1. Scope and legal framework of the Regulation
The Regulation laying down provisions on the processing of personal data carried out by the Bank of Italy in the context of its support activities for the Banking and Financial Ombudsman – Arbitro Bancario Finanziario, ABF – (hereinafter, ‘the Regulation’), approved in September 2026, is a general administrative act through which the Bank of Italy, acting as data controller, identifies the types of personal data processed, the operations that may be performed, and the security measures adopted to protect the rights and freedoms of data subjects. The processing takes place in the context of the support provided for the functioning of the ABF, which is carried out in fulfilment of a legal obligation and in the performance of a task carried out in the public interest, pursuant to Article 6(1)(c) and (e) GDPR. Such a legal obligation and public-interest task are identified by sector-specific legislation and entrusted to the Bank of Italy.
The Regulation focuses on the purposes and methods of processing, providing specific measures to protect the rights and freedoms of data subjects. Its provisions, which also take into account the use of AI systems, govern purposes, legal bases, data, data subjects, operations, rights and safeguards.
In this context, the Bank of Italy considers it advisable to use AI systems to facilitate the administrative and support activities it carries out for the functioning of the ABF. These systems would be used to identify the decision-making patterns followed in the ABF’s activities and any deviations from them, as well as for preliminary or otherwise purely ancillary activities, which do not affect the decision-making processes in individual disputes or, more generally, the administration of justice by the ABF.
To carry out its support activity, the Bank of Italy must process personal data, special categories of data, and data relating to criminal convictions and offences. For this reason, the Regulation lays down specific provisions on the operations that may be performed, the protective measures and the methods of processing, including with reference to the use of AI systems, with a view to ensuring clarity and transparency towards data subjects. The operations include the collection, storage, retrieval, comparison and interconnection of case files, the search for earlier complaints by the same claimant, the preparation of technical reports by the Technical Secretariats for the ABF decision-making bodies (the Panels), statistical analysis, publication of anonymised decisions and transmission to other authorities. The protective measures include access restricted to authorised staff, access logs, annual reviews of access privileges, disk and transmission encryption, AI applications isolated from external networks, enhanced safeguards for documents on Panel appointments (which may refer to criminal matters), operator training, a multidisciplinary team in charge of monitoring, retraining and versioning the models, retention periods ranging from 5 to 30 years (10 for data in AI applications), and the removal of unnecessary third-party data by the Panel. Finally, the Regulation clarifies that the use of AI tools does not entail a significant risk to the rights and freedoms of data subjects, as they do not materially influence the outcome of the decision-making process.
Following this overview, Section 2 examines the high-risk classification under the AI Act and related GDPR questions, Section 3 assesses purpose limitation and data subjects’ rights, and Section 4 draws together four areas of vulnerability.
2. Artificial Intelligence and the high-risk classification
The classification of the systems as not high-risk is a decisive issue. This is because point 8 of Annex III to the AI Act includes among high-risk AI systems those used in alternative dispute resolution to research and interpret facts and the law. The Bank of Italy’s Regulation therefore relies on the derogation provided for in Article 6(3) and on Recital 61: the systems are said to perform limited procedural, preparatory or decision-pattern identification tasks without replacing human assessment.
The validity of this argument varies depending on the specific function performed.
The retrieval of previous decisions and assistance with admissibility checks have a plausibly auxiliary nature. Identifying centroids and outliers (the typical and the most deviant decisions in each cluster), however, is more problematic. Although point (c) of Article 6(3) covers systems that detect decision-making patterns without replacing or influencing a previously completed human assessment, the issue remains that indicating to a Panel which decisions are considered “typical” and which are “deviant” may influence the actual decision-making process, particularly through anchoring effects and the risk of automation bias (Tversky and Kahneman, 1974; Parasuraman and Manzey, 2010).
The statement that the decision is “entirely entrusted to human assessment” and the description of governance measures (such as a multidisciplinary team, periodic retraining, explainability and model versioning), does not answer the question of whether human oversight is effective in practice (Green, 2022). The fact that this classification is a self-assessment, contained in an act adopted by the very entity using the system, also deserves critical scrutiny (for a critical analysis of the AI Act’s reliance on self-assessment, Veale and Zuiderveen Borgesius, 2021).
A further issue concerns the profiling exception. Under the last subparagraph of Article 6(3) AI Act, a system that performs profiling of natural persons is always considered high-risk, and the AI Act adopts the GDPR definition of profiling.
The Bank of Italy’s Regulation expressly excludes any form of profiling. However, searching for complaints brought by the same claimant against the same financial intermediary over the preceding ten years entails the processing of personal data and the reconstruction of an individual’s case history. The purpose is narrowly defined, namely ensuring compliance with the ne bis in idem principle. Nevertheless, it is worth considering whether the definition of profiling under Article 4(4) GDPR, which refers to the evaluation of personal aspects relating to an individual, can genuinely be regarded as unrelated to this activity (Wachter and Mittelstadt, 2019; Kuner et al., 2020).
As regards Article 22 GDPR, the absence of a solely automated decision-making process appears plausible, given the role played by the ABF Panel (Binns and Veale, 2021). A question nevertheless remains as to whether Article 15 of Law No. 132/2025, which reserves every decision to the judge, could apply, if only as an underlying guiding principle. The ABF Panel is not a judicial body, but the reference to this provision in the recitals suggests that the Regulation intends to incorporate its underlying rationale.
There also appears to be a tension between different statements in the Regulation. On the one hand, it repeatedly emphasises that clustering is performed exclusively on anonymised decisions and that AI-based analysis “does not concern personal data”. On the other hand, it provides that AI tools analyse complaints and their attachments for admissibility checks and verification of compliance with the ne bis in idem principle. These documents contain personal data, including special categories of data, and are retained in the relevant applications for ten years. Furthermore, AI is also used to assist in verifying anonymisation. Such statements can be reconciled only by drawing a much clearer distinction between the individual functions performed by the systems.
Anonymisation also warrants critical examination. In this respect, Recital 26 GDPR and Article 29 Working Party Opinion 05/2014 require that identification be impossible by means reasonably likely to be used. In practice, merely removing the cover page and the names of the parties may not be sufficient where the body of a decision contains indirectly identifying elements (Ohm, 2010; Finck and Pallas, 2019). The Regulation relies on precisely these measures: the system strips the cover page carrying the parties’ identities and the case number, operational instructions tell the Technical Secretariats and Panels to avoid personal data in the body of decisions, and an AI function assists in verifying anonymisation. These safeguards could usefully be complemented. A documented re-identification risk assessment based on the Opinion 05/2014 criteria (singling out, linkability and inference) would be a suitable starting point. It may also be worth considering removing rare attributes and quasi-identifiers, and generalising others such as dates, places or amounts, to coarser granularity.
3. Purpose limitation and data subjects’ rights
The first issue is purpose limitation. The Regulation allows the further use of analytical results to inform supervisory activities and consumer protection initiatives. The compatibility of this further processing must be assessed under Articles 5(1)(b) and 6(4) GDPR.
The safeguards identified include the aggregation of results, the impossibility of identifying individuals and the fact that the analyses do not, in themselves, trigger supervisory inspections but merely constitute an additional source of information. Nevertheless, the risk of function creep remains, precisely because the same institution performs both support and supervisory functions (Koops, 2021). More generally, formal separation of purposes offers limited protection where the controller is the same, since data collected for one task tend to become attractive for another (Custers and Uršič, 2016). A documented compatibility assessment under Article 6(4) would therefore strengthen the Regulation.
The second issue concerns the position of data subjects, particularly third parties. A privacy notice is provided when a complaint is submitted, or when a substitute declaration is filed in the case of candidates for appointment. Acknowledgement is recorded either by flagging on the Portal or by signing the paper form. The notice expressly refers to the use of AI tools. Such a reference informs data subjects that AI is being used but does not explain how or what the consequences are, which is where meaningful transparency lies (Edwards and Veale, 2017).
Data subjects may exercise the rights set out in Articles 15–22 of the GDPR, subject to the limitations in Article 2-undecies of the Italian Data Protection Code. They may also object to processing under Article 21 of the GDPR and lodge a complaint with the Italian Data Protection Authority. The main critical issue concerns third parties, whose data may be processed without their receiving any direct privacy notice. Article 14 of the GDPR allows for certain exemptions, for instance when providing information proves impossible or would require a disproportionate effort. However, the Regulation does not explain whether or how these exemptions apply.
4. Between transparency and self-assessment
The Regulation deserves credit for making the use of AI by a public authority transparent and for anticipating the intersection between the GDPR and the AI Act. However, four main areas of vulnerability emerge from the analysis above.
First, the robustness of the classification of the systems as not high-risk, which relies on self-assessment, particularly with regard to the outlier identification function.
Second, the internal consistency between the assertion that no personal data are processed and their actual presence in the relevant data flows, including the reliability of anonymisation.
Third, the effectiveness of human oversight in the face of the risk of automation bias.
Fourth, given that the legal basis for processing is a general administrative act, the combination of function creep, long retention periods and limited information to third parties raises particular concerns for criminal offence data and calls for clearer justification in light of the principle of substantive legality.
As for future developments, further guidelines from the Italian Data Protection Authority, the progressive implementation of the AI Act and possible judicial review can be expected to clarify whether the approach adopted by the Bank of Italy will become a model for other public authorities.