF, the World Bank and the Intra-European Organisation of Tax Administrations (IOTA). It should be noted that the data sources rely predominantly on publicly available information and administrative self-reporting, which may affect both the precision and cross-country comparability of the estimated effectiveness and fiscal outcomes. In addition, the measurement of AI adoption is constrained by the absence of a universally accepted operational definition suitable for empirical research, requiring the use of proxy indicators (Wirtz et al., 2019). Although the EU AI Act (European Union, 2024) establishes a regulatory definition of an AI system, it does not eliminate conceptual ambiguities relevant to cross-country comparisons or quantitative measurement, given the diversity of AI applications across public organizations. Nevertheless, a Tax AI Index (TAI) has been developed to provide a comparative quantitative assessment of the extent to which AI systems have been operationally integrated into EU tax administrations. The index is designed to capture both taxpayer-facing AI applications and internal AI deployments related to data management, decision support and process oversight. It consists of four thematically structured subindices, each representing a distinct dimension of AI adoption (Figure 1).

Data on specific indicators from the Inventory of Tax Technology Initiatives (OECD, n.d.) are used to construct and calculate each subindex. The ISORA dataset reflects the most recent update as of 1 October 2025. The application of four subindices facilitates a more flexible and analytical examination of the components of AI integration. This makes it possible to identify the strengths and weaknesses of AI application. Moreover, the selection of these analytical dimensions is based on the literature review and on core structural elements of the conceptual framework of the Tax Administration 3.0 model developed by the OECD (2020), which outlines the strategic foundations of digitally enabled tax administration.

3.1. General Framework for Calculating the Subindices

The subindices (TAIIS, TAIDS, TAIRES, TAIGS) for the tax administration of a given Member State, ${SUBINDEX}_{i,j}$, are constructed within a unified methodological framework, according to baseline Equation (1):

$${SUBINDEX}_{i,j}={\displaystyle \frac{1}{\mathrm{N}}}\sum _{z=1}^{\mathrm{N}}{\mathrm{S}\mathrm{B}\mathrm{I}\mathrm{N}\mathrm{D}}_{z,j}$$

(1)

where the following definitions are used:

${\mathrm{S}\mathrm{B}\mathrm{I}\mathrm{N}\mathrm{D}}_{z,j}$—binary value (0 or 1) assigned to each indicator z included in the subindex for a given Member State j; ${\mathrm{S}\mathrm{B}\mathrm{I}\mathrm{N}\mathrm{D}}_{z,j}\in \left\{\mathrm{0,1}\right\}$;

N—number of indicators z included in the subindex.

The use of a single aggregation mechanism ensures full methodological comparability across all four subindices, while allowing them to capture distinct functional dimensions of AI deployment in tax administration. After calculating the subindex value for each of the 27 EU Member States, the scores are normalised in order to ensure comparability between states. In their initial form, the subindices are based on binary indicators and reflect the presence or absence of specific initiatives. Due to the varying number of relevant indicators across functional domains and the uneven diffusion of practices among states, observed values may remain significantly below the theoretical maximum of 1. This does not imply a weak relative position but rather reflects the early stage of institutionalisation of AI solutions in most tax administrations. To address this, normalisation is performed with respect to the highest observed value for each subindex, as adopted in the present study. This approach constitutes linear scaling, whereby each value is expressed as a proportion of the empirically established maximum, ${(\max }_{\delta }\left({SUBINDEX}_{i,\delta }\right),\delta \in EU)$. This procedure does not alter the relative differences between EU Member States or their ranking. It merely transforms the scale so that the highest observed value is fixed at 1, with all other values distributed proportionally.

Consequently, normalization does not “inflate” results but preserves structural relationships and relative distances while improving interpretability. The methodological rationale for this approach lies in the comparative design of the study, which seeks to evaluate relative performance within the observed sample rather than absolute technological maturity against an abstract theoretical benchmark. In such a comparative framework, the use of the empirical maximum as a reference point is methodologically justified, as it reflects the highest level actually achieved within the EU at the time of observation. Furthermore, normalization is applied symmetrically across all Member States and subindices, ensuring methodological neutrality in the assessment of their tax administrations. Because the transformation is linear, it does not distort the distribution, the relative variance is preserved and no artificial differences are introduced. In this sense, normalization simply ensures comparability across scales and does not serve to inflate scores.

It should be emphasised that, even after normalisation, a value of 1 denotes the empirically leading performer rather than an absolute frontier of development. Accordingly, the results should be interpreted as relative measures of comparative positioning, not as absolute indicators of progress in AI governance or technological adoption.

The normalised SUBINDEX values for each Member State are calculated as specified in Equation (2):

$${SUBINDEX}_{i,j}^{\prime }={\displaystyle \frac{{SUBINDEX}_{i,j}}{{{\mathrm{m}\mathrm{a}\mathrm{x}}_{\delta }(SUBINDEX}_{i,\delta })}}$$

(2)

where the following definitions are used:

${SUBINDEX}_{i,j}$—subindex i value for Member State j;

${{\mathrm{m}\mathrm{a}\mathrm{x}}_{\delta }(SUBINDEX}_{i,\delta })$—highest empirical value of the subindex i among all 27 EU Member States ($\delta \in EU$).

3.2. Structural Composition and Description of the Subindices

The calculation of the four subindices is based on a comprehensive set of 42 indicators. Seven indicators form the TAI Interaction subindex, the TAI Data subindex is the most extensive, encompassing 13 indicators, the TAI Rule Enforcement Subindex includes 10 indicators and the TAI Governance subindex comprises 12 indicators. The overall structural framework is summarized in

Table 1. Indicators comprising the structure of the subindices.

TAIIS		TAIDS
TAIIS_1	Rule-based virtual assistant	TAIDS_1	AI in data governance
TAIIS_2	AI virtual assistant (personalised)	TAIDS_2	Real-time fraud detection
TAIIS_3	Filing assistance	TAIDS_3	AI/ML big data analysis
TAIIS_4	Other interactions	TAIDS_4	Big data for compliance
TAIIS_5	Suggested responses	TAIDS_5	Big data trend analysis
TAIIS_6	Live chat support (officials)	TAIDS_6	Policy forecasting
TAIIS_7	Automated service responses	TAIDS_7	Revenue forecasting
		TAIDS_8	New service development
		TAIDS_9	Robotic process automation
		TAIDS_10	Artificial intelligence use
		TAIDS_11	Machine learning
		TAIDS_12	Network analysis
		TAIDS_13	DataOps approach
TAIRES		TAIGS
TAIRES_1	Personalised information provision	TAIGS_1	AI use limitations
TAIRES_2	Virtual assistants (use)	TAIGS_2	Ethical framework
TAIRES_3	Risk assessment	TAIGS_3	Internal code review
TAIRES_4	Fraud detection	TAIGS_4	Internal input review
TAIRES_5	Decision support	TAIGS_5	Internal testing
TAIRES_6	Action recommendations	TAIGS_6	Internal monitoring
TAIRES_7	Automated decisions	TAIGS_7	Other internal measures
TAIRES_8	Dispute resolution	TAIGS_8	External code review
TAIRES_9	System integrity	TAIGS_9	External input review
TAIRES_10	Other applications	TAIGS_10	External testing
		TAIGS_11	External monitoring
		TAIGS_12	Other external measures

.

To enhance clarity and support interpretation, each indicator is assigned both a code (e.g., TAIIS_1, TAIDS_2, TAIRES_3, TAIGS_4, etc.) and a concise short label, with full descriptions provided in Appendix A.

• TAI Interaction subindex

Through the TAI Interaction subindex (TAIIS), the objective is to measure the degree of integration of algorithmic and automated systems within communication processes between the tax administration and external stakeholders (taxpayers and intermediaries), as well as within internal communication channels with staff. Conceptually, the subindex is part of the theoretical framework of digital public administration transformation. It reflects the transition from traditional electronic services (e.g., e-government) towards AI-enabled administrative governance. It captures not only the presence of digital interfaces but also the degree of algorithmic adaptability and personalised service provision. According to OECD (2020), future tax administrations should be primarily oriented towards supporting taxpayers.

The indicators distinguish between rule-based automation (pre-programmed responses) and AI-driven personalization in virtual assistants. They also capture functional AI applications across interaction phases. These include support for tax declaration submission, automated response generation and assistance for staff during live communication.

• TAI Data subindex

The TAI Data subindex (TAIDS) captures the degree of integration of artificial intelligence and related analytical technologies within the tax administration’s data governance and management architecture. Conceptually, the subindex builds on the data-driven governance paradigm, which treats data as a strategic asset and analytical tools as core drivers of institutional effectiveness.

The indicators presented in Table 1 cover three interrelated dimensions: (1) AI within data governance, reflecting the integration of AI into data management processes, big data analytics and real-time fraud detection; (2) technological infrastructure, including the use of robotic process automation (RPA), machine learning, network analysis, DataOps and other innovative approaches; and (3) the functional application of big data to enhance revenue collection, identify behavioural patterns, support revenue and policy forecasting and enable the development of new services.

In essence, the subindex measures the analytical capacity of the administration. It assesses the ability to transform large volumes of structured and unstructured data into actionable managerial information and support strategic decision-making.

• TAI Rule Enforcement subindex

The TAI Rule Enforcement subindex (TAIRES) assesses the use of algorithmic systems in the enforcement of tax legislation, administrative control and decision-making processes. This subindex is most directly linked to the core regulatory function of the tax administration. It covers the full spectrum of algorithmic involvement—from analytical support to partially or fully automated decision-making.

The selected indicators of the TAI Rule Enforcement Subindex (TAIRES) include activities related to risk assessment and profiling, as well as fraud detection. They also cover decision-support functions for administrative staff, including recommendation generation and the potential for automated administrative decisions.

• TAI Governance subindex

The TAI Governance subindex (TAIGS) evaluates the presence of mechanisms for control, accountability, transparency and ethical oversight in the use of artificial intelligence. It focuses on the administration’s capacity to manage risks associated with automated decision-making, including algorithmic bias.

The subindex encompasses processes such as AI code review, the validation of input data, the testing of AI outputs, the monitoring of results and other risk management approaches. Its primary objective is to capture the extent to which tax administrations ensure the safe, ethical and reliable use of AI through oversight of system development, deployment and operation.

The indicators comprising the subindex reflect both technical and organisational measures for the control and governance of AI. The indicators can be grouped into two analytical dimensions. The first relates to internal mechanisms of control, accountability and transparency. These typically include restrictions on AI use, the establishment of an ethical framework, code review procedures and testing and monitoring. The second dimension concerns external oversight mechanisms, specifically the possibility of review and monitoring by a third-party entity.

3.3. Standardized Interpretation Framework for TAI Subindices

The results may be interpreted using a balanced scale based on the normalised values, as shown in

Table 2. Standardized comparative interpretation scale for TAI sub-indices.

${\mathbf{S}\mathbf{U}\mathbf{B}\mathbf{I}\mathbf{N}\mathbf{D}\mathbf{E}\mathbf{X}}_{\mathbf{i},\mathbf{j}}^{\prime }$	$\mathbf{D}\mathbf{e}\mathbf{s}\mathbf{c}\mathbf{r}\mathbf{i}\mathbf{p}\mathbf{t}\mathbf{i}\mathbf{o}\mathbf{n}$
0.00–0.20	Very low relative positioning, indicating minimal or no systematic use of AI-based tools and practices.
0.21–0.40	Low level of development, characterised by limited and fragmented adoption of AI solutions.
0.41–0.60	Moderate relative positioning, with AI supporting selected processes in a structured but not comprehensive manner.
0.61–0.80	Relatively high level of development, with broader and more systematic integration of AI across functions.
0.81–1.00	Highest observed level within the sample, indicating a leading relative position without implying full maturity or optimality.

. It should be noted that the interval-based classification (0.20 increments) is not derived from the empirical distribution of the data but is intentionally defined as a uniform interpretative scale. This approach is consistent with the normalised [0,1] structure of the indices and ensures symmetry and comparability across all subin-dices.

The same classification framework is consistently applied to all subindices examined in the study. The purpose of the scale is not to provide a statistically determined clustering of Member States but to facilitate analytical interpretation and relative positioning within the sample. Accordingly, the thresholds should be understood as heuristic and not as data-driven breakpoints.

Higher normalised values indicate a stronger relative position within the sample in terms of AI integration across these core functions. This reflects comparative, not absolute, differences among tax administrations. Therefore, for example, a higher normalised TAIIS score indicates a stronger relative position within the sample in terms of automated adaptability and AI deployment in front-office operations.

3.4. Internal Consistency of the Subindices

To ensure the robustness of the constructed subindices, it is recommended to evaluate their internal consistency. While internal consistency is most commonly assessed using Cronbach’s α, the binary nature of the indicators in the present study necessitates the use of the Kuder–Richardson Formula 20 (KR-20, Equation (3)). As a special case of Cronbach’s α for dichotomous variables, KR-20 provides an appropriate measure of reliability, capturing the extent to which the indicators consistently reflect the same underlying construct.

$$\mathrm{K}\mathrm{R}-20={\displaystyle \frac{k}{k-1}}\left(1-{\displaystyle \frac{{\sum }_{z=1}^k{p}_z{q}_z}{{\sigma }_{X_i}^2}}\right)$$

(3)

where the following definitions are used:

$k$—number of binary indicators z within the subindex;

$p_z$—proportion of the 27 EU Member States that exhibit a positive outcome (i.e., a value of 1) for indicator z;

$q_z$—proportion of the 27 EU Member States that exhibit a negative outcome (i.e., a value of 0) for indicator z;

${\sigma }_{X_i}^2$—variance of the total score $X_{i,j}$ for subindex i across the 27 Member States, where $X_{i,j}={\sum }_{z=1}^N{\mathrm{S}\mathrm{B}\mathrm{I}\mathrm{N}\mathrm{D}}_{z,j}$ is the total sum of binary indicator values within subindex i for Member State j.

Assessing internal consistency in this way helps to strengthen the validity of subsequent analytical findings.

3.5. TAI

The composite TAI is calculated as a weighted average of the normalized subindex values ${SUBINDEX}_{i,j}^{\prime }$ for each Member State j, as specified in Equation (4).

$$TAI=\sum _{i=1}^4{w}_i \ast {SUBINDEX}_{i,j}^{\prime }$$

(4)

where the following definitions are used:

$w_i$—weight assigned to each subindex i, ${\sum }_{i=1}^4{w}_i=1$, $w_i\ge 0$;

${SUBINDEX}_{i,j}^{\prime }$—normalised value of subindex i for Member State j, ${Subindex}_{i,j}^{\prime }$ ∈ [0,1].

From a methodological perspective, the present study applies differentiated weights to the individual subindices in the calculation of the composite TAI. The use of differentiated weights reflects the functional significance of the various dimensions of artificial intelligence adoption in tax administration. The objective of the index is not merely to capture the presence of technological tools but to assess the strategic depth and institutional maturity of AI integration. Therefore, the weights $w_i$ specified in

Table 3. Subindex weights in the composite TAI calculation.

${\mathit{S}\mathit{U}\mathit{B}\mathit{I}\mathit{N}\mathit{D}\mathit{E}\mathit{X}}_{\mathit{i},\mathit{j}}^{\prime }$	${\mathit{w}}_{\mathit{i}}$
TAIIS	15%
TAIDS	25%
TAIRES	35%
TAIGS	25%

are applied.

The highest weight is assigned to the TAI Rule Enforcement subindex (35%), as it captures core regulatory and fiscal functions of tax administration. These include risk assessment, fraud detection and algorithmically supported decision-making, which directly affect revenue collection. AI integration strengthens control and reduces information asymmetries, justifying its dominant weight.

The TAI Data subindex and the TAI Governance subindex are each assigned a weight of 25%. The data dimension reflects the infrastructural and analytical foundation of enforcement, as effective fraud detection depends on interoperable data systems and advanced analytical models. However, it retains an instrumental role, supporting rather than constituting enforcement outcomes. In parallel, the governance dimension reflects the growing importance of accountability, ethical standards and oversight of algorithmic decision-making. These mechanisms ensure institutional legitimacy, mitigate risks such as bias and opacity and support compliance with the rule of law.

The Interaction subindex receives the lowest weight (15%), as it primarily concerns communication and service delivery functions. Although digital interfaces improve efficiency and reduce transaction costs, their impact on revenue performance and enforcement capacity remains indirect and comparatively limited.

Within the evaluative component of the study, alternative computational specifications are employed. These include scenarios based on equal weighting of the subindices ($w_i$ = 25%, or 0.25) and scenarios with differentiated weights, where the most influential subindex receives greater emphasis and the least weighted subindex is reduced in importance. This procedure serves as a robustness check aimed at verifying the stability and sensitivity of the composite TAI under varying methodological assumptions.

The final determination of the degree of artificial intelligence integration and process automation may be carried out using the assessment scale presented in

Table 4. Comparative scale of overall AI integration in EU tax administrations.

TAI	Description
0.00–0.20	Lower relative position within the sample. AI tools are used minimally and in few processes, most administrative and communication functions remain predominantly traditional. This reflects comparatively low digitalisation rather than absence of AI initiatives.
0.21–0.40	Early stage of integration. AI deployment is limited to isolated functions and experimental applications. Partial automation exists, while systemic coordination remains limited. The administration is below the sample average.
0.41–0.60	Moderate relative position. AI is systematically applied in selected processes, supporting communication, analytics and control functions. Subindices indicate partial institutionalisation, though AI is not fully embedded in managerial structures.
0.61–0.80	Relatively advanced position. Intelligent systems are institutionalised across multiple processes and support decision-making. Governance mechanisms are in place, with a comparatively high level of automation within the sample.
0.81–1.00	Highest observed relative position within the sample. AI is strategically embedded in core functions, automation is widespread and governance is established. This reflects leading performance among Member States, without implying absolute technological maturity.

.

4. Results and Discussion

Each EU Member State maintains a functioning tax administration, which may be institutionally structured either as an autonomous body (most commonly an agency) or as a department directly subordinated to the Ministry of Finance. Irrespective of institutional design, tax administrations across the EU are increasingly pursuing the adoption and deployment of AI-based instruments, while those with prior experience in this field are seeking to further expand and refine their implementation. As outlined in the preceding literature review, the theoretical contributions of AI to tax administration are increasingly corroborated by empirical practice.

This section of the study applies the previously developed methodological framework to conduct a comparative assessment of EU Member States with respect to the degree of AI integration within their tax administrations. The evaluation is structured around a set of subindices, with the overall assessment derived from the composite TAI. It incorporates a total of 42 indicators, each coded in binary form (0 or 1) for the calculation of the respective subindices, thereby ensuring methodological consistency and cross-country comparability.

The results of the conducted empirical analysis indicate that the integration of artificial intelligence within EU tax administrations remains at a comparatively early and challenging stage. Even following normalization of the composite TAI, the average score across Member States amounts to 0.41. According to the scale presented in Table 4, this corresponds to a “moderate relative position”. When non-standardised subindex values are applied, the estimated TAI score decreases to approximately 0.30, further substantiating the conclusion that most Member States remain in the initial phase of AI-driven digital transformation within their tax administrations. Similar conclusions are reached by Munjeyi and Schutte (2024) based on previous empirical studies, in their search for the key factors influencing these processes. The results also lend partial support to the findings of Chen et al. (2017), who argue that some tax administrations remain cautious about the premature adoption of AI technologies. This reluctance is primarily driven by reputational considerations, potential risks to revenue collection performance and the possibility of incurring substantial additional costs due to unforeseen implementation errors.

Finland stands out among EU Member States as exhibiting one of the most advanced levels of AI integration within its tax administration. The results of the present study, derived from the composite TAI, corroborate the conclusions of the European Commission (2025b) concerning the high level of institutional and technological development of the Finnish tax authority. This finding is further supported by evidence from the Digital Economy and Society Index (DESI), which indicates that Finland is among the leading EU countries in terms of the share of AI specialists employed in the private sector, particularly within large enterprises (Eurostat, 2026). A similar pattern is observed in Denmark and Sweden, whereas Romania shows the opposite trend, with a relatively low share of AI specialists in the private sector. These differences may help explain the variation in TAI scores across Member States, as stronger private sector capabilities in AI appear to facilitate more effective adoption and integration within public administrations. Moreover, Finland’s model of AI deployment operates within a clearly defined regulatory framework and reflects a strong commitment to procedural transparency. Such features remain relatively uncommon at this stage of AI integration in public revenue administrations (European Commission, 2025b). The calculated TAI score for Finland reaches 0.88 (Figure 2), positioning it as the only Member State within the highest category defined in Table 4: “Highest observed relative position within the sample. AI is strategically embedded in core functions, automation is widespread and governance is established. This reflects leading performance among Member States, without implying absolute technological maturity.”

If the procedural logic proposed by Junquera et al. (2025) is followed, successful AI implementation in tax administration progresses through three principal phases:

• Creation (establishment) of an initial framework incorporating AI;
• Consolidation, characterised by the broadening of AI applicability and the transition from process design toward practical deployment and enhancement;
• Optimization of processes and continuous improvement.

The analysis of AI tool applicability within the tax administrations of EU Member States indicates that most remain in a pilot (project) phase, while for certain administrations, no information is available regarding the implementation of concrete initiatives in this area.

The group of administrations exhibiting higher levels of AI integration includes Spain (0.80), Italy (0.74), Ireland and France (0.70), Sweden (0.69), Lithuania (0.67), Denmark (0.66) and Austria (0.65). These data support the publicly available information from TaxAdmin.AI (n.d.), according to which the tax administrations of Spain and Sweden demonstrate the highest level of institutional maturity with respect to the number and functional scope of deployed AI systems. In a recent study, Jiménez and Mółka (2025) analytically examine a portion of the AI tools used by the tax agency of Spain, confirming the substantial progress observed in digitalization processes. The findings of the present study for Spain are consistent with these prior analyses, further corroborating the high level of institutional maturity and the functional scope of AI deployment within the agency. Additionally, in 2024 the Spanish tax agency published a strategic document aimed at promoting more extensive use of AI technologies (Agencia Tributaria, 2024). The tax administration of Denmark may also be classified among institutions characterised by high levels of public trust, attributable primarily to effective communication, transparent evaluation models and a robust institutional culture. Similar institutional characteristics (high public trust, transparent evaluation practices and a strong administrative culture) are observed across the other Scandinavian countries, including Sweden and Finland. These factors provide a compelling explanation for the relatively higher TAI scores recorded in these countries, reflecting their capacity to effectively implement and govern AI-based tools within tax administrations. In addition, Butt (2024) emphasizes that the high level of digitalization in Denmark and Sweden is also the result of prudent and consistent government policies, which have facilitated the establishment of specialized public institutions supporting the effective integration of digital and AI-based tools in the public sector.

Moderate rates of AI utilization in tax processes are observed in Poland (0.54), Estonia (0.52) and Hungary (0.46). The result for Estonia appears somewhat inconsistent with its classification, as it is identified, alongside Finland, by the European Commission (2025b) as one of the most advanced countries in terms of AI development in tax administration. The fourth group of countries, characterised by an initial stage of AI adoption, comprises Bulgaria (0.39), Greece and Slovenia (0.38), Germany (0.37), Latvia (0.27), the Netherlands (0.25) and Croatia (0.23). The lowest-performing group, in which administrative processes remain largely non-automated and data management and communication retain traditional modalities, includes countries such as Portugal (0.20), the Czech Republic (0.17), Slovakia (0.14), Cyprus and Malta (0.10), Belgium (0.06) and Luxembourg and Romania (0.04). According to data from selected indicators of the OECD (n.d.), these countries make only limited use of AI tools within their tax administrations. Available evidence suggests that, where such tools are implemented, they primarily confined to algorithmic or automated systems supporting communication with stakeholders or facilitating the processing of fiscal information. Somewhat surprisingly, Luxembourg, Malta and Belgium show low TAI scores. According to DESI (Eurostat, 2026), however, these countries display relatively high levels of digitalization in specific areas: Malta and Luxembourg lead in digital public services and service transparency, while Belgium, Luxembourg and Malta rank high in the availability of AI specialists in the private sector. These differences between TAI and DESI do not indicate a contradiction, but reflect the distinct scope and sources of the indices. DESI captures overall digital readiness, whereas TAI measures the specific deployment of AI in tax administrations. Thus, countries with high DESI values may still show low TAI scores, due to institutional or legal constraints and the self-reporting practices in OECD databases.

To provide additional context for the variation in TAI scores across Member States, the analysis is extended by incorporating the eGovernment Benchmark (European Commission, 2025a). This indicator reflects the general level of digital development in public administrations across the EU. It does not measure the performance of tax administrations directly, but captures the broader institutional environment in which they operate. This extension allows the results to be interpreted from two perspectives. One perspective relates to the use of AI in tax administrations. The other refers to the overall digital readiness of the public sector.

By combining these two dimensions (

Table 5. AI Adoption in Tax Administrations and Digital Government Maturity in the European Union: A Comparative Matrix.

		Digital Government Maturity
		Lower	Medium	High
Artificial intelligence adoption in tax administrations	Lower relative position within the sample	Cyprus, Romania, Slovakia	Belgium, Czech Republic	Luxembourg, Malta, Portugal
	Early stage of integration	Bulgaria, Croatia, Germany, Greece	Slovenia	Latvia, the Netherlands
	Moderate relative position	-	Hungary, Poland	Estonia
	Relatively advanced position	France, Italy	Denmark, Ireland, Spain, Sweden	Austria, Lithuania
	Highest observed relative position within the sample	-	-	Finland

), the study constructs a two-dimensional analytical matrix that distinguishes between levels of AI adoption in tax administrations (TAI) and overall digital government maturity. This approach enables a more nuanced comparative analysis of EU Member States and facilitates the identification of convergence and divergence patterns between administrative digital capacity and sector-specific AI implementation.

Countries are classified into three groups based on their position in the observed distribution of eGovernment Benchmark scores. These groups reflect low, medium and high levels of digital government maturity. A data-driven approach is applied in this classification. The sample is divided into tertiles, ensuring that each group contains an equal number of countries. This approach avoids arbitrary cut-off values and improves consistency across the classification.

A joint interpretation of both dimensions shows that high levels of general digital development do not always correspond to high TAI scores. This pattern is particularly evident in Luxembourg, Malta, Portugal, Latvia and the Netherlands. In general, countries with more developed digital administrations tend to achieve higher levels of AI adoption in tax systems. However, several exceptions are observed. At the same time, the results also reveal cases of strong alignment between the two indicators. Finland represents a clear example, combining the highest relative TAI score with a high level of digital maturity in the public sector. A similar consistency is observed at the lower end of the distribution. Cyprus, Romania and Slovakia display low levels of AI adoption in tax administrations, which correspond to lower levels of digital maturity. Comparable patterns, although less pronounced, can also be identified in Bulgaria, Greece, Germany and Croatia, where relatively modest TAI scores broadly align with their overall level of digital development. These patterns suggest that overall digital maturity may be associated with the level of AI adoption in tax administrations, without implying a direct causal relationship. Institutional settings, organisational structures and policy choices may also influence the extent to which AI is adopted and integrated within public administrations.

• Detailed analysis of the results by subindices

The values of the TAI for individual countries can be explained in greater depth through analysis of subindex performance. As previously noted, Finland demonstrates the most advanced integration of AI in the governance of tax processes. Across all four analytical dimensions (subindices), the country records high values compared to other EU Member States (Appendix B). Among the principal strengths of the Finnish tax administration are its data-driven governance model and the extensive use of algorithms to support strategic planning, monitoring and control, thereby reducing the scope of manual intervention and human oversight. At the same time, the Finnish administration appears to lag behind the administration of France in the degree of AI integration within communication processes involving stakeholders (taxpayers and administrative personnel). The tax administration of Finland also exhibits comparatively lower results (TAIGS = 0.67) relative to the administrations of Ireland (TAIGS = 1), Sweden (TAIGS = 0.78) and Austria (TAIGS = 0.78) with respect to indicators of accountability, transparency and ethical oversight of AI utilization, although the score remains high within the framework of the scale presented in Table 2. More extensive mechanisms of external oversight have not yet been implemented. Furthermore, with respect to TAIGS and TAIRES, a substantial number of EU Member States (eight in total) record a value of 0, indicating the absence of targeted initiatives employing AI in enforcement, control, accountability or transparency functions. These countries include the Netherlands, the Czech Republic, Slovakia, Cyprus, Malta, Belgium, Luxembourg and Romania. Among the countries mentioned, the Netherlands demonstrates particularly strong AI integration in data processing functions, a feature also observed in the tax administration of the Czech Republic and, to a more limited extent, Cyprus. The administrations of Malta, Romania and Slovakia remain at an early stage of AI integration, primarily focusing on applications involving algorithmic and automated systems within communication processes.

A significant number of tax administrations across EU Member States employ AI-based interfaces to automate taxpayer service processes. Most commonly, these interfaces take the form of chatbots designed to assist taxpayers in completing periodic (monthly, quarterly, annual) tax returns or in addressing specific tax inquiries. Through artificial intelligence technologies, chatbots process and interpret taxpayers’ queries (natural language processing), identify key entities within the request (entity recognition) and generate responses based on database retrieval or trained language models. The tax administration of Austria operates a chatbot known as “Fred,” which assists citizens in addressing tax inquiries. A comparable solution is provided by the tax administration of Finland through its chatbot “Virtanen”. Similar services are also offered by the tax administrations of Germany, Latvia, France, Spain, Estonia, Sweden and Slovakia, primarily in support of mandatory tax filings (Directorate-General for Structural Reform Support, 2023). In some administrations, chatbot responses are primarily directed at corporate entities, while in others they address individual taxpayers. A commonly noted limitation concerns language coverage, which is often confined to the official national language. Although multilingual support remains relatively limited, its gradual expansion may offer an advantage over traditional forms of service delivered exclusively by administrative staff.

Figure 3 shows that only France demonstrates a highly integrated interactive system in which AI manages communication and user interactions with minimal need for manual intervention. This assessment is supported by a report of the Direction générale des Finances publiques (2025), which indicates that the tax administration ranks among the most positively evaluated public services with approximately 82% of citizens expressing support.

In Estonia, Italy, Spain and Finland, systems supporting communication through artificial intelligence are well developed and used across the principal channels of interaction. The largest group of EU tax administrations (nine in total), including Malta, Slovakia, Latvia, Poland, Denmark, Austria, Lithuania, Ireland and Sweden, demonstrates a medium level of integration. In these countries, administrative personnel remain responsible for additional verification and validation of results. Nevertheless, the composite TAI indicates comparatively higher levels of digital development in Poland, Austria, Denmark, Lithuania and Ireland compared with Malta, Slovakia and Latvia. In Ireland the tax administration employs an instrument that assists in categorizing taxpayer inquiries, reducing the time required for routing requests to the appropriate administrative units and thereby shortening response times (OECD, 2025b). In Romania, Portugal, Croatia, the Netherlands, Germany, Greece and Bulgaria initial stages of process automation are observed in the examined functional area. By contrast, the tax administrations of Luxembourg, Belgium, Cyprus, the Czech Republic, Slovenia and Hungary make limited use of artificial intelligence in taxpayer communication, with service provision primarily dependent on human administration. Although OECD data (current to October 2025) do not record a virtual assistant service in Romania, a pilot system named “ANA” has been introduced as of early 2026. The service initially provides information concerning individual tax obligations, with plans for broader availability in subsequent phases.

Another important characteristic of the instruments relying on artificial intelligence that are employed within EU tax administrations is their capacity to analyse selected information by combining multiple sources and extracting data that are relevant for tax administration purposes. In practice, these systems draw upon a wide spectrum of sources, including external registers (such as publicly accessible databases, public and commercial registers and banking records) as well as internally generated datasets, most frequently derived from submitted tax declarations. They may also incorporate information produced through the application of other analytical instruments based on artificial intelligence. Through the combination of diverse sources of information, such technological solutions significantly accelerate the processing of large volumes of data, enhance overall data quality and reinforce the supervisory and control functions of tax administrations. Such instruments are frequently regarded as constituting a foundational infrastructure upon which several of the functional domains discussed above can be effectively developed and implemented.

Figure 4 illustrates the relationship between the comparative score achieved under TAIDS and the overall performance of individual tax administrations as measured by the composite TAI. The results indicate that, in contrast to the comparatively moderate use of artificial intelligence in communication and interaction with participants in the tax process, its application is considerably more advanced in the domain of data governance and management. The figure shows that in 7 of the 27 EU Member States (Finland, Lithuania, Spain, Denmark, the Netherlands, Estonia and Hungary) there exists a highly developed administrative environment in which artificial intelligence is systematically applied across procedures related to planning, forecasting and decision-making concerning the processing and use of data. OECD (2017) highlighted Finland’s progress in deploying Robotic Process Automation (RPA) for data processing, which was estimated to reduce the workload of these tasks by the equivalent of 52 person-years of effort. Nevertheless, it should be emphasised that in most of these countries comprehensive centralised standards and an integrated data governance framework remain insufficiently developed. This institutional gap may generate inconsistencies or errors and, as a consequence, reduce the overall effectiveness of systems relying on artificial intelligence. As previously highlighted in this study, ensuring an exceptionally high level of protection of processed data constitutes a major and continuing institutional challenge.

An analysis based on the indicators incorporated in TAIDS shows that 19 countries report the use of AI in two principal areas. These include the processing of extensive datasets to strengthen tax compliance and the conduct of network analysis. Seventeen countries report the use of machine learning techniques in data modelling and analytical procedures, while an equal number employ extensive datasets for the purpose of forecasting tax revenues. Approximately half of the analysed tax administrations state that they have introduced mechanisms enabling the detection and prevention of tax fraud in real time.

In the tax administrations of Bulgaria, Italy and the Czech Republic, artificial intelligence is likewise employed in the processing of extensive datasets, including in supporting managerial decision-making and in enhancing operational efficiency. Although comparable technological applications can be observed in Ireland, France, Austria, Greece, Sweden, Poland, Slovenia, Germany and Cyprus, the degree of integration in their tax administrations remains comparatively lower than in the previously identified group. Jiménez and Mółka (2025) also highlight the significant progress in the digitalization of the Polish tax administration, emphasizing the large volume of data collected through a well-established system of data exchange between banks and public authorities. These data are subsequently processed using algorithmic tools, enabling risk assessment and supporting more targeted and efficient tax control.

The analysis may be further complemented by data from TaxAdmin.AI (n.d.), indicating that Austria, Belgium, Spain, Denmark, Germany, Greece, Ireland, Luxembourg, the Netherlands, Portugal and France employ techniques such as web scraping to collect raw data that subsequently inform analytical systems relying on artificial intelligence. These data are obtained primarily from online sources, commercial platforms and social media environments. The tax administration of France has adopted an even more innovative approach, involving the analysis of satellite imagery of properties (or specific components thereof, such as swimming pools) and the systematic comparison of these observations with declared information. This method enables the identification of discrepancies and supports more accurate property assessments, thereby contributing to enhanced fiscal equity (Ay & Söylemez, 2025). The initial outcomes of this integration have proven substantial. Pilot implementations identified more than 20,000 undeclared swimming pools and at a subsequent stage this number exceeded 120,000. As a result, approximately EUR 40 million in additional public revenue was generated in 2023 (Direction générale des Finances publiques, 2024). One of the tools more frequently used for collecting publicly available information is XENON. Its purpose is to gather and analyse such information, compare it with declared data and identify inconsistencies, including undeclared income or assets. The software was first developed and implemented within the tax administration of the Netherlands more than two decades ago and was later adopted by other administrations, including Belgium, Austria, Sweden and Denmark.

In the methodological section above, it was noted that the TAI Rule Enforcement Subindex covers indicators measuring processes related to the core regulatory and fiscal functions of tax administrations. The indicators included in TAIRES examine various potential applications of AI in tax administration. These cover automated information provision to taxpayers, responses to queries concerning tax legislation, risk assessment and profiling, dispute resolution and monitoring of system anomalies. The results of the comparative assessment indicate that most tax administrations do not yet employ artificial intelligence in the processes described above. This contributes to the comparatively lower values of the composite index. For two of the ten indicators comprising TAIRES no EU Member State has declared the use of artificial intelligence. One indicator concerns the adoption of final administrative decisions and the other concerns the provision of recommendations for dispute resolution between tax administrations and taxpayers. Final decisions at this stage are taken and validated by administrative staff. This confirms the previously expressed observation by Richmond et al. (2023) and others, that the limited interpretative capacity of AI systems in administrative decision-making remains a critical concern, particularly when AI is used to support complex judgment-based processes. In the analysis of challenges related to the integration of artificial intelligence in tax administrations, the example of risk modelling in the Netherlands was discussed, illustrating that inappropriate technological solutions may generate negative social consequences.

Within the TAIRES subindex, the largest number of EU tax administrations (13) declared the use of artificial intelligence for risk profiling, risk assessment and fraud detection. These areas of application contribute to administrative efficiency and cost reduction. The assessment process is typically based on patterns derived from historical data and relational analyses of financial transactions (OECD, 2025a).

Risk assessment may be structured along several dimensions. These include the risk of non-compliance with tax legislation (according to TaxAdmin.AI (n.d.), this approach is observed in Lithuania, Denmark, Ireland and Slovenia), the risk of inconsistencies in declared information (Poland) and ex ante assessments of the probability that an economic agent will enter insolvency or bankruptcy proceedings (Belgium). However, the OECD database on the Belgian tax administration does not explicitly indicate the use of AI in such risk assessment procedures. The Estonian tax authorities provide an electronic service enabling taxpayers to check their tax behaviour rating, which consists of two components: a tax compliance rating and a tax behaviour adequacy rating.

On the other hand, available data show that AI is a significant instrument in combating tax fraud with a view to enhancing revenue collection (Greece, Hungary, Poland, Spain). In these cases, its application (primarily through machine learning techniques) focuses on modelling transactions and analysing company and household profiles in order to detect anomalies. In Sweden and Poland, the examination of potential tax fraud is further reinforced by network analysis tools designed to identify systemic interdependencies among related taxpayers and recurrent behavioural patterns indicative of coordinated fraudulent schemes. A defining feature of these processes is their predominant focus on combating non-compliance with Value-added tax (VAT) legislation, with certain administrations (e.g., Austria) further integrating them into broader risk assessment models. The Federal Ministry of Finance of Austria (Bundesministerium für Finanzen, 2025) reports that AI-supported analyses conducted by the Ministry’s Predictive Analytics Competence Centre (PACC) covered millions of non-compliance cases, generating additional tax revenues exceeding EUR 350 million through targeted audits. A significant limitation, however, lies in the need to process extensive data volumes, which poses substantial operational challenges and entails considerable financial and infrastructural investment.

Although the tax administration of Finland achieves the highest normalised TAIRES score (Figure 5), AI integration is identified in only six of the ten indicators comprising the subindex. This indicates that measures related to AI application in tax enforcement are present, but at a relatively moderate level of implementation. The methodology is comparative and focuses on relative performance across Member States. This provides a reasonable basis for concluding that the Finnish tax administration exhibits the highest degree of AI integration in tax enforcement within the examined sample.

Alongside Finland, algorithms in the tax administrations of Spain, Italy, Sweden and France support strategic planning, monitoring and control functions, complementing human oversight. This is also confirmed in a report by Direction générale des Finances publiques (2025), which states that in 2024 AI-based tools were used in 56% of audit procedures concerning business taxpayers in France. As a result of these audits, the administration collected approximately EUR 1 billion in additional revenue generated through tax audits compared with the preceding period.

For numerous tax administrations, particularly in Central and Eastern Europe (Lithuania, Estonia, Hungary, Greece, Bulgaria, Latvia and Croatia), the TAIRES subindex score remains below the composite TAI score. By contrast, several Member States (the Netherlands, Belgium, Luxembourg, Romania, Slovakia, Cyprus, Malta and the Czech Republic) record a TAIRES score of zero, indicating that no indicators within this subindex provide documented evidence of AI use in tax rule enforcement and compliance.

Importantly, these zero values should not be interpreted as an absence of AI deployment. Rather, they likely reflect a lack of publicly available information, sometimes referred to as a documentation bias. This interpretation is consistent with OECD evidence on the use of advanced analytics in compliance risk management (OECD, 2016) and the importance of trust-based tax administration (OECD, 2023). Limited disclosure may therefore be intended to preserve enforcement effectiveness and prevent behavioural adaptation by taxpayers.

The zero values of the TAIRES subindex may also relate to structural characteristics of national tax systems. For instance, countries such as the Netherlands, Belgium and Luxembourg are characterised by high institutional trust and tax morale, which have been associated with greater voluntary compliance (Alm & Torgler, 2006). In such contexts, reliance on algorithmic enforcement may be lower.

At the same time, the group of countries with zero TAIRES scores shows heterogeneity in administrative capacity. Luxembourg exhibits a relatively centralised model and higher human resource allocation per capita, whereas Romania, Malta, Cyprus, Slovakia and the Czech Republic operate with more limited resources (Pîrvu, 2021). Evidence from Mergel et al. (2019) and Dener et al. (2021) supports the link between organisational capacity, digital skills and the pace of digital transformation. Similarly, findings from the European Commission Joint Research Centre highlight data availability and scale limitations, which are particularly relevant for smaller jurisdictions such as Malta, Cyprus and Luxembourg (Van Noordt et al., 2020).

Legal and regulatory frameworks provide an additional explanatory dimension. The General Data Protection Regulation limits fully automated decision-making in cases producing legal or similarly significant effects (European Union, 2016). The EU AI Act imposes obligations on high-risk AI systems, including applications in public administration that affect individuals’ rights or obligations, such as in risk assessment or compliance monitoring (European Union, 2024). It should be noted, however, that not all AI applications in tax administrations fall within the high-risk category, as classification depends on specific use cases and potential impact.

Overall, the results suggest that the zero values observed for the TAIRES subindex should be interpreted with caution. Rather than indicating the absence of AI-driven enforcement practices, they appear to reflect a combination of limited public disclosure, regulatory constraints, differences in administrative capacity and structural factors such as country size and data availability. In this sense, the empirical findings of the present study are consistent with existing evidence on the determinants of AI adoption in the public sector.

The assessment of control, accountability and transparency in the use of AI within tax administrations represents a central analytical dimension (Yordanova, 2024). As highlighted in the literature review, these issues are widely recognised as significant challenges for public administrations, including tax authorities. In the present study, the assessment is conducted through indicators capturing the development and implementation of ethical standards and norms within administrative structures, as well as mechanisms of external oversight concerning digital transformation processes.

Limitations on AI deployment (legal, ethical or technical) constitute an additional key aspect of the analysis. Internal control mechanisms primarily involve verification of algorithmic design and output consistency, including measures to identify potential biases and ensure alignment with administrative objectives. External oversight mechanisms, by contrast, enable independent actors outside the administrative system to evaluate these processes, thereby enhancing transparency and institutional accountability. The objective of such controls is to ensure that tax enforcement is supported by artificial intelligence without generating risks for the tax system or individual taxpayers. Seventeen Member States of the EU report internal controls concerning data utilised for AI training and analytical purposes, typically aimed at mitigating algorithmic bias or analytical incompleteness. In sixteen Member States, restrictions on the applicability of AI have been identified.

Regarding regulated possibilities for external oversight, only a small number of countries (three in total—Ireland, Lithuania and Austria) report the existence of such an option in their tax systems. Figure 6 consequently shows Ireland achieving the highest comparative score in this functional dimension. This result reflects the presence of structured oversight mechanisms aimed at transparency and process supervision. In particular, Ireland exhibits the most extensive framework for external monitoring (with the exception of AI system code). This aligns with its role as a major technology hub within the EU and the development of regulatory capacities relevant to oversight.

• Internal consistency of the subindices

To further evaluate the robustness of the constructed subindices, their internal consistency is examined. Following the methodology outlined in Equation (3), internal consistency is assessed using the Kuder–Richardson Formula 20 (KR-20). Each subindex is analysed separately. KR-20 values range from 0 to 1, with higher values indicating stronger internal consistency among the indicators comprising each subindex.

The robustness check for the subindices (

Table 6. Internal Consistency (KR-20) of Subindices.

Subindex	KR-20
TAIIS	0.406681
TAIDS	0.842222
TAIRES	0.720049
TAIGS	0.849946

) shows that three (TAIDS, TAIRES and TAIGS) of the four subindices exhibit high internal consistency (KR-20 > 0.7), indicating that their indicators are well-aligned. The TAIIS subindex, however, displays a lower KR-20 value (0.406681), reflecting weaker coherence among its components. This lower consistency should not necessarily be seen as a weakness, as TAIIS captures diverse aspects of AI use in communication between tax administrations and economic agents, spanning different functional dimensions such as technological features, applications and institutional arrangements. It may also reflect an early stage of AI adoption, where communication-oriented applications constitute the most accessible entry point. More advanced uses, such as AI in tax law enforcement and administrative decision-making (TAIRES) and AI governance frameworks (TAIGS), often remain largely undeveloped or absent at this stage (which is why several EU countries register zero values for these subindices).

The implications of this lower internal consistency for the overall composite index are further explored through robustness analysis.

• Robustness assessment of the composite TAI
• Contribution of the TAIIS subindex to the composite TAI

The role of the TAIIS subindex is assessed by recalculating the composite TAI both with and without it, allowing evaluation of its influence on country rankings. Before this, correlations between the individual subindices and the composite index are presented using the correlation coefficient (

Table 7. Correlation matrix between the composite TAI and the subindices.

	TAIIS	TAIDS	TAIRES	TAIGS	TAI
TAIIS	1
TAIDS	0.326737	1
TAIRES	0.636804	0.50573	1
TAIGS	0.571871	0.561908	0.802658	1
TAI	0.705538	0.728384	0.9282	0.903028	1

) to illustrate the strength of these relationships.

Despite its lower internal consistency, the TAIIS subindex exhibits a relatively high correlation with the overall index (r = 0.705538), indicating that it contains meaningful information on the level of AI integration. We further compare the composite TAI calculated with and without TAIIS. When excluded, its 15% weight is redistributed evenly among the remaining subindices, resulting in weights of 30% for TAIDS and TAIGS and 40% for TAIRES.

The recalculated TAI shows deviations across countries within the range of −0.1 to +0.1, most notably in France and Hungary due to differences in TAIIS scores. The correlation between the TAI with three subindices and the full TAI with four subindices is 0.987. These results confirm that removing TAIIS does not substantially alter the overall rankings, demonstrating the robustness of the composite index. Thus, the lower KR-20 value for TAIIS reflects its multidimensional nature rather than a methodological issue and does not compromise the validity of the overall index. TAIIS also provides an additional perspective on the broader applicability of AI in tax administrations.

• Sensitivity of the composite TAI to subindex weighting

To further assess the robustness of the composite TAI, alternative weighting schemes for the four subindices were applied. Specifically, the TAI is assessed by comparing its value with a hypothetical equal-weighting scenario (w_i = 25% for each subindex). An alternative robustness check increases the weight of the TAI Rule Enforcement Subindex (TAIRES) from 35% to 40%, while preserving the relative shares of TAIDS and TAIGS and reducing the communication-oriented subindex (TAIIS) from 15% to 10%. The comparative results for the three scenarios are presented in

Table 8. Robustness assessment of the composite TAI.

Country	WTAIIS = 25% WTAIDS = 25% WTAIRES = 25% WTAIGS = 25%	Rank	WTAIIS = 10% WTAIDS = 25% WTAIRES = 40% WTAIGS = 25%	Rank	WTAIIS = 15% WTAIDS = 25% WTAIRES = 35% WTAIGS = 25%	Rank
Finland	0.85	1	0.89	1	0.88	1
Spain	0.79	2	0.8	2	0.8	2
Italy	0.73	3	0.74	3	0.74	3
Ireland	0.69	5	0.71	4	0.7	4 *
France	0.72	4	0.69	6	0.7	4 *
Sweden	0.65	7	0.7	5	0.69	6
Lithuania	0.67	6	0.67	7	0.67	7
Denmark	0.64	8	0.66	8 *	0.66	8
Austria	0.63	9	0.66	8 *	0.65	9
Poland	0.53	11	0.55	10	0.54	10
Estonia	0.58	10	0.49	11	0.52	11
Hungary	0.42	12	0.47	12	0.46	12
Bulgaria	0.4	13	0.39	14 *	0.39	13
Greece	0.38	14	0.39	14 *	0.38	14 *
Slovenia	0.33	16	0.41	13	0.38	14 *
Germany	0.34	15	0.38	16	0.37	16
Latvia	0.31	17	0.26	17	0.27	17
Netherlands	0.27	18	0.23	18	0.25	18
Croatia	0.24	19	0.22	19	0.23	19
Portugal	0.19	20 *	0.21	20	0.2	20
Czech Republic	0.17	22	0.17	21	0.17	21
Slovakia	0.19	20 *	0.11	22	0.14	22
Cyprus	0.1	24	0.1	23	0.1	23 *
Malta	0.15	23	0.07	24	0.1	23 *
Belgium	0.06	25 *	0.06	25	0.06	25
Luxembourg	0.04	27	0.04	26	0.04	26 *
Romania	0.06	25 *	0.03	27	0.04	26 *

Note: Countries marked with * share the same position due to identical index values.

. This analysis complements the previous examination of the TAIIS subindex, providing additional evidence of the stability and reliability of the composite measure.

The data in Table 8 show a high degree of consistency across the three weighting approaches. Under all approaches, Finland, Spain and Italy occupy the top positions. Differences in ranking for certain countries are not associated with substantial differences in index scores, which suggests that variations in relative position do not necessarily reflect meaningful performance gaps.

5. Conclusions

The results of the present study confirm substantial cross-country heterogeneity in AI utilisation within EU tax administrations, with notably higher levels observed in the Scandinavian countries. In most Member States, AI implementation remains at an experimental or pilot stage, primarily aimed at enhancing taxpayer communication, improving service delivery and increasing administrative efficiency. This pattern is reflected in the zero values of the TAIRES and TAIGS subindices for several countries, as these indicators are typically associated with more advanced stages of AI integration. By contrast, a number of Member States have initiated these processes earlier, a difference clearly captured by the comparative quantitative indicators.

A joint interpretation of the results suggests that the level of AI adoption does not always correspond to the overall level of digital development in the public sector. While some countries display strong alignment between the two dimensions, others show notable differences. These patterns indicate that broader digital capacity may be associated with AI adoption, but it does not fully explain the observed variation. At the same time, the findings also reveal cases of clear consistency between the two indicators. Finland represents a notable example, combining the highest relative TAI score with a high level of digital maturity in the public sector. Similar alignment is observed at the lower end of the distribution. Cyprus, Romania and Slovakia display low levels of AI adoption in tax administrations, which correspond to lower levels of digital maturity. Comparable patterns, although less pronounced, can also be identified in Bulgaria, Greece, Germany and Croatia.

Although the use of AI offers considerable potential benefits, it also entails a multidimensional risk profile, including technological vulnerabilities, capacity and expertise constraints, and legal and regulatory challenges. In this regard, it is essential to define a coherent roadmap for structured integration, while simultaneously assessing the expected benefits against the potential economic, institutional and compliance costs associated with their deployment.

Such a roadmap should be developed across several dimensions. First, clearly defined strategic priorities are required to guide the integration of AI within tax administrations. Second, measures adopted at the institutional level should (where feasible) be aligned with broader public sector digitalisation efforts. Such alignment may enhance institutional coherence, improve interoperability between systems and lead to more effective data exchange, thereby contributing to the development of an integrated data governance framework. A coordinated approach may also reduce the risk associated with fragmented digital architectures, duplication of investments and technological incompatibilities. At the same time, it can support the adoption of common standards in cybersecurity, data protection, and algorithmic transparency and accountability frameworks.

In the broader European context, such coordination should be aligned with the EU regulatory framework, including the requirements of the Artificial Intelligence Act. Such alignment may strengthen governance mechanisms for high-risk AI systems and contribute to increased institutional trust, both among public authorities and between the state and taxpayers. Most broadly, coordinated implementation may enhance compliance and service delivery, reduce administrative burdens and support more efficient public resource management.

The present study provides a detailed and structured comparative assessment of AI adoption in the EU tax administration. It combines a quantitative measurement framework with a comparative perspective, highlighting both the progress achieved and the challenges that remain, particularly in relation to governance, transparency and institutional capacity.

While the results provide a detailed comparative perspective on AI adoption in EU tax administrations, they should be interpreted with caution. Differences in reporting practices and levels of digital development across countries may influence the observed patterns, while the use of proxy indicators limits the precision of the measurement. Moreover, the comparative design identifies associations rather than causal relationships and the findings should not be interpreted as evidence of direct effects. Finally, the rapidly evolving nature of AI technologies implies that the results may become outdated over time.

Regardless of the analytical depth of the study, the subject matter remains open to further elaboration through the incorporation of additional analytical dimensions. Among these, particular attention should be devoted to the behavioural implications of AI deployment in tax processes, especially its influence on taxpayers’ perceptions of impartiality, fairness and decision-making transparency. While algorithmic selection mechanisms may reinforce perceptions of objectivity, they may also generate heightened institutional distrust in the absence of robust explain ability safeguards and accountability mechanisms.

A further research priority concerns the systematic economic appraisal of public sector investments in AI applications, explicitly grounded in Cost–Benefit analysis (CBA), with particular emphasis on their net social returns, fiscal sustainability and distributional implications. Such an approach would contribute to evidence-based policy decisions concerning the appropriate scale, sequencing and strategic orientation of future digital transformation initiatives.

In addition, future research could further deepen the analysis of institutional and structural factors affecting AI adoption in tax administrations. The present study provides initial insights into the importance of country-level characteristics. However, these factors could be examined in greater detail in future work. A more detailed analysis of national tax system features and administrative frameworks may help explain the observed variation in TAI scores across countries.

Finally, the analytical framework in this study could be extended beyond the EU context. It may also be applied to non-member countries. This would enable a broader international comparative assessment of AI adoption in tax administrations. Such an extension would allow the identification of global patterns and institutional differences. It would also support the identification of best practices in digital transformation. Overall, this would strengthen the empirical basis for understanding the diffusion of AI in public sector governance.