Algorithmic Tax Justice in Peru

Abstract

Peru’s tax dispute system—administrative claim, Tax Court appeal, and contentious-administrative review—has increasingly migrated toward electronic files, e-invoicing, interoperable databases, and data-driven oversight. This article examines whether artificial intelligence can reduce avoidable tax litigation without weakening taxpayers’ rights and identifies the institutional conditions required to reconcile administrative efficiency with due process, reason-giving, and effective contestation. Using a legal-doctrinal and policy-analytical design, the study analyzes Peru’s tax dispute architecture, digital evidence environment, and AI-related risks in compliance and administrative litigation. The findings show that only bounded decision-support applications are institutionally appropriate, including audit triage, anomaly detection, document classification, workflow prioritization, compliance assistance, and consistency checks, provided that they do not replace legally attributable human judgment. AI is compatible with digital tax justice only when six safeguards are institutionalized: legally meaningful explainability, evidentiary and computational traceability, meaningful human oversight with override authority, lifecycle auditability, effective contestation, and distributional equality. The analysis further demonstrates that facially neutral digital requirements and risk models may generate unequal effects when disparities in connectivity, digital literacy, record-keeping capacity, and access to professional assistance translate into differences in audit exposure, compliance costs, evidentiary burdens, and practical contestability. The article proposes a rights-compatible framework for AI-supported tax enforcement in Peru.

1. Introduction

Tax administrations across the world are undergoing a structural digital transformation. Electronic invoicing, prefilled returns, interoperable databases, automated risk analysis, and digital payment traces have expanded the informational capacity of tax authorities and changed the evidentiary environment in which compliance decisions are made. The recent literature shows that digital tax administration is no longer limited to electronic filing; it increasingly involves data integration, automated monitoring, and analytics-based compliance strategies that reshape the relationship between taxpayers and the state (Bassey et al. 2022; Hesami et al. 2024).

This transformation is particularly relevant in developing and emerging economies, where tax administrations face persistent constraints related to informality, limited enforcement capacity, and revenue mobilization needs. Evidence from Ethiopia shows that electronic sales registration machines can increase reported income tax and VAT revenue, although taxpayer responses may also shift through reported costs (Mascagni et al. 2021). Research on Rwanda suggests that e-invoicing can improve net VAT payments and audit efficiency, especially when digitalization is linked to enforcement capacity rather than treated as a purely technological reform (Kotsogiannis et al. 2025). In China, stronger VAT enforcement has been associated with spillover effects on payroll-tax evasion, showing that digital and enforcement reforms may redistribute compliance behavior across tax bases (Li et al. 2021). Evidence from Uruguay also indicates that digital payment incentives can increase card transactions without necessarily improving tax compliance, which cautions against assuming that transaction digitalization automatically produces compliance gains (Brockmeyer and Sáenz Somarriba 2025). Together, these studies suggest that digital tax tools can improve administrative capacity, but their effects are heterogeneous and institutionally conditioned.

Latin America offers a particularly important setting for this debate because the region has been at the forefront of electronic invoicing and data-driven VAT control. Ecuadorian evidence on ghost firms and fraudulent invoices shows that transaction-level tax data can help detect sophisticated evasion schemes, but it also illustrates the need for careful targeting and procedural safeguards when enforcement relies on digital traces (Carrillo et al. 2023). At the same time, broader evidence on fiscal redistribution in Latin America indicates that tax and transfer systems have differential effects across groups and development dimensions, reinforcing the need to analyze digital enforcement not only from the perspective of revenue, but also from the perspective of fairness and distributional impact (Claveria 2025).

Peru’s current digital tax administration is the result of a gradual institutional transition from conventional documentary control toward electronic reporting, digital records, electronic notification, and large-scale transactional data processing. This trajectory has been especially visible in the expansion of SUNAT’s electronic invoicing ecosystem, which transformed VAT control from ex post documentary verification into a data-intensive environment based on transaction-level information (Bellon et al. 2022, 2023). In parallel, Peru’s digital government framework, electronic document rules, and tax procedural regulations have progressively strengthened the legal basis for electronic records, digital notifications, interoperability, and digitally supported administrative procedures (Ministry of Economy and Finance (Perú) 2013; Presidency of the Council of Ministers (Perú) 2008, 2018, 2021). Therefore, AI would not enter a paper-based tax administration, but rather an already digitized evidentiary infrastructure in which electronic invoices, digital records, and interoperable systems increasingly shape compliance monitoring and dispute formation.

Peru is an analytically relevant case for three reasons. First, the country has developed a mature electronic invoicing infrastructure that has already generated measurable compliance effects. Bellon et al. (2022) show that the transition from paper to VAT e-invoicing in Peru increased reported firm sales, purchases, and VAT liabilities, particularly among smaller firms and sectors with higher non-compliance. Second, subsequent evidence indicates that e-invoicing reforms also generate compliance spillovers through firm-to-firm transactional networks, confirming that Peru’s digital tax infrastructure affects not only individual taxpayers but also broader reporting ecosystems (Bellon et al. 2023). Third, Peru’s tax dispute system combines administrative claims, Tax Court appeals, and subsequent judicial review, creating a setting in which digital evidence, automated triage, and AI-supported risk assessment may influence both enforcement and contestation.

Against this background, artificial intelligence introduces a specific governance dilemma. On the one hand, AI may help SUNAT and dispute-resolution bodies prioritize cases, detect anomalies, classify documents, identify inconsistent filings, and reduce avoidable litigation. On the other hand, AI may intensify opacity, information asymmetries, evidentiary burdens, discriminatory targeting, and weak contestability if algorithmic outputs become operationally decisive without being legally explainable or reviewable. The problem is not whether technology can make tax enforcement faster, but whether speed can be reconciled with reason-giving, due process, and effective defense rights.

This article therefore asks two research questions: (i) which uses of AI are institutionally appropriate to reduce tax litigation without eroding procedural safeguards? and (ii) which minimum conditions are required to reconcile administrative efficiency with due process, reason-giving, and effective contestation? The article argues that AI can contribute to dispute prevention and faster resolution only when it is confined to auditable decision-support functions and embedded within a governance framework based on legally meaningful explainability, traceability, meaningful human oversight, lifecycle auditability, effective contestation, and distributional equality and accessibility.

The contribution of this article is threefold. First, it develops the concept of algorithmic tax justice as a normative framework that integrates distributive, procedural, corrective, and institutional dimensions of justice in AI-supported tax administration. Second, it connects Peru’s electronic tax infrastructure and administrative litigation pathway with recent international evidence on digital tax administration, VAT enforcement, and AI accountability. Third, it proposes a minimum safeguards package that is institutionally implementable in the Peruvian context and relevant to other jurisdictions where tax authorities are moving from digital records toward algorithmic oversight.

1.1. Conceptualizing Algorithmic Tax Justice

This article understands algorithmic tax justice as the set of normative and institutional conditions under which AI-supported tax administration can improve compliance and dispute management without undermining taxpayers’ rights. The concept is not reducible to technological efficiency, predictive accuracy, or revenue optimization. Rather, it refers to the extent to which algorithmic tools used in tax administration remain compatible with fairness, reason-giving, reviewability, non-discrimination, and effective contestation. In this sense, the concept builds on recent debates on tax algorithmic governance, taxpayers’ rights in AI-enabled tax administration, explainable AI in tax law, and procedural fairness in algorithmic decision-making (Decker et al. 2025; Faúndez-Ugalde et al. 2020; Górski et al. 2025; Hadwick 2022).

First, algorithmic tax justice has a distributive dimension. AI-supported enforcement must not produce systematically unequal exposure to audits, sanctions, procedural burdens, or compliance costs across categories of taxpayers. This concern is especially relevant in countries where compliance capacity is uneven: larger taxpayers may have sophisticated digital compliance infrastructures, whereas smaller taxpayers may face higher relative costs, lower technical capacity, and weaker ability to rebut algorithmic classifications. Evidence on digital tax reforms shows that e-invoicing and digital enforcement may improve compliance, but their effects vary according to firm size, sector, audit exposure, transaction networks, and institutional design (Bellon et al. 2022, 2023; Kotsogiannis et al. 2025; Mascagni et al. 2021). Therefore, algorithmic tax justice requires attention not only to aggregate compliance gains, but also to the distribution of burdens and benefits across taxpayers. Importantly, facial neutrality should not be conflated with distributive neutrality. A uniform digital obligation, an identical response period, or a common risk-scoring model may impose substantially different real costs depending on a taxpayer’s connectivity, digital literacy, record-keeping infrastructure, access to professional advice, and ability to contest data-driven inferences. The relevant equality inquiry therefore concerns not only whether the same formal rule applies to all taxpayers, but also whether that rule predictably transfers greater learning, compliance, and psychological costs to those with fewer administrative and technological resources. From this perspective, unequal impact may arise without any explicit discriminatory classification because apparently neutral institutional arrangements interact with pre-existing differences in capacity and access (Chudnovsky and Peeters 2021; Moynihan et al. 2015).

Second, algorithmic tax justice has a procedural dimension. Taxpayers must be able to know, understand, and challenge the reasons that justify administrative action. In this sense, explainability is not merely a technical preference; it is a legal condition for reason-giving, defense, and reviewability. An AI-supported risk score cannot replace the duty to articulate legally relevant reasons. Recent scholarship on explainable AI in tax law argues that XAI should not be understood as a merely technical feature, but as a minimum legal standard that enables taxpayers and reviewing authorities to understand how algorithmic outputs are produced and how they affect tax decisions (Kuźniacki et al. 2022). This position is reinforced by more recent work showing that explanations in the tax domain must be legally meaningful, context-sensitive, and usable for contestation rather than limited to technical transparency (Górski et al. 2025). Likewise, procedural fairness in algorithmic decision-making requires that affected persons have genuine opportunities to understand, participate in, and challenge decisions that affect them (Decker et al. 2025).

Third, algorithmic tax justice has a corrective dimension. If an AI-supported process contributes to an erroneous selection, an unjustified assessment, or an excessive evidentiary burden, the taxpayer must have access to effective mechanisms for review, correction, and remedy. Contestation must therefore include access to relevant data, model-related explanations, and meaningful human reassessment. This is particularly important because AI may deepen information asymmetries between tax authorities and taxpayers when the administration controls the data infrastructure, the model, and the institutional interpretation of risk (Hadwick 2022; Guglyuvatyy 2025). Without corrective mechanisms, algorithmic tax administration may increase procedural vulnerability rather than strengthen digital tax justice.

Fourth, algorithmic tax justice has an institutional dimension. AI systems must be embedded in accountable governance arrangements that define who designs, validates, deploys, audits, supervises, and reviews them. In the Peruvian context, this requires coordination among SUNAT, the Ministry of Economy and Finance, the Tax Court, the judiciary, and the data protection authority. The recent public-sector AI governance literature emphasizes that trustworthy automated decision-making depends on institutional requirements such as auditability, documentation, human oversight, risk management, and lifecycle monitoring (Agbabiaka et al. 2025; Green 2022; Koshiyama et al. 2024; Laine et al. 2024). Accordingly, algorithmic tax justice is best understood as a rights-compatible governance framework for digital tax administration, rather than as a claim that algorithmic tools are inherently fair.

1.2. Peruvian Legal Framework for Algorithmic Tax Justice

Peru’s architecture of tax dispute resolution is built on a sequential pathway that combines administrative adjudication with subsequent judicial review. At the constitutional level, contentious-administrative review functions as the mechanism through which “resoluciones administrativas que causan estado” are challenged before the Judiciary, and the statutory framework emphasizes that the purpose of this action is the judicial control of administrative conduct and the effective protection of rights and legally protected interests.

Within this framework, tax controversy is structured by the Tax Code (Ministry of Economy and Finance (Perú) 2013), which regulates the contentious-tax procedure and establishes the Tax Court (Tribunal Fiscal) as the body that resolves tax claims in the last administrative instance. The Code also regulates key procedural guarantees that become particularly relevant in digital contexts: (i) notification rules (including electronic notifications), (ii) admissibility and filing requirements for appeals, and (iii) the administrative stages that must be exhausted before judicial review.

General administrative legality and due process guarantees are provided by the General Administrative Procedure Act (Ministerio de Justicia y Derechos Humanos (Perú) 2019). This statute is especially salient for AI-enabled administration because it embeds the duty to motivate administrative acts and, crucially, requires that when administrative acts are produced through automated systems, the administration must ensure that the affected party can identify the name and position of the authority issuing the act—an explicit legal anchor for accountability in semi-automated decision pipelines.

Digital transformation in the public sector is further shaped by the Digital Government Law (Presidency of the Council of Ministers (Perú) 2018) and its regulation (Presidency of the Council of Ministers (Perú) 2021), which establish governance principles for digital identity, interoperability, security, and data management, and provide a legal basis for the transversal use of digital technologies and electronic means in administrative procedures.

Finally, the evidentiary and rights environment of digital tax justice is reinforced by complementary regimes. The legal validity of digital signatures and digitally signed electronic documents is recognized under Law No. 27269 and its regulation (Presidency of the Council of Ministers (Perú) 2008), including their admissibility as evidence in judicial and administrative proceedings under specified conditions, an essential element when disputes hinge on electronic invoices, records, and audit trails. The processing of taxpayer data within AI-enabled tax oversight is constrained by the Personal Data Protection Act (Congreso de la República (Perú) 2011) and its updated regulation (Autoridad Nacional de Protección de Datos Personales (Perú) 2024), which foreground purpose limitation, proportionality, and safeguards for rights-holders in data processing operations. Appendix A presents four concrete applications in the Peruvian tax dispute pathway.

2. Methodology

This article adopts a legal-doctrinal and policy-analytical design. Its purpose is not to measure the empirical performance of a specific AI system used by Peruvian tax authorities, but to determine the legal and institutional conditions under which AI-supported tools may be compatible with digital tax justice. The study was conducted between February and March 2026, and the literature review was updated through April 2026.

2.1. Corpus and Data Sources

The primary corpus consisted of Peruvian legal materials governing tax controversy, administrative procedure, digital government, electronic evidence, data protection, and judicial review. These materials were selected because they define the procedural and evidentiary environment in which AI-supported tax tools would operate. The secondary corpus consisted of the peer-reviewed literature on digital tax administration, AI in tax enforcement, explainable AI, algorithmic accountability, e-invoicing, VAT compliance, BEPS/Pillar Two, and procedural fairness. The secondary literature was selected using relevance, recency, and disciplinary proximity criteria, with preference given to publications from the last five to seven years and to studies published in indexed journals with verifiable DOI.

2.2. Methodological Positioning

The study should be understood primarily as a Peruvian doctrinal and institutional analysis, supplemented by functionally selected international insights. It is not presented as a full comparative law study. The international evidence is used for two limited purposes: first, to contextualize Peru within broader digital tax administration trends; and second, to identify risks and safeguards that are functionally comparable across jurisdictions using e-invoicing, VAT analytics, digital payments, or AI-supported tax compliance tools. This reformulation avoids treating the comparative component as systematic when the article’s main analytical object is the Peruvian tax dispute pathway.

2.3. Analytical Procedure

The analysis proceeded in five steps. First, a functional map of Peru’s tax dispute architecture was constructed, identifying the administrative claim, Tax Court appeal, and contentious-administrative judicial review stages. Second, the study identified the digital evidence environment in which tax disputes increasingly arise, including electronic invoices, electronic records, administrative notifications, and digital case files. Third, AI-related risks were classified into six categories: opacity, bias, information asymmetry, evidentiary burden shifting, traceability deficits, and unequal distribution of technological access and compliance burdens. Fourth, the article derived safeguards by translating general AI-governance principles—explainability, traceability, meaningful human oversight, lifecycle auditability, contestability, and distributional equality and accessibility—into legally operative requirements for tax administration and review. Fifth, the proposed safeguards were tested for institutional implementability by asking which Peruvian institution would be responsible for each safeguard and at which procedural stage it should operate.

The analytical procedure was designed to ensure doctrinal coherence, functional mapping, and normative consistency between the Peruvian legal framework, AI-related risks, and the proposed safeguards.

2.4. Scope and Limitations

The study does not evaluate proprietary systems, internal SUNAT models, or confidential tax authority algorithms. Its contribution is normative and institutional: it identifies the minimum conditions under which AI-supported tax enforcement and dispute management would remain compatible with due process, reason-giving, and effective contestation. Therefore, the analysis should be read as a safeguards framework for AI-enabled tax administration, not as an empirical audit of current AI deployment.

To improve transparency in the structure of the argument,

Table 1. Relationship between research questions and manuscript sections.

Research Question	Main Sections Addressing the Question	Function in the Argument
RQ1. Which AI uses are institutionally appropriate to reduce litigation without eroding safeguards?	Section 3, Section 4 and Section 6, and Appendix A	Identifies acceptable AI uses as bounded decision-support functions: audit triage, anomaly detection, compliance assistance, document classification, workflow prioritization, and consistency checks.
RQ2. Which minimum conditions reconcile efficiency with due process, reason-giving, and contestation?	Section 5, Section 7 and Section 8	Derives safeguards for legally meaningful explainability, traceability, meaningful human oversight, lifecycle auditability, effective contestation, and distributional equality and accessibility.
Integrative answer	Conclusion	Establishes that AI can reduce avoidable litigation only conditionally, when embedded in a verifiable rights-compatible governance framework.

shows how the sections of the article respond to the two research questions.

3. Artificial Intelligence in Tax Administration and Compliance

Artificial intelligence (AI) refers to a set of computational techniques—particularly machine learning, natural language processing, and predictive analytics—that enable the automated classification of information, pattern detection, and decision support in high-volume administrative environments. In taxation, these capabilities are attractive because tax administrations process heterogeneous and massive datasets (e.g., returns, invoices, third-party reporting, accounting records), making AI a plausible instrument to improve timeliness, consistency, and targeting in oversight and service functions (Salah and Awwad 2024).

In the Peruvian case, it is important to distinguish between existing digital tax infrastructure and fully implemented AI-based decision-making. Peru has consolidated electronic invoicing, digital records, electronic notifications, and interoperable administrative systems that create the data environment in which AI-supported tools may operate. However, the present article does not claim that binding tax determinations are currently made by autonomous AI systems. Rather, it analyzes the legal conditions under which AI may be introduced or expanded as decision support in audit selection, anomaly detection, document triage, compliance assistance, and dispute management. This distinction is essential because the legal risk does not arise only when AI issues a final decision; it also arises when AI meaningfully influences which taxpayers are selected, which evidence is prioritized, and which explanations become available in the administrative file.

Recent empirical evidence from Peruvian customs agencies further indicates that regulatory digitization improves organizational performance primarily when it is accompanied by professional competency development. Digital transformation increased job proficiency both directly and indirectly through employees’ competencies, suggesting that effective human oversight requires not only formal decision authority but also sufficient institutional and technical capacity to interpret, question, and, where necessary, override technology-supported outputs (Chavez Diaz et al. 2026).

A first, well-documented contribution of AI is automation and operational efficiency. AI-based systems can streamline routine tasks such as responding to taxpayer inquiries, organizing and classifying documents within electronic files, and supporting internal workflows that reduce administrative workload and processing time (Akhila et al. 2024; Mehdiyev et al. 2021). Closely related, AI can strengthen audit selection and compliance monitoring through risk scoring and anomaly detection, helping administrations prioritize high-risk cases while avoiding unnecessary interventions (Shakil and Tasnia 2022; Vivian et al. 2023). In practice, such tools may also support continuous auditing environments and data-intensive oversight strategies, including in public-sector contexts where computer-assisted audit techniques are used to enhance monitoring (Gloria Robel et al. 2024).

A second line of application concerns predictive modeling for fraud and noncompliance. Predictive models—implemented through methods such as bagging, support vector machines, or other machine learning architectures—can estimate the probability of tax evasion or irregular behavior, which may improve audit effectiveness and reduce enforcement costs (Han 2022; Huang et al. 2022). Likewise, AI-supported detection models have been discussed as potentially improving tax fraud identification compared to purely traditional approaches (Radhi et al. 2024). However, these gains depend on data quality, model governance, and safeguards that prevent the operationalization of biased or legally opaque criteria.

Despite these opportunities, the use of AI in tax administration raises substantive legal and procedural concerns. The most pressing is explainability and transparency: if AI models are not interpretable, taxpayers and reviewing bodies may be unable to understand why an audit was triggered or why a certain risk assessment was produced, undermining the duty to provide reasons and the practical possibility of contestation (Kuzniacki et al. 2022; Kuźniacki et al. 2022; Mehdiyev et al. 2021). This is why explainable AI (XAI) should be treated as a minimum legal standard in tax administration when AI outputs influence audit selection, risk assessment, or other high-impact administrative actions (Kuźniacki et al. 2022). When such outputs affect taxpayers’ rights, this minimum standard also connects with constitutional and human-rights requirements of reason-giving, accountability, and effective review (Kuzniacki et al. 2022).

A second concern relates to taxpayers’ rights and non-discrimination. Studies focused on Latin America warn that AI can deepen information asymmetries and weaken defense rights unless administrations provide meaningful access to relevant information and ensure effective channels to challenge both the data and the logic underpinning AI-supported actions (Faúndez-Ugalde et al. 2020). Moreover, risks of discriminatory bias are salient in the inspection domain: if training data encode historical enforcement patterns, AI tools can reproduce unequal impacts, making bias control and procedural safeguards a core legal requirement (Martín López 2022; Palomino Guerrero 2022). At the governance level, debates on “tax algorithmic governance” emphasize that legality review must address not only outcomes but also the design of decision pipelines, documentation, and accountability mechanisms (Hadwick 2022).

Finally, AI expansion intersects privacy and regulatory limits. Because AI performance often relies on extensive data integration, it increases the need for proportionality, purpose limitation, secure access controls, and auditable governance. Recent developments in European tax law underline a growing trend toward restricting or conditioning tax authorities’ use of AI systems, reinforcing the general principle that innovation must remain compatible with rights-protective constraints (Peeters 2024). In sum, AI can support compliance and reduce procedural burden, but only as legally auditable decision support—i.e., explainable, traceable, and contestable—rather than as an opaque mechanism that effectively predetermines administrative outcomes (Calderón and Ribeiro 2020).

4. Tax Compliance and Artificial Intelligence as a Dispute-Prevention Architecture

Tax compliance should be understood not merely as an internal corporate policy, but as an ex ante governance layer capable of reducing tax disputes by improving the quality, consistency, and traceability of tax-relevant information. In this sense, compliance operates upstream of administrative litigation: when reporting processes are standardized and risk is managed systematically, the likelihood of inconsistencies that later trigger audits, assessments, and appeals is reduced. Fernández de la Cigoña (2023) conceptualizes tax compliance as a set of policies, controls, and procedures aimed at ensuring fulfillment of tax obligations and managing tax risk, thereby decreasing sanction-related contingencies. This preventive rationale becomes particularly salient in digital audit environments, where enforcement increasingly depends on cross-checking large datasets and identifying anomalies across invoices, books, and returns.

A balanced analysis must also recognize the institutional perspective of the tax administration. In Peru, as in many developing and emerging economies, tax enforcement operates under conditions of informality, limited administrative resources, high volumes of transactions, and pressure to mobilize revenue without increasing unnecessary compliance costs. From this perspective, AI-supported tools are institutionally attractive because they may improve audit targeting, reduce false positives, prioritize high-risk cases, and allow scarce enforcement resources to be allocated more effectively. The legitimacy problem, therefore, is not that SUNAT seeks to use data-driven tools to improve enforcement; rather, it is whether those tools are governed in a way that preserves legal accountability, prevents discriminatory targeting, and maintains taxpayers’ capacity to understand and challenge administrative action.

4.1. Digital Tax Enforcement, VAT, and International Tax Pressures

The institutional appeal of AI-supported oversight should also be read in light of broader international tax pressures. The OECD/G20 BEPS agenda and the Pillar Two global minimum tax have increased the importance of reliable reporting infrastructures, cross-border information exchange, and administrative capacity. Recent scholarship shows that Pillar Two may reshape incentives for investment, profit shifting, and domestic minimum taxation, while developing countries may face vulnerabilities if treaty shopping or top-up tax allocation reduces their effective ability to protect their own tax base (Schjelderup 2024; van’t Riet and Lejour 2025). Although this article does not analyze Pillar Two implementation in Peru, this literature is relevant because it shows that domestic tax administration is increasingly linked to global reporting, transparency, and anti-avoidance architectures.

VAT and indirect taxation are equally central to the Peruvian case. E-invoicing systems create the transactional data layer on which risk analysis, anomaly detection, and compliance-by-design strategies depend. However, international evidence cautions that VAT digitalization does not produce uniform results. In Peru, e-invoicing improved reported VAT-related variables and generated compliance spillovers through firm networks (Bellon et al. 2022, 2023). In Rwanda, e-invoicing improved VAT audit efficiency, but its effects were closely linked to enforcement practices (Kotsogiannis et al. 2025). In Uruguay, electronic payment incentives increased card transactions but did not significantly improve tax compliance (Brockmeyer and Sáenz Somarriba 2025). These findings support a cautious conclusion: digital tools can strengthen tax administration, but their fairness and effectiveness depend on institutional design, taxpayer capacity, and procedural safeguards.

4.2. AI-Enabled Compliance as “Compliance-by-Design”

AI can strengthen compliance precisely because it is well suited to continuous reconciliation and early detection of deviations. AI-driven systems may automate the classification of transactions, flag inconsistencies between e-invoicing records and tax returns, identify outliers relative to sectoral baselines, and recommend corrective actions before formal enforcement begins (Akhila et al. 2024; Yordanova 2025). In VAT and similar regimes, AI-based tools have been proposed to support the modeling of regulatory requirements and transactional data, improving the consistency of reporting and the timeliness of corrections (Yordanova 2025). In parallel, predictive modeling techniques (e.g., bagging, support vector machines) have been discussed as mechanisms to forecast tax risk, allowing organizations to implement targeted controls in higher-risk operational areas (Huang et al. 2022).

Importantly, these applications do not simply “make compliance cheaper”. They restructure the evidentiary base that later feeds administrative files: standardized data trails, automated reconciliations, and documented control steps can reduce ambiguity about how information was generated and why certain corrections were made. When disputes do arise, a mature compliance architecture can therefore improve the taxpayer’s ability to explain figures and defend positions, because it produces auditable documentation rather than ad hoc post hoc justifications.

4.3. From Compliance to Fewer Disputes: A Causal Pathway in Administrative Litigation

The contribution of integrating compliance with AI in a tax justice framework lies in articulating a realistic pathway for reducing litigation without sacrificing procedural guarantees. In administrative tax litigation, disputes often escalate because (i) inconsistencies are detected late, (ii) records are fragmented across systems, and (iii) explanations and supporting evidence are reconstructed under time pressure and informational asymmetry. AI-supported compliance can mitigate each of these factors.

First, early warning reduces escalation. By flagging anomalies prior to submission or during ongoing accounting cycles, AI can enable correction before the Tax Administration initiates enforcement or formalizes a position that then becomes difficult to reverse (Shakil and Tasnia 2022). Second, improved traceability reduces interpretive conflict. A compliance program that standardizes documentation, version control, and validation logs can make it easier to demonstrate the provenance and integrity of electronic records, which is increasingly decisive in digital audit contexts. Third, better targeting can reduce unnecessary interventions. Where administrations use risk-based selection tools, improved data quality and consistent reporting can reduce false positives—cases selected due to superficial anomalies rather than substantive noncompliance—thereby lowering both enforcement costs and the volume of contested cases (Han 2022; Huang et al. 2022).

This dispute-prevention logic is compatible with the broader policy aspiration of making tax enforcement more efficient. Yet, it must be framed carefully: the objective is not to replace adjudication with automation, but to reduce avoidable disputes by improving the reliability of tax-relevant data and the capacity to provide reasoned explanations.

4.4. Limits and Risks of AI-Driven Compliance in a Rights-Based System

The same features that make AI attractive for compliance also create legal and institutional risks that must be addressed explicitly.

(a) Compliance asymmetries and unequal capacity. AI-enabled compliance infrastructures are not equally accessible or equally burdensome across taxpayer categories. Large corporations are more likely to possess integrated accounting systems, specialized tax departments, legal counsel, and internal data-analytics capabilities. By contrast, micro and small enterprises, self-employed taxpayers, and individuals with limited administrative capacity may depend on basic software, mobile devices, external accountants, shared connectivity, or manual record-keeping. Consequently, the same digital obligation may operate as a relatively minor extension of existing systems for a large taxpayer while requiring a smaller taxpayer to acquire software, purchase professional assistance, reorganize records, or devote a disproportionate amount of time to compliance. This distinction is consistent with evidence showing that tax compliance costs tend to be regressive in relation to business size and remain comparatively high for smaller firms (Evans et al. 2014).

Technological inequality also concerns access to devices, connectivity, digital skills, accessibility, and confidence in online interactions. Tax-specific research has warned that the digitization of taxpayer services can adversely affect low-income taxpayers, older persons, persons with disabilities, and other groups that encounter barriers to internet access or digital proficiency (Bevacqua and Renolds 2019). This concern is particularly relevant in Peru. In the first quarter of 2025, household internet access reached 80.3% in Metropolitan Lima but only 20.5% in rural areas; access to a computer was reported by 55.0% of households in Metropolitan Lima and 8.2% of rural households (INEI 2025). Although household indicators cannot be treated as direct measures of business compliance capacity, they reveal the broader infrastructure within which many individual taxpayers, self-employed workers, and microenterprises interact with digital tax services.

These disparities create several channels through which facially neutral AI-supported measures may produce differentiated effects. First, taxpayers with incomplete or less standardized digital records may generate more anomaly flags even when the underlying problem is limited administrative capacity rather than deliberate noncompliance. Second, variables such as filing irregularity, transaction density, geographical location, cash intensity, or dependence on intermediaries may operate as proxies for socio-economic vulnerability. Third, taxpayers with greater access to accountants, lawyers, and technical experts are better positioned to correct data errors, reconstruct digital transactions, and rebut algorithmic inferences. Finally, historical enforcement data may generate a feedback loop: sectors or taxpayer groups that were more frequently audited in the past produce more recorded irregularities, which may then be interpreted by a predictive model as evidence justifying continued scrutiny (Chudnovsky and Peeters 2021; Martín López 2022; Palomino Guerrero 2022).

Accordingly, aggregate improvements in reporting or enforcement cannot, by themselves, demonstrate that digitalization is distributionally fair. Increased reported compliance among smaller firms may coexist with higher adaptation costs, greater dependence on intermediaries, or weaker capacity to contest erroneous classifications. The relevant evaluation must therefore consider both benefits and burdens: who receives simplified services or earlier error detection, who bears the cost of technological adaptation, who is selected for additional scrutiny, and who can effectively challenge the resulting administrative position (Bevacqua and Renolds 2019; Evans et al. 2014).

In the Peruvian context, substantive equality requires more than formally identical digital procedures. It requires proportional documentation demands, accessible explanations, assisted-compliance channels, mobile-compatible and disability-accessible services, reasonable alternatives when digital interaction is not feasible, and monitoring of whether audit selection, information requests, sanctions, and dispute outcomes disproportionately affect taxpayers with lower compliance capacity. An inability to use a particular digital channel should not, by itself, generate an adverse inference regarding the taxpayer’s substantive compliance (Bevacqua and Renolds 2019; Giest and Samuels 2023; INEI 2025).

(b) Shifting the practical burden of proof. When AI systems generate risk labels, anomaly flags, or predictive inferences, taxpayers may face an epistemic burden: they are required to refute not only factual claims but also model-driven suspicions that are difficult to interpret or replicate. This is especially problematic if the system’s logic is not explainable in legally operative terms. The literature on XAI in taxation emphasizes that explainability should operate as a minimum legal standard when AI influences high-impact tax processes, because taxpayers must be able to understand and challenge the logic that shapes audit selection, risk classification, or evidentiary expectations (Kuźniacki et al. 2022). Without this standard, AI-supported compliance may shift complexity onto taxpayers, weaken the practical exercise of defense rights, and turn compliance into a mechanism for externalizing algorithmic opacity onto the affected party.

(c) Bias and selective enforcement through proxies. AI models trained on historical data can encode structural patterns of enforcement and reporting. Even if a compliance system is internal to a taxpayer, the broader ecosystem matters: administrative models that react to certain patterns may indirectly pressure taxpayers toward behaviors aligned with those models, potentially embedding biased heuristics into compliance practice. In inspection settings, concerns about bias and non-discrimination are therefore inseparable from AI governance (Martín López 2022; Palomino Guerrero 2022).

(d) Privacy and proportionality constraints. Compliance-by-design can expand data aggregation and monitoring. This raises proportional questions: which data are necessary, who can access them, and how long they are retained. These are not merely technical issues; they condition the legality and legitimacy of a compliance ecosystem in which private systems and public enforcement infrastructures increasingly interact (Calderón and Ribeiro 2020; Hadwick 2022).

4.5. A Minimum Governance Position for Peru: Compliance as Auditable Support, Not Automated Authority

Against this backdrop, the regulatory challenge in Peru is not simply to “adopt compliance” or “adopt AI,” but to define minimum governance conditions under which AI-enabled compliance contributes to reducing litigation while preserving safeguards. A practical and rights-compatible position is to treat AI as auditable decision support within compliance and enforcement pipelines. At the compliance layer, this entails four immediate requirements, which operate within the broader six-safeguard framework developed in Section 8.

Explainability in legally meaningful terms: AI-driven compliance outputs that may later shape assessments or disputes should be explainable as legally relevant reasons, not merely as opaque scores or technical classifications. This approach is consistent with Kuźniacki et al. (2022), who frame XAI in tax law as a minimum legal standard rather than as an optional technical improvement.

Traceability and documentation: compliance processes should preserve logs, provenance, and change histories for tax-relevant records so that explanations are verifiable in litigation contexts (Han 2022).

Human accountability: AI should not displace professional judgment in determining reporting positions; rather, responsible personnel must remain accountable for decisions and able to justify them with evidence (Calderón and Ribeiro 2020).

Non-discrimination and proportionality controls: risk and monitoring practices must be subject to governance measures that detect biased impacts and prevent excessive data collection (Martín López 2022; Palomino Guerrero 2022).

Under these constraints, AI-enabled compliance becomes a credible lever for reducing avoidable disputes—by improving data quality and traceability—while maintaining contestability and legal reviewability. This framing also anticipates the risks analyzed in Section 5: opacity, information asymmetry, and evidentiary burden increase when AI outputs are not explainable or reconstructable. Consequently, the compliance layer discussed here functions as the preventive counterpart of the safeguards framework developed in Section 8, where legally meaningful explainability, traceability, meaningful human oversight, lifecycle auditability, effective contestation, and distributional equality and accessibility are formalized as conditions for digital tax justice.

5. Risks: Opacity, Bias, Unequal Impact, Information Asymmetries, Evidentiary Burden, and Traceability

Opacity and the duty to give reasons. AI-supported tax enforcement frequently relies on complex models (e.g., risk scoring, anomaly detection, network analytics) whose internal logic is not readily intelligible to taxpayers or even to decision-makers. This creates a black-box effect that can undermine the administrative duty to provide reasons and the practical possibility of judicial review. Even where “explainability” is invoked as a remedy, the literature warns that explanations can be context-dependent and strategically incomplete, particularly in “wicked” governance problems where there is no consensus on what counts as an adequate justification (de Bruijn et al. 2022). A key institutional risk is that algorithmic outputs become de facto determinative: audit selection or liability assessments are treated as objective “technical” results, while the underlying criteria remain unchallengeable. In the tax domain, this concern is amplified by evidence that taxpayers may be “at a loss” regarding whether AI-based determinations respect legal procedures—hence the insistence that XAI must be aligned with reason-giving obligations, not merely technical transparency (Górski et al. 2025). More broadly, work on defining explainability requirements shows that transparency must be translated into verifiable system requirements and documentation practices—otherwise, “ethics” remains aspirational and non-justiciable (Balasubramaniam et al. 2023).

Bias, unequal impact, and selective enforcement. AI systems trained on historical tax and audit data can reproduce—and sometimes amplify—pre-existing enforcement patterns. In practice, this risk is not limited to overt discrimination; it also includes proxy effects (variables correlated with protected or vulnerable categories) and structural bias embedded in administrative routines. In tax inspection contexts, the non-discrimination problem is particularly salient because risk scoring may change the probability of being audited, the intensity of scrutiny, and the evidentiary expectations placed on the taxpayer. The emerging literature in tax fraud detection illustrates both the promise of AI-based oversight and the governance challenge of ensuring that model performance does not translate into unfair targeting (Belahouaoui and Alm 2025). A parallel survey of tax risk detection methods highlights that high-performing approaches (e.g., graph-based learning) can be poorly interpretable and computationally intensive—features that complicate accountability and increase the likelihood that biased signals remain hidden (Zheng et al. 2024). In governance terms, this is precisely why algorithmic auditing and ethics-based AI auditing have been framed as lifecycle practices (design–deployment–monitoring) rather than one-off compliance checks (Koshiyama et al. 2024; Laine et al. 2024).

Information asymmetries and procedural disadvantages. AI can deepen asymmetries between the tax authority (which controls data infrastructures, models, and institutional knowledge) and taxpayers (who usually lack access to the same informational resources). In administrative litigation, this asymmetry can translate into unequal “procedural capacity”: the authority can justify its actions through proprietary analytics or internally unavailable risk indicators, while the taxpayer is expected to rebut conclusions without access to comparable information. Recent work on AI-driven decision-making in the public sector shows that AI adoption also reconfigures internal power and expertise dynamics, reinforcing the dependence on specialized analysts who control model interpretation (Mahroof et al. 2025). From a procedural fairness perspective, this matters because procedural justice depends not only on outcomes but on whether affected individuals have meaningful opportunities to understand and contest decisions (Decker et al. 2025). In the tax sphere, studies that explicitly link AI-based tax administration to taxpayers’ rights also emphasize that a lack of transparency and accountability undermine due process; therefore, it is necessary to institutionalize appropriate oversight and review mechanisms (Guglyuvatyy 2025).

Shifts in the burden of proof and epistemic overload. When AI is introduced into audit selection or assessment pipelines, the practical burden may shift toward the taxpayer: contesting a decision increasingly requires technical understanding, data reconstruction, and evidentiary rebuttal of model-driven inferences. This “epistemic burden” is not a minor inconvenience; it can effectively reduce contestability, especially for individuals and smaller firms with limited resources. Empirical reviews of “human oversight” policies caution that merely adding a human-in-the-loop does not solve this problem if decision-makers are not equipped to meaningfully interrogate AI outputs or to override them in a reasoned manner (Green 2022). Complementarily, work on institutional trust and asymmetric information underscores that governance must create conditions for meaningful choice and accountability without imposing excessive informational burdens on users—a point directly relevant to taxpayers confronting algorithmically mediated enforcement (Dowding and Taylor 2024). In tax algorithmic governance, these dynamics resemble a “one-way mirror,” where authorities can see and classify taxpayers, while taxpayers cannot see the operational logic applied to them (Hadwick 2022).

Traceability, evidentiary integrity, and auditability. Digital tax justice depends on whether AI-supported actions are reconstructable: what data were used, which model/version was applied, what thresholds were active, what human interventions occurred, and how outputs were incorporated into the administrative file. Without such traceability, judicial review becomes largely formal and the taxpayer’s right to defense is weakened. Recent archival and governance scholarship argues for preserving paradata—records of the processes, tools, and agents that produce outputs—as a foundation for accountability in semi-autonomous systems (Cameron and Hamidzadeh 2024). In parallel, forensic and security research on digital chain-of-custody frameworks emphasizes that integrity depends on verifiable logs and tamper-resistant evidence handling, including blockchain-based approaches for preserving multimedia evidence until adjudication (Sakshi and Sharma 2023). For tax litigation, the normative implication is straightforward: AI use must be paired with mandatory logging, access records, data provenance, and model documentation so that decisions can be audited internally and challenged externally, consistent with broader trustworthy-AI requirements in public-sector automated decision-making (Agbabiaka et al. 2025) and with transparency-by-design approaches for public policymaking (Papadakis et al. 2024).

6. Artificial Intelligence and the Judicial Process

Judicial review in tax matters is designed to restore legality and protect procedural guarantees once administrative remedies have been exhausted. In this context, the growing digitization of judicial case management (e-files, electronic submissions, and interoperable records) makes it increasingly plausible that courts will adopt AI not only to handle volume but also to support legally relevant tasks such as document triage, information retrieval, and consistency checks across large case files (Mingtsung and Shuling 2020; Xu 2022).

A useful way to delimit legitimate judicial uses is to distinguish where AI intervenes in adjudication. Empirical research suggests that social acceptance and perceived fairness vary by stage: AI support tends to be viewed as less problematic when it assists information acquisition (e.g., search, extraction, clustering of documents) than when it moves toward decision selection or decision implementation, where it can be perceived as displacing judicial responsibility (Barysė and Sarel 2024). For tax disputes, this distinction is not merely perceptual; it maps onto the intensity of rights impact. Tools that improve navigation of electronic records may reduce delay and improve consistency, but once AI outputs begin to steer outcomes—through risk labels, suggested holdings, or automated reasoning—courts face heightened risks of opacity and contestability deficits.

These risks become more acute with the introduction of generative AI. Even if used for drafting, summarizing, or proposing arguments, generative systems may introduce errors that are difficult to detect, weaken evidentiary traceability, and create accountability gaps regarding who is responsible for the content that ultimately grounds a judicial decision. Regulatory analysis of judicial uses of generative AI emphasizes that accountability requires a clear allocation of roles, documented oversight practices, and risk-based safeguards—precisely because “efficiency” is not a sufficient justification when fundamental procedural rights are at stake (Carnat 2024). In tax litigation, where the dispute often turns on technical evidence and the reasoning that connects facts to legal consequences, any tool that affects how facts are filtered or how arguments are framed must remain auditable and open to contradiction.

For this reason, “human-in-the-loop” language is insufficient if it remains purely formal. Evidence from governance studies shows that many oversight policies assume—without support—that humans can effectively monitor algorithmic outputs; in practice, limited time, expertise gaps, and automation bias can convert oversight into a symbolic shield rather than a real safeguard (Green 2022). Applied to judicial review, meaningful oversight requires that judges and court staff retain the capacity to interrogate AI-assisted outputs, demand the underlying basis for summaries or classifications, and explicitly justify departures from or reliance on those outputs. Otherwise, AI risks generating argumentation gaps—i.e., decisions that appear reasoned but in fact rest on opaque or unchallengeable intermediate steps—undermining the legitimacy of adjudication (Fouquet 2021).

Accordingly, in digital tax justice the appropriate role of AI in the judiciary is supportive and bounded. AI may legitimately enhance case management and information handling, particularly at early stages of processing and analysis; however, it should not be treated as a source of authority capable of substituting judicial reasoning. A rights-compatible framework therefore requires: (i) documented and reproducible workflows (what the tool does, when, and with what data), (ii) traceable records of human intervention, (iii) disclosure and contestation pathways when AI meaningfully shapes fact selection or argument framing, and (iv) an explicit reaffirmation that the judicial decision—its evaluation of evidence and legal justification—remains attributable to human decision-makers and fully reviewable on the record (Xu 2022; Xu et al. 2022).

7. Digital Evidence in Tax Administrative Litigation and Judicial Review

The shift toward electronic tax administration has transformed the evidentiary landscape of tax disputes. In contemporary tax controversy, the “file” is increasingly built from native-digital records: e-invoicing and electronic receipts, digital accounting books, e-returns, third-party reporting, interoperable datasets, and the transactional traces produced by platforms used to issue, transmit, store, and validate tax-relevant information. As a result, administrative tax litigation and subsequent judicial review no longer revolve primarily around paper documents, but around data integrity, provenance, and the reconstructability of digital records within electronic case files.

This transformation has two immediate implications. First, the probative value of digital tax evidence depends less on its existence than on its verifiable authenticity and integrity over time. Digital objects are comparatively easy to copy and modify, and their evidentiary strength depends on whether parties can demonstrate that the record presented for evaluation is the same as the record originally generated or obtained, without alteration (Perales 2016). Second, digitalization expands the number of intermediaries that may affect evidence quality (information systems, service providers, automated validations, data transformations), which increases the need for documented procedures that preserve reliability and prevent contestation from becoming a purely technical battle.

Within this setting, traceability becomes the functional equivalent of the classic evidentiary chain of custody. The admissibility and weight of electronic tax documents depend on whether the dispute forum can reconstruct who generated the record, when and how it was stored, whether and how it was transmitted, who accessed it, what transformations were applied, and how it entered the administrative file. The chain of custody must therefore be conceived as a chronological and verifiable record of custody, access, and processing, ensuring that the evidence evaluated by the reviewing body corresponds to the evidence originally obtained (Shah et al. 2017). This requirement is not limited to “documents” in the narrow sense. In tax disputes, it extends to system logs, audit trails, metadata, and validation records, which often become decisive to establish whether a given transaction, declaration, or registry entry is attributable and reliable.

Because electronic evidence is frequently distributed across systems and repositories, proposals for tamper-resistant custody mechanisms have gained attention. Blockchain-based approaches, for example, have been discussed to preserve integrity and document custody events for digital files, precisely by making later alterations detectable and by providing verifiable histories of access and transfer. Similar custody-preservation frameworks have been developed for digitally generated evidence in IoT environments, again emphasizing that integrity depends on robust logging and verifiable handling (Sakshi and Sharma 2023). While these technical solutions are not automatically transferable to tax litigation, they reinforce a core point: digital tax justice requires evidence governance that is auditable and resistant to manipulation, not merely digitized.

The evidentiary challenge becomes more complex once AI-supported tools enter the tax dispute ecosystem. Risk scores, anomaly flags, clustering outputs, or automated consistency checks can influence which records are collected, highlighted, or prioritized; consequently, AI affects not only administrative efficiency but also the construction of the evidentiary narrative. This creates a procedural risk: if the taxpayer must rebut an assessment shaped by AI-driven inference without access to the underlying data lineage and processing steps, the practical burden of proof can shift in ways that weaken contestability. In other words, the evidentiary problem is not only whether a particular invoice or ledger entry is authentic, but whether the pipeline that transformed raw data into an actionable enforcement or litigation position is reviewable.

For this reason, traceability must cover both evidence and computation. In addition to preserving the integrity of tax-relevant records, institutions must preserve the technical and organizational traces that explain how those records were processed and interpreted. Archival governance scholarship has proposed preserving paradata—records about processes, tools, and agents that generate output as a basis for accountability in semi-autonomous systems (Cameron and Hamidzadeh 2024). Transposed to tax litigation, this means that when AI-supported analytics influence the dispute, the file should include: data provenance (what sources were used), transformation logs (how data were cleaned or aggregated), model/version identifiers (what system produced the output), thresholds or rules applied, and records of human intervention. Without these traces, judicial review risks becoming formalistic, because neither the reviewing body nor the taxpayer can test the reliability of the AI-mediated evidentiary path.

These observations serve as a bridge to the governance framework outlined in Section 8. If digital tax justice depends on reconstructable evidence and reviewable computation, then safeguards cannot be limited to abstract principles. They must be operationalized through: (i) explainability compatible with legal reasoning; (ii) meaningful human oversight over AI-mediated evidentiary steps; (iii) traceability of both evidence and computation; (iv) lifecycle auditability through documented logs and independent review; (v) effective contestation mechanisms that guarantee access and the practical ability to challenge both data and inference; and (vi) distributional equality and accessibility controls that identify and correct unequal technological access, compliance burdens, and enforcement effects.

8. AI Governance, Transparency, and Safeguards in Digital Tax Justice

The expansion of electronic records and advanced analytical tools in tax administration and contentious litigation necessitates a rethinking of the governance of artificial intelligence (AI) within parameters compatible with the rule of law. In this field, institutional legitimacy depends not only on efficiency or the reduction of procedural burdens, but—primarily—on whether automated or semi-automated systems preserve legally meaningful reason-giving, responsible human oversight, the right to defense, evidentiary traceability, effective auditing, and substantive equality across taxpayers with different technological and compliance capacities. Consequently, a governance framework for digital tax justice must be conceived as a set of verifiable conditions—not mere declarations of principles—aligned with the system’s lifecycle, including design, training, deployment, operation, and review, so that the adoption of technology does not undermine procedural safeguards or the standards of reasonableness and proportionality required of administrative and judicial actions (Agbabiaka et al. 2025).

First and foremost, explainability serves as the essential link between automated decision-making and legal oversight. However, the literature cautions that “explaining” is not the same as “convincing” nor does it guarantee social acceptance: explanations are contextual, they can be challenged, and in scenarios of regulatory conflict or “wicked problems,” even a technically correct explanation may be perceived as insufficient or strategically biased (de Bruijn et al. 2022). Therefore, in the context of digital tax justice, it is advisable to adopt an action-oriented approach to explainability: (i) describe the system’s purpose (what it is used for), (ii) define its role in the decision-making process (recommendation, prioritization, detection of inconsistencies, etc.), (iii) identify the relevant data and its quality, and (iv) provide a comprehensible explanation of the result in legally operative terms (e.g., risk criteria or detected patterns), avoiding the imposition of unreasonable burdens on the taxpayer (e.g., “the neural network indicated it”). At the technical-organizational level, it is preferable to prioritize models and architectures that support transparency “by design”—interpretable rules or reproducible explanations—especially in high-impact applications, as proposed by transparent AI approaches applied to public decision-making environments (Papadakis et al. 2024). In other words, when a decision may affect rights (sanctions, objections, denials), explainability should be treated as a requirement for justiciability, not as a secondary attribute.

Second, human oversight must be understood as “meaningful” oversight, not merely a ritualistic formality that legitimizes opaque outcomes. Critical evidence shows that policies based solely on “human-in-the-loop” approaches can fail if real capabilities for oversight, correction, and accountability are not designed; in particular, human oversight can become a mechanism for shifting blame or a symbolic cover for algorithmic decisions that, in practice, no one is in a position to question (Green 2022). Applied to tax litigation, this means that the decision-maker (whether administrative or judicial) must have the expertise, time, and substantive authority to review the system’s assumptions, request additional evidence, deviate from the recommendation, and provide a reasoned explanation of why they are doing so. Furthermore, human oversight must be institutionalized through internal rules: separation of duties (who develops, who operates, who decides), review committees for high-impact systems, and escalation protocols when there are alerts regarding bias, systematic errors, or model degradation (Agbabiaka et al. 2025).

Third, traceability must cover both the evidentiary record and the computational process through which AI-supported outputs are produced. A tax-related output is legally reviewable only when the relevant processing path can be reconstructed: which data sources were used, how the data were cleaned or transformed, which model and version were applied, which thresholds or business rules were active, and what human interventions occurred. Accordingly, institutions should preserve data provenance, transformation logs, model and version identifiers, decision thresholds, records of queries, audit trails, and evidence of integrity, including appropriate version control and digital chain-of-custody mechanisms. These records should form part of the administrative governance structure and, when materially relevant to a dispute, should be available in a form that permits the taxpayer and the reviewing body to assess the reliability of the AI-mediated evidentiary path (Cameron and Hamidzadeh 2024; Han 2022; Koshiyama et al. 2024).

Fourth, lifecycle auditability must be institutionalized as a continuous and multi-level governance requirement. Auditing should not be confined to technical performance at the deployment stage. It should include internal audits of policy compliance, data quality, access controls, and institutional responsibilities; technical audits of robustness, error rates, model drift, and discriminatory effects; legal audits of compatibility with due process, reason-giving, proportionality, and data-protection requirements; and independent external audits for high-impact systems. Lifecycle auditability requires documented risk assessments, monitored performance and bias indicators, incident-reporting and remediation protocols, public documentation proportionate to the system’s risk, and periodic determinations of whether the system should be recalibrated, restricted, suspended, or withdrawn (Agbabiaka et al. 2025; Koshiyama et al. 2024; Laine et al. 2024). In the Peruvian public sector, this requirement should incorporate cybersecurity and organizational risk-management controls, since uneven institutional implementation may affect information quality, process reliability, and decision effectiveness (Olivos Estrada et al. 2025).

Fifth, effective contestation mechanisms must enable taxpayers to understand, challenge, and correct both the data and the inferences used in AI-supported tax processes. From the perspective of procedural justice, fairness is not limited to statistically neutral outcomes; it requires procedures that recognize informational and institutional power asymmetries and provide affected persons with a genuine opportunity to challenge decisions. In digital tax administration, this requires notification when AI materially influences case selection, prioritization, anomaly detection, evidentiary assessment, or workflow management; timely access to the relevant electronic case file; disclosure of the legally relevant criteria supporting the administrative action; and the ability to dispute inaccurate data, inappropriate assumptions, proxy variables, or unsupported algorithmic inferences. Contestation must therefore operate as a practical right rather than as a formal possibility (Decker et al. 2025; Hadwick 2022).

Where risk profiles, anomaly flags, or automated classifications influence an administrative position, taxpayers must be able to request meaningful human reassessment, submit exculpatory or contextual evidence, obtain correction of inaccurate information, and receive a reasoned response within a reasonable period. The reviewing official must possess sufficient authority to modify or disregard the AI-supported output and must explain whether and how the taxpayer’s objections were evaluated. Access to source code will not necessarily be required in every case; however, trade secrecy, cybersecurity concerns, or technical complexity should not prevent disclosure of the information necessary to challenge the factual and inferential basis of an adverse administrative action (Górski et al. 2025; Guglyuvatyy 2025).

Sixth, distributional equality and accessibility must operate as lifecycle governance requirements. Distributional equality should be treated as an independent and cross-cutting safeguard rather than as an assumed consequence of procedural transparency. Before an AI-supported tax system is deployed, and periodically throughout its operation, the responsible authority should conduct a distributional impact and administrative-burden assessment. This assessment should identify the taxpayer categories likely to be affected, examine the technological and professional resources required to comply, and determine whether the system transfers learning, compliance, evidentiary, or psychological costs from the administration to taxpayers with lower institutional capacity (Chudnovsky and Peeters 2021; Moynihan et al. 2015).

The assessment should include disaggregated monitoring of audit-selection rates, false-positive rates, the frequency and complexity of information requests, response and correction times, sanction rates, access to human assistance, appeal rates, and administrative or judicial reversal rates. Where legally permissible and statistically meaningful, the analysis should distinguish outcomes by taxpayer size or turnover, economic sector, geographical location, digital access channel, and other indicators relevant to compliance capacity. This is particularly important because compliance costs tend to represent a greater relative burden for smaller businesses, while digitalized public services may create additional obstacles for taxpayers with limited technological access or proficiency (Bevacqua and Renolds 2019; Evans et al. 2014; Giest and Samuels 2023).

Material and unexplained disparities should trigger an investigation of input data, proxy variables, model thresholds, historical enforcement patterns, and human implementation practices. Depending on the source and seriousness of the disparity, corrective measures may include model recalibration, revision of risk criteria, exclusion of problematic variables, simplified evidentiary requirements, additional taxpayer assistance, longer or more flexible response periods, or temporary suspension of the affected system (Koshiyama et al. 2024; Laine et al. 2024).

Finally, digital-by-default administration should not become digital-only administration. Taxpayers who face connectivity, accessibility, linguistic, or digital-literacy barriers must retain an effective route to obtain information, submit evidence, correct data, and request human review. Assisted and alternative channels should therefore be treated as components of rights-compatible digital administration rather than as exceptional concessions (Bevacqua and Renolds 2019; Giest and Samuels 2023).

The preceding analysis can be summarized in terms of institutionally appropriate AI uses and their minimum safeguards.

Table 2. Institutionally appropriate AI uses in digital tax justice.

AI Use	Potential Advantage	Main Risk	Minimum Safeguard
Audit triage and risk scoring	Prioritizes scarce enforcement resources and reduces random or inefficient audits	Opaque selection, proxy discrimination, over-targeting of vulnerable taxpayers	Explainable risk categories, human validation before enforcement action, testing for proxy effects, and periodic comparison of selection and false-positive rates across taxpayer size, sector, region, and other legally permissible indicators of compliance capacity.
Anomaly detection in e-invoicing and returns	Detects inconsistencies earlier and may prevent avoidable disputes	False positives and evidentiary burden shifting	Access to detected inconsistency, opportunity to correct or explain, documented thresholds
Document classification and case-file organization	Reduces processing time and improves electronic file navigation	Relevant evidence may be deprioritized or omitted	Full file access, logs of classification, human review of contested documents
Workflow prioritization in administrative appeals	Reduces backlog and improves case management	Hidden prioritization criteria may affect procedural equality	Public prioritization criteria, audit logs, review by responsible officials
Compliance assistance and taxpayer guidance	Supports voluntary compliance and reduces errors	Incorrect or misleading automated guidance	Disclaimer, escalation to human assistance, preservation of taxpayer reliance evidence
Judicial information retrieval and summarization	Supports judges and staff in complex digital files	Hallucination, selective summarization, loss of evidentiary nuance	Disclosure of AI assistance, source-linked summaries, human responsibility for reasoning
Taxpayer-facing digital compliance, notification, and AI-assisted services	Reduces filing errors, provides earlier guidance, and may lower average compliance costs	Digital exclusion, regressive adaptation costs, missed notifications, dependence on professional intermediaries, and reduced ability to correct or contest data	Assisted and non-digital access channels; mobile-compatible and accessible design; proportional documentation and response periods; human escalation; no adverse inference based solely on digital incapacity; disaggregated monitoring of access, errors, sanctions, and dispute outcomes

distinguishes between low-risk supportive uses and higher-risk uses that require stronger procedural controls.

Limitations and Future Research

This study has limitations that should be acknowledged. First, it is a legal-doctrinal and policy-analytical study, not an empirical audit of AI systems currently used by SUNAT, the Tax Court, or the judiciary. Therefore, its conclusions concern the legal and institutional conditions under which AI-supported tools should operate, rather than the measured performance, accuracy, or bias of specific systems. Future research should examine concrete AI or analytics tools used in Peruvian tax administration, including their input data, model design, governance arrangements, human oversight practices, and effects on audit selection or dispute outcomes.

Second, the comparative dimension of this article is functional and illustrative rather than systematic. International evidence from Peru, Ethiopia, Rwanda, Uruguay, Ecuador, and China is used to contextualize risks and institutional design choices, but the article does not conduct a full comparative legal analysis. Future research should develop systematic comparative studies on algorithmic tax governance across Latin American jurisdictions, especially in countries with mature e-invoicing systems. Such studies should assess whether digital tax enforcement produces unequal procedural burdens across taxpayer categories, whether smaller taxpayers face disproportionate compliance costs, and whether contestation mechanisms are sufficiently accessible when enforcement relies on algorithmic risk assessment.

9. Conclusions

This article examined AI-supported tax compliance and administrative tax litigation in Peru through the lens of algorithmic tax justice. The analysis answers the first research question by concluding that the institutionally appropriate uses of AI are those confined to bounded and auditable decision support. In tax administration, these include audit triage, anomaly detection, compliance assistance, document classification, workflow prioritization, and consistency checks. In judicial or quasi-judicial settings, AI may assist information retrieval, file organization, and preliminary summarization. However, AI should not substitute legally attributable human reasoning, nor should opaque risk scores become de facto grounds for tax assessments, sanctions, or adverse procedural treatment.

The second research question concerns the minimum conditions required to reconcile efficiency with due process, reason-giving, and effective contestation. The answer is that AI can reduce avoidable litigation only if six safeguards are institutionalized. First, explainability must be legally meaningful: affected taxpayers and reviewing bodies must understand the relevant criteria, not merely receive technical descriptions. Second, traceability must cover both evidence and computation, including data provenance, logs, model and version information, thresholds, and human interventions. Third, human oversight must include real authority to interrogate, modify, or override AI-supported outputs. Fourth, lifecycle auditability must be mandatory, including internal, technical, legal, and independent audits for high-impact systems. Fifth, contestation mechanisms must allow taxpayers to access, rebut, and correct the data and inferences used against them. Sixth, distributional equality and accessibility must require ex ante and periodic assessment of technological access, compliance burdens, false-positive rates, and disparate effects across taxpayer categories.

The article’s theoretical contribution is to conceptualize algorithmic tax justice as a framework that integrates distributive, procedural, corrective, and institutional dimensions of justice. This framework clarifies that digital tax justice is not achieved by improving efficiency alone. Rather, efficiency becomes normatively acceptable only when AI-supported enforcement remains explainable, reviewable, non-discriminatory, and contestable.

The practical contribution is the proposed allocation of institutional responsibilities. SUNAT should ensure explainable audit selection, documented risk criteria, taxpayer access to relevant digital evidence, and internal monitoring of bias and false positives. The Ministry of Economy and Finance should define cross-cutting regulatory standards for AI-supported tax enforcement and compliance-by-design. The Tax Court should require that AI-influenced administrative files include sufficient traceability for appellate review. The judiciary should ensure that AI-supported evidentiary paths remain open to contradiction in contentious-administrative proceedings. Finally, the data protection authority should supervise proportionality, purpose limitation, security, and data-subject rights when taxpayer data are processed through AI-enabled analytics.

The distributive dimension modifies the criterion by which administrative efficiency must be evaluated. An AI-supported system cannot be considered consistent with digital tax justice merely because it reduces average processing time, improves aggregate revenue collection, or lowers the administration’s operational costs. Those gains may conceal a transfer of costs and risks toward taxpayers who have weaker connectivity, less standardized records, lower digital literacy, or limited access to professional assistance. A facially neutral system may therefore remain substantively unequal when it systematically changes who is audited, who must produce additional evidence, who depends on intermediaries, and who can effectively challenge an erroneous inference.

For this reason, the minimum safeguards framework includes distributional equality and accessibility as a sixth requirement. SUNAT should conduct ex ante and periodic impact assessments, monitor relevant outcomes across taxpayer categories, provide assisted and alternative service channels, and adjust documentation or response requirements in proportion to taxpayers’ actual compliance capacity. The Ministry of Economy and Finance should incorporate these obligations into cross-cutting standards for AI-supported tax administration, while the Tax Court and the judiciary should examine whether technological barriers or unequal informational capacity have impaired effective defense. Digital tax justice thus requires differentiated institutional support where formally uniform procedures would otherwise reproduce existing socio-economic inequalities.

Accordingly, the conclusion is conditional in both procedural and distributive terms. AI may contribute to reducing avoidable disputes and improving compliance only when it operates as auditable decision support and when its benefits and burdens are not distributed in a manner that systematically disadvantages less-resourced taxpayers. Explainability, traceability, human oversight, auditability, and contestation remain indispensable, but they must be complemented by accessibility, proportionality, and continuous monitoring of unequal effects.