AI Act: analysis of Italian Legislative Decree 179/2026 for businesses, public administrations and professionals

1. Responsibility for decisions under the new Italian legislation

A business purchases a system that selects information, ranks priorities and suggests decisions. The contract describes the software’s performance; the sales presentation promises time savings; the relevant department values its speed. The legal issue begins when that result informs a choice affecting a person, a service or the exercise of a power. Who can verify it? Who has the information needed to challenge it? What alternative remains practicable if the result is unconvincing?

Legislative Decree No. 179 of 7 October 2026, published in the Official Gazette on 8 October and due to enter into force on 23 October 2026, should be read with these questions in mind. The authorities’ powers, the supervisory system, penalties, training and employment provisions define the conditions under which the legal system permits a technical capability to contribute to decisions with legal significance. The measure’s importance therefore extends beyond the introduction of new documentation requirements.

The argument I wish to advance is that artificial intelligence compliance must also be assessed in terms of the quality of the organisation using it. A system may be accurate for a particular function and become unsuitable when the context changes. A decision may formally be signed by a person yet depend substantively on an output that the person has been unable to examine. The presence of a human decision-maker is meaningful only if the organisation provides that person with knowledge, time, information and powers to intervene.

This interpretation does not justify constructing obligations other than those laid down in the legal sources. On the contrary, a distinction must be maintained between the applicable provision, the persons to whom it applies, its commencement date and any further organisational choices that may be prudent. Artificial intelligence law is particularly susceptible to confusion between rules in force, legislative announcements, voluntary standards and market expectations. For a director or manager, confusing these levels means spending resources without understanding which risk is actually being managed.

The analysis focuses on Decree No. 179 and its relationship with the European regulation on artificial intelligence, as amended in 2026, with Law No. 132 of 2025, and with the rules that continue to govern personal data, contracts, employment, healthcare and administrative decision-making. It also considers the procedural connection with Legislative Decree No. 160 of 2026. It does not purport to address every issue in the digital economy: its subject is the transition from adopting a system to responsibility for the decisions that the system enables.

For businesses, organisations and professionals, this perspective has practical value. The legal issue can be addressed while the contract is still negotiable, the procedure can still be configured and internal responsibilities can still be allocated clearly. Waiting for an inspection, an appeal or a dispute with an employee often means having to reconstruct retrospectively information that should have accompanied the entire project.

References [1] [2] [3] [16]

2. The European framework and the timetable for obligations

Reading the new national measure requires an initial clarification. The legislative decree does not transpose a European regulation in the manner applicable to directives: it organises its implementation within the domestic legal system in areas left to national competence and coordinates other rules with it. This distinction affects how the obligations are established. For each form of conduct, the applicable European or national source, the person to whom it is addressed and the date from which compliance can be required must be identified. The system cannot be described as a single requirement that becomes mandatory for everyone on the same date.

As at 17 October 2026, the European reference point is Regulation (EU) 2024/1689 as amended by Regulation (EU) 2026/1744 of 8 July 2026, published on 24 July and in force from 27 July. The so-called Digital Omnibus on artificial intelligence is therefore enacted law, not a proposal still under discussion. Article 113 retains 2 August 2026 as the general application date, but defers Chapter III, Sections 1, 2 and 3, except Article 6, paragraph 5: until 2 December 2027 for systems classified as high-risk under Article 6, paragraph 2, and Annex III; and until 2 August 2028 for those falling within Article 6, paragraph 1, and Annex I. It is therefore incorrect to attribute full application of all high-risk obligations indiscriminately to 2026.

Conversely, the deferral does not amount to a general suspension of artificial intelligence law. Account must still be taken of the provisions already applicable, the prohibitions, the rules on general-purpose models, literacy and transparency obligations, together with sectoral rules. Article 50, for example, distinguishes information about interaction with an AI system, the technical marking of synthetic content and the obligations of those disseminating certain content. For systems placed on the market before 2 August 2026, Article 111, paragraph 4, gives providers until 2 December 2026 to comply with the specific marking required by Article 50, paragraph 2. This transitional provision does not extend all transparency obligations wholesale.

The regime for existing systems also requires a separate examination. In certain circumstances, Article 111 links application to high-risk systems already placed on the market or put into service to significant changes made after the relevant Chapter III date; it also establishes specific deadlines for systems intended for public authorities and certain general-purpose models. The assessment of timing therefore requires the history of the product and its use: when it was introduced, what functions it had, what changes have occurred and who decided upon them.

The argument advanced here is that the timetable must form part of the dossier assessing the technological investment’s regulatory status. A contract concluded today may produce effects beyond the new deadlines; an initially limited solution may be extended, updated or integrated into more sensitive functions. Purchasing a system without making provision for documentation, updates, cooperation and compliance costs means postponing a decision that may become more onerous to negotiate once replacing the provider has become difficult.

Two matters must be precisely distinguished: what the law already requires and what it is advisable to design in order to make subsequent use sustainable. Confusing them generates either alarmism or inertia. A legally sound compliance plan makes this distinction explicit, allocates responsibilities and deadlines, and enables decision-making bodies to understand the risks they assume by approving the transaction. Compliance is not a date in a diary: it is also the ability to keep the relationship between the technology purchased and the intended activity manageable.

References [2] [3] [4]

3. Provider and deployer in the value chain

The classifications and obligations examined in this section must be coordinated with the Chapter III commencement dates described in the preceding section; analysing them is already necessary for projects intended to operate beyond those dates.

One of the most significant issues for businesses and organisations is understanding their position in the artificial intelligence value chain. Purchasing a service does not settle the classification. A distinction must be drawn between the system’s provider, the person using it under their authority, any importers and distributors, and operators developing or making components available. Nor does deployer status under the AI Act necessarily coincide with controller status under the GDPR: the two assessments rest on different criteria and both must be undertaken.

The intended purpose is decisive. Classification does not derive solely from the technology used or the sector in which the purchaser operates. A healthcare business may use an assistant to reorganise documents and a system intended for a clinical function: their use within the same organisation does not make the applicable rules identical. Article 6 distinguishes the connection with products subject to harmonisation legislation from the uses listed in Annex III; it provides exceptions for certain functions that do not entail significant risks, subject to conditions that must be documented. Profiling natural persons in the cases covered by Annex III precludes reliance on that exception. The product’s commercial terminology alone neither establishes nor excludes high-risk status.

Article 25 also identifies situations in which a distributor, importer, deployer or other third party assumes the obligations of a provider: placing its own name or trade mark on a high-risk system, without prejudice to the provision’s specific reservation concerning contractual arrangements for allocating obligations; substantial modifications that preserve that classification; or changing the intended purpose of a system not initially classified as high-risk where its new use brings it within that category. The Omnibus clarifies the original provider’s cooperation, including relevant documentation, knowledge of limitations and targeted technical access. However, account must also be taken of the conditions in which cooperation is not required, particularly where the provider has clearly excluded transforming the system into a high-risk use.

A practical contractual consequence follows. A clause describing the customer as a mere deployer does not neutralise an operational arrangement that gives it responsibility for designing a new intended purpose or making a relevant modification. Contracts must be read alongside actual configurations, instructions and internal decisions. Equally, not every customisation automatically turns the customer into a provider: classifying every adaptation as substantial would be the opposite simplification, and equally unhelpful.

Consider, by way of illustration, a company that purchases a generic tool for analysing documents and subsequently integrates it into a recruitment process, giving it the function of ranking candidates. The legal issue changes when what the system determines within the procedure changes. Before extending its use, the AI’s actual contribution, the classification of the use and any potential change in roles must be examined. The same reasoning applies where an administrative platform is progressively used to affect access to benefits or services.

The argument is that compliance requires change management as well as an initial product assessment. The organisation should be able to identify who authorises a new intended purpose, on the basis of what information and with what involvement of the relevant functions. The relationship between the legal function, process owners, security and procurement acquires an operational purpose here: preventing an update presented as routine from inadvertently changing the business’s regulatory position.

Within the system of requirements under Articles 9–15, risk management, data, documentation, records, transparency, oversight and security contribute to verifiable compliance. Article 9 conceives it as a continuous process throughout the lifecycle. The interpretation proposed is that the contract must enable this continuity: access to the necessary information, notification of changes, evidence of checks, incident management and the availability of alternative solutions. These matters must be proportionate to the particular case, not standard wording copied indiscriminately into every IT purchase.

References [6] [4]

4. Human oversight and the right to an explanation

The presence of a person in the procedure is not, on its own, sufficient to establish that the decision has remained human. The issue is whether that person has the knowledge, information, time and authority needed to assess the system’s result and depart from it. Article 14 of the AI Act, within the high-risk regime and according to the relevant commencement dates, requires proportionate oversight measures and expressly considers the danger of automatically relying on the output. Article 26 places responsibility on the deployer to assign oversight to competent, trained persons with effective authority and support.

This architecture does not coincide with the prohibition in Article 22 GDPR. The latter concerns decisions based solely on automated processing that produce legal effects concerning a person or similarly significantly affect them. The provision contains exceptions and safeguards; it does not establish that every use of an algorithm is prohibited or that a general contractual authorisation resolves every issue. The AI Act and the GDPR operate on coordinated but distinct levels: the system’s classification, the lawfulness of processing and the legality of the individual decision require separate assessments.

The SCHUFA judgment of 7 December 2023, Case C-634/21, makes clear the substantive nature of this examination. In the circumstances considered, the Court of Justice treated the production of a probability value by a credit information agency as an automated decision where that value played a determining role in the third party’s subsequent decision. The existence of a different person who formally concludes the procedure is therefore insufficient to interrupt the legal significance of the automated processing. The principle does not permit indiscriminate equivalence: it requires the actual weight of the output in the choice to be established.

The Dun & Bradstreet Austria judgment of 27 February 2025, Case C-203/22, completes the picture in relation to explanation. The data subject must be able to understand which personal data were used and through what procedure and principles they contributed to the outcome. Merely communicating a complex formula does not meet that need. Reliance on trade secrecy does not permit an automatic refusal: where the controller invokes protected information, the competent authority or court must carry out the necessary balancing exercise on the basis of the information made available. Technical complexity does not thereby become an area beyond legal scrutiny.

The interpretation advanced is that oversight must be designed as an organisational capability that can actually be exercised, not as a signature added at the end of the process. In a selection, service allocation or risk assessment procedure, it is necessary to ask what the responsible person can genuinely check, which documents they can consult, which anomalies require further investigation and how disagreement with the output is handled. Requiring oversight without providing access to information may formally transfer responsibility without enabling it to be discharged.

This also entails a different approach to documentation. Retaining only the final result may not allow the decision to be reconstructed; indiscriminately retaining every item of data is not, in itself, a response consistent with proportionality and data minimisation. The necessary evidence, the respective legal bases and retention periods must be identified. The relationship with the provider must enable information useful for oversight and for responding to challenges to be obtained, without confusing the need to explain a decision with a general demand for publication of source code.

On this point, the business’s interest converges with that of the person concerned: being able to demonstrate how a decision was made is also essential to defending the propriety of its conduct. An organisation able to explain and reconsider its choices deals with complaints, inspections and disputes with greater understanding. It is in this reconstruction of the procedure, responsibilities and evidence that legal advice can have an impact before a dispute arises, while maintaining the principle that no organisational safeguard amounts to an abstract guarantee of immunity from liability.

References [7] [8] [9] [10]

4.1. Algorithmic manipulation, psychological conditioning and freedom of decision

Responsibility for decisions also requires consideration of the conditions in which a person’s will is formed. The presence of a human interlocutor, their ability to confirm a proposal or the existence of formally expressed consent do not exhaust that assessment. A system may leave the individual to perform the final act of choosing while, at the same time, interfering profoundly with the process through which that choice is reached. The legal issue is then the effectiveness of self-determination.

Article 5, paragraph 1, point (a), of the European regulation prohibits the placing on the market, putting into service and use of systems employing subliminal or purposefully manipulative or deceptive techniques where the further conditions laid down in the provision are met: material distortion of behaviour, appreciable impairment of the ability to make an informed decision, inducement to make a decision that would not otherwise have been taken, and the causing, or reasonable likelihood of causing, significant harm. Point (b) separately governs the exploitation of vulnerabilities associated with age, disability or a specific social or economic situation, likewise requiring material distortion of behaviour and actual or reasonably foreseeable significant harm.

These prohibitions have applied since 2 February 2025. Their legal basis is therefore directly European; Legislative Decree No. 179/2026, due to enter into force on 23 October 2026, addresses the national organisation of supervision and the imposition of penalties. This distinction also matters when establishing the timing of liability.

The reference to significant harm prevents every influence on choice from being indiscriminately equated with a prohibited practice. Recital 29 refers to adverse consequences for physical and psychological health and financial interests, including harm that may accumulate over time. It also clarifies that legitimate commercial practices, including advertising compliant with the applicable rules, do not in themselves constitute prohibited manipulation. The assessment therefore requires attention to the system’s actual operation, those to whom it is directed and the consequences of the interaction.

The interpretation advanced here identifies freedom in forming a decision as an essential criterion for assessing uses of AI. It becomes relevant to understand whether personalisation helps a person find their way or exploits information about them to weaken their ability to assess matters. This question is particularly significant where the interface presents itself as a trustworthy interlocutor, accompanies the user over an extended period and progressively adapts its responses to their reactions. The issue must be examined in the overall relationship as well as in the content of the individual message.

Consider, by way of illustration, a commercial assistant that uses knowledge of a user’s financial difficulties to exert personalised pressure towards seriously detrimental financial commitments. Or a digital service that exploits fears expressed by a person about their health to steer them towards costly services through misleading representations. These situations must be assessed against every condition of the prohibition: the use of AI or personalisation of communication, considered in isolation, does not permit automatic classification.

For businesses, public authorities and professionals, this perspective suggests that the preliminary assessment should also include the objectives assigned to the system, the criteria used to measure its effectiveness and the conditions in which it interacts with people. If an application is assessed solely on its ability to secure acceptance, prolong interaction or increase purchases, it is necessary to examine the behaviour those objectives incentivise. The issue concerns commercial functions, procurement, service managers and governing bodies, with input from technical and legal expertise.

In contractual and organisational terms, it is advisable to define permitted uses, seek relevant information from the provider about persuasive features and the system’s limitations, provide for proportionate checks, and regulate modifications, reporting and suspension of use. Measures must be designed around the specific service. A general declaration of compliance becomes useful when accompanied by material that allows it to be understood and verified. Likewise, a notice about the use of AI cannot, on its own, be regarded as an adequate response to the risk of manipulation.

The importance of this assessment also emerges from the penalty regime. Article 23, paragraph 1, of Legislative Decree No. 179/2026 provides, unless the conduct constitutes a criminal offence, for penalties of up to EUR 35 million or, if higher, up to 7 per cent of total worldwide annual turnover in the preceding financial year for breaches of the prohibitions in Article 5. The provision refers to the European criteria for determining the amount and reserves for SMEs, including microenterprises, and innovative start-ups the specific criterion in Article 99, paragraph 6, of the regulation. The actual amount therefore requires an assessment of the individual infringement.

Legal advice can contribute as early as the design of these relationships: identifying where the choice is formed, what information influences it, who controls the manner of interaction and what evidence allows it to be reconstructed. From this perspective, protecting personal autonomy also means helping the organisation recognise and manage the risk arising from the way it builds its digital relationships.

5. The national authorities and coordination of powers

The regulation of artificial intelligence raises an issue long familiar to public economic law: the same activity may concern several authorities, without that plurality of responsibilities permitting an undifferentiated overlap of controls. An algorithm used by a bank for a credit assessment, one incorporated into a medical device and one used to manage infrastructure do not acquire the same classification merely because they share a technology. The context of use helps determine both the operator’s obligations and the public authority responsible for verifying compliance.

Legislative Decree No. 179/2026 organises this plurality by distinguishing the notifying function from market surveillance. Article 4 gives AgID responsibility for the procedures for assessing, designating, notifying and monitoring conformity assessment bodies. It therefore does not introduce a general requirement for prior administrative authorisation to use any AI tool. Confusing the two matters would mean imagining a centralised control that the decree does not establish and, for a business, seeking an authorisation where it must instead identify a sequence of obligations distributed among several parties.

Article 5 designates ACN as the market surveillance authority for AI systems, without prejudice to the powers of the European AI Office and specific national competences. The Bank of Italy, CONSOB and IVASS act within their respective remits in relation to the high-risk systems described in the provision; the Italian Data Protection Authority assumes the role of surveillance authority within the limits of Article 74, paragraph 8, of the European regulation. The powers that the same Authority exercises in relation to personal data naturally remain distinct. The fact that a system falls under another authority’s supervision does not remove the associated processing from data protection rules.

ACN also acts as the single point of contact under Article 6. A single contact does not, however, mean that all powers are vested in one body. The legislature entrusts agreements between public authorities and cooperation mechanisms with making a distributed system of supervision workable: Article 7 refers to the agreements under Article 15 of Law No. 241/1990, information sharing and responses to serious incidents; Article 8 governs coordination with sectoral authorities for products also subject to European harmonisation legislation.

The issue has practical legal significance. Effective information exchange can prevent repetitive requests for documents and inconsistent assessments; it cannot, however, obscure the legal basis of the power exercised, the subject of the investigation or the position of its addressee. On a systematic reading, coordination should increase the public authority’s capacity to obtain information without making the procedure less transparent to the operator. This is where effective oversight meets the rights of the defence.

The business should therefore be able to identify which authority is acting, in relation to which system and use, for which obligations and in the exercise of which powers. A generic technological description such as ‘intelligent platform’ does not allow this to be established. A reliable description of the functions, a distinction between provider and professional deployer, identification of modifications made to the system and the connection with the business processes concerned are required. The mapping of public competences depends, to a large extent, on the precision with which private activities are mapped.

The decree also provides a significant safeguard in relation to complaints. Article 18 requires a full opportunity to be heard and provides that a complaint sent to an authority lacking competence must be forwarded by that authority to the competent one, with the complainant being informed. Institutional plurality must not become an unsustainable burden of navigation for those seeking protection. Equally, for the person under investigation, the opportunity to be heard must extend to the system’s classification, the establishment of the facts and the relevance of the proposed measures.

Within this framework, legal assistance is useful even before an inspection. Defining contractual roles, establishing what documentation is available from providers and preparing consistent engagement with the authorities enables the organisation to present its conduct accurately. Technical knowledge remains indispensable; it acquires institutional effectiveness, however, when translated into verifiable facts, identifiable responsibilities and legally relevant arguments.

References [1]

6. Supervision, penalties and protection of the business

Effective regulation depends on the ability to establish what a system does and the conditions in which it is used. For this reason, the AI supervision procedure cannot be confined to examining a contractual declaration of compliance. It is necessary to establish what information was available, what risks could be identified, what instructions were given and what decisions were taken when the system’s behaviour departed from expectations. Compliance thus acquires a documentary dimension encompassing the entire organisational decision-making cycle.

Articles 15 and 16 of the decree govern ACN’s powers, referring both to the AI regulation and to Regulation (EU) 2019/1020 on market surveillance. Remote or on-site checks and inspections, including without notice, are provided for real-world testing. The procedure may be initiated following a serious incident, a complaint that is not manifestly unfounded, a report from another authority or on the authority’s own initiative. The issue may therefore arise outside the contractual relationships within which the business believed it had contained it.

The ability to produce coherent documentation must not be confused with assembling a defensive archive after the event. The decisive element is contemporaneous evidence: system versions, acceptance criteria, user reports, checks performed and reasons for interventions. These elements make it possible to distinguish unforeseeable error from organisational deficiency, an infringement attributable to the deployer from a defect attributable to the provider, and prompt correction from inaction. They do not confer immunity, but enable an individualised assessment.

Article 23 provides for several penalty bands. For breaches of the prohibitions in Article 5 of the AI Act, the maximum is EUR 35 million or 7 per cent of the preceding year’s worldwide annual turnover, whichever is higher; for the obligations listed in paragraph 2, EUR 15 million or 3 per cent; and for inaccurate, incomplete or misleading information supplied in response to a request, EUR 7.5 million or 1 per cent. A further band, up to one million euros or 0.5 per cent, concerns Articles 27 and 86 of the regulation: the fundamental rights impact assessment and the right to an explanation. In each band, the percentage maximum applies where it is higher. For SMEs, including microenterprises, and innovative start-ups, the reference to the special criterion in Article 99, paragraph 6, must be considered. The infringements in Article 23 are formulated subject to the proviso that the conduct does not constitute a criminal offence. These maxima must always be read alongside the conditions of the individual infringement and the application dates of the European obligation concerned.

The amounts do not exhaust the significance of the rules. Article 23 refers to the European criteria for determining the penalty and distinguishes the substantive or formal nature of the obligation breached. For infringements of low harmfulness or risk, Article 22 permits alternatives to a financial penalty: an order to end the infringement and refrain from repeating it or, where it has ceased, a public statement. Failure to comply with the order exposes the person concerned to an increased penalty within the maxima. Cooperation and prompt remediation therefore carry weight within an evaluative procedure, without creating an automatic entitlement to leniency.

Under Article 24, the penalty procedure requires notification of the allegations, reasons and conclusion within a reasonable period. Specific arrangements are left to the authorities’ procedural rules; there is also a residual reference to Law No. 689/1981, insofar as compatible, excluding the reduced payment provided for by its Article 16. The appropriate remedy and competent court must be identified by reference to the authority, the decision and the legislation referred to: it would be misleading to present every AI dispute as following a single procedural route.

Publication of the penalty also requires attention. The decree provides for disclosure of any appeal and its outcome, as well as circumstances permitting anonymisation linked to data protection, cybersecurity risks, criminal investigations and identifiable disproportionate harm. The business’s reputation therefore forms part of the procedure, but must be addressed through the relevant statutory criteria, rather than through a general claim to exemption from transparency.

An effective defence begins with the ability to distinguish the issues: technical challenge, legal classification, proportionality of the measure, preservation of evidence and protection of business continuity. For a regulated business, an essential services operator or a healthcare organisation, a supervisory measure may affect processes that cannot be interrupted without consequences. Legal risk management must therefore connect the response to the authority with organisational solutions that are practicable, allowing both the public interest in safety and the means of achieving it to be assessed.

References [1] [19] [4]

7. Corporate governance of AI uses

The board of directors does not need to become a machine learning laboratory. It must, however, be enabled to understand which business functions depend on AI systems, what consequences may arise from their improper use and who is able to intervene. The gap between technical expertise and managerial responsibility must be bridged by an adequate flow of information, proportionate to the function performed and the size of the organisation.

Article 2086 of the Civil Code provides a systematic reference point for corporate or collective enterprises: the adequacy of organisational arrangements relates to the nature and size of the activity. Integrating tools that affect recruitment, the management of a plant or the terms offered to customers may change that operational nature. What is proposed here is an organisational interpretation of the provision, not the identification of a separate obligation introduced by Decree No. 179 to establish an AI committee in every company. The choice of structure must remain proportionate to actual needs.

A first mistake is to assign a single function responsibility for resolving disparate legal issues. The IT manager assesses integration and operation; the security manager considers vulnerabilities and continuity; the DPO, where appointed, performs the tasks assigned by the GDPR with the corresponding independence; the human resources department understands employment decisions; and the process owner knows the consequences of an error. Legal advice connects these assessments with the system’s classification, the applicable legal sources and the measures the organisation must adopt.

The operation of information flows also requires a choice. A report stating only that the system is compliant says little. Senior management needs to know the permitted uses, relevant limitations, dependencies on the provider, events requiring a fresh assessment and conditions in which use must be discontinued. For the most significant projects, these elements can accompany the investment resolution and subsequent review. Operational functions must be able to report a concern without the report being perceived as a failure to be concealed.

Consider a system initially purchased to summarise complaints. If it is subsequently used to prioritise action or suggest rejecting requests, the relationship between information and decision changes. The original assessment may no longer be sufficient, even if the software’s commercial name remains the same. An adequate organisation must make this transition identifiable and subject it to an appropriate assessment.

Useful documentation arises from this need for understanding. Inventories, instructions and minutes have value when they describe what happens and allow it to be corrected. Extensive documentation disconnected from actual uses may make responsibility harder to identify and weaken the business’s defence. AI governance requires, first and foremost, a reliable representation of the processes into which the technology has been introduced.

References [11] [6] [8]

8. The contract with the provider as a condition for oversight

A significant part of the ability to comply is determined by the contract. A deployer without access to the information needed to understand the system’s limitations, preserve evidence or manage an update may face obligations it is materially unable to fulfil. Negotiations must therefore also address the conditions for oversight, alongside price, licences and service levels.

The first issue is the agreed function. The description of the intended use should identify processes, persons concerned, permissible data and expected results, specifying which uses require a new agreement or assessment. An all-embracing formula presenting the system as suitable for any business decision makes the allocation of responsibilities uncertain. The contractual framework cannot replace classification under the AI Act, but can provide the information needed to undertake it and keep it up to date.

Changes to the service must then be addressed. The provider may modify the underlying model, parameters, sources that can be consulted or the location where data are processed. Continuity of the interface may conceal a substantive discontinuity in performance. For sensitive uses, it is reasonable to negotiate prior notifications, documentation of changes, proportionate checks and remedies where an update makes the system incompatible with the authorised use. These are contractual choices requiring justification, not a mandatory clause identical for every service.

The availability of evidence is equally important. The contract should clarify which records are generated, who can access them, how long they are retained and how they are extracted if a challenge arises. The period must be coordinated with the applicable rules, the purpose of processing and personal data minimisation. Retaining everything indefinitely is not a legally neutral solution. Conversely, automatically deleting the only available evidence may make a defence impracticable.

Data issues require separate provisions. Data protection roles depend on the actual activities; the designation used in the contract is not decisive. The business must understand whether documents, requests and results are used to train the service further, whether subcontractors can access them and what transfers may occur. Lawful possession of a document does not always entail the right to use it for any form of training. Personal data protection, professional secrecy, contractual confidentiality and intellectual property rights must still be assessed.

Liability clauses must be read alongside operational remedies. An indemnity may govern the financial relationship between the parties within permitted limits, but does not transfer public powers, the employer’s obligations or the responsibility of a person making a professional decision to an external party. The right to suspend use, obtain assistance following an incident and migrate data and configurations may have greater immediate value than a general promise of compliance.

Article 25 of the AI Act, within its scope and application timetable, makes managing responsibilities in the value chain particularly important. It is advisable, even at the negotiation stage, to examine whether customisation, changes of intended purpose or marketing under a different name could result in a different classification of the role. When a company purchases a technological capability, it also chooses the information and counterparties on which the legal soundness of its processes will depend.

References [6] [8]

9. Data, algorithms and mathematical methods as trade secrets

Article 52 of Decree No. 179 inserts into Article 98 of the Industrial Property Code a paragraph 1-bis concerning data, algorithms and mathematical methods for training AI systems. The reference includes model architectures, optimisation functions, training procedures and configurations, as well as the further technical and computational elements specified in the provision. The rule expressly recognises the potential economic value of knowledge that does not correspond to a physically identifiable product.

This inclusion is subject to the conditions already laid down in paragraph 1: secrecy, economic value deriving from secrecy, and the adoption of reasonably adequate measures to preserve it. The decree does not confer a general monopoly over algorithms or turn every item of business data into a trade secret. A known method, freely accessible information assets or an unprotected dataset do not automatically satisfy the requirements for protection merely because they are used in AI.

The most interesting doctrinal aspect concerns the relationship between value and organisation. An intangible asset is also protected through the way the business identifies it, safeguards it and regulates access to it. Organisational measures thus help make the assets to be defended legally identifiable. It is not enough to claim, when a dispute arises, that a body of information was important: it must be possible to explain which information was confidential and how that confidentiality was actually maintained.

For a business developing or customising systems, the review should distinguish source data, processing outputs, configurations, documentation and results, identifying ownership or lawful control and usage restrictions for each component. Protection may require tiered access, confidentiality agreements and rules for employees and other personnel; these tools must reflect the reality of the activity. A general prohibition in a contract does not always compensate for the indiscriminate circulation of information inside or outside the business.

The relationship with transparency cannot be resolved by asserting the absolute priority of either interest. Trade secrecy is not a general exemption from obligations towards authorities and persons protected by the legal system. On the other hand, explaining a decision does not necessarily entail publishing code, weights, training data and know-how in full. The relevant information and the form of access capable of making protection effective must be identified, while safeguarding what need not be disclosed.

Article 53 confirms the jurisdiction of the specialised business divisions of the courts over the disputes referred to in Article 98. The connection with procedural tools for protecting confidentiality and obtaining evidence, examined below, demonstrates the need for a coherent strategy. The business must be able to establish its right to secrecy and, at the same time, enable the required checks. Addressing these two objectives in separate departments without a common plan may produce incompatible clauses and avoidable difficulties in defending its position.

References [1] [12] [10] [16]

10. Employment and effective decision-making power

Article 41 of the decree identifies the final decision on the employment relationship as a point at which responsibility cannot be dissolved into the result of a calculation. Without prejudice to Article 22 GDPR, the provision requires decisions concerning the establishment, modification and termination of the relationship, including disciplinary decisions and performance assessments, not to be taken solely through automated processing. The final decision is reserved to a natural person with effective and independent power. The employee’s consent does not constitute an exception to this prohibition.

The most significant point is the description of human power. In my view, a merely formal intervention does not deliver the safeguard described by the provision. It is not enough for a manager to confirm on screen the solution proposed by the system if they lack the information, time and authority needed to depart from it. Effectiveness requires the organisation to make an individual assessment possible; independence means that the algorithmic result must not become, through internal procedures or incentives, a determination that is binding in practice.

This criterion does not require analytical tools to be abandoned. It requires their place in the decision to be clarified. Assessment may be assisted by indicators, comparisons and forecasts; it remains necessary to distinguish the indicator from the fact, correlation from the justifying reason, and measured performance from conduct actually attributable to the person. A uniform score may be particularly misleading if derived from data concerning tasks, working conditions or operational opportunities that are not comparable.

The text nevertheless contains a limitation that must not be omitted: activities involving the search for and selection of candidates do not fall within final decisions concerning the establishment of the relationship, even where they result in exclusion from subsequent stages. It is therefore incorrect to derive from Article 41 an indiscriminate national prohibition of all automated selection. It would be equally incorrect to turn this exclusion into an unregulated area. Data protection rules, safeguards against discrimination and the European provisions applicable to the systems used must still be assessed according to their respective timetable.

The distinction raises an important issue of interpretation. In a multi-stage process, preliminary exclusion may decisively affect access to employment without coinciding with the recruitment decision reserved to a natural person. Legal classification cannot therefore depend solely on the name given to the stage in the software. Its actual function must be established and the safeguards arising from the different legal sources assessed separately. This interpretation respects the limit expressly set by the legislature without attributing to it an exonerating effect that the text does not contain.

Paragraph 3 connects organisational authority with the intelligibility of the decision. In addition to information before processing, the employee may obtain, on request and through a natural person, intelligible reasons including the influence of AI and the main parameters considered. An explanation does not necessarily mean handing over the source code. It must, however, enable the employee to understand how the system contributed to the specific case: a list of general formulas or a mere commercial description of the product would not fulfil the purpose of the safeguard.

The provision rendering decisions taken in breach of paragraph 1, and specifically dismissal, void makes advance planning essential. Before introducing a tool for assessing or managing personnel, the business should establish who decides, what checks they perform, how anomalies are handled and how a record of human examination is retained. The procedure must be designed with human resources, technical managers and data protection functions: purchasing the product does not exhaust the organisational choice, and the contract with the provider does not transfer the employer’s duties in full.

Article 42 extends attention to the practical organisation of work. Systems affecting work rates, working methods and processes relevant to safety must be considered in the risk assessment required by Article 28 of Legislative Decree No. 81/2008; information and training on specific risks are also required. In logistics, healthcare, waste management or industrial plants, AI may therefore affect both the assessment of the person and the conditions in which they work. Legally adequate governance must be capable of understanding both dimensions.

References [1] [8] [9]

11. Literacy and the quality of judgement

Training in artificial intelligence is often presented as acquiring skills in using the tools. The law asks a further question: are the organisation and its professionals able to recognise the limits of use, assess the consequences of outputs and identify decisions that must remain subject to independent judgement? Literacy matters precisely at this point, because the ability to obtain an answer from the system does not demonstrate the ability to verify it.

The starting point must be the updated text. Article 4 of the AI Act, as amended by the 2026 Omnibus, requires providers and deployers to take measures supporting the development of literacy among staff and others acting on their behalf. The assessment takes account of knowledge, experience, education, training, the context of use and the persons concerned. The provision clarifies that there is no requirement to guarantee each individual a particular level of literacy. It also provides for support from the Commission and Member States, with particular attention to small and medium-sized enterprises. It is incorrect to reproduce the original wording while ignoring this amendment, or to infer from it that training has become irrelevant.

The European approach distinguishes the adoption of organisational measures from the guarantee of an individual cognitive outcome. This does not mean that any course, regardless of its content, satisfies the statutory requirement. The argument advanced here is that the quality of the measures must be assessed in relation to the tasks assigned to people and the risks of actual uses. A common introductory module may provide a starting point; it does not necessarily complete the preparation of those selecting a provider, authorising the use of confidential data or overseeing a decision intended to affect others.

For professionals, the issue is particularly sensitive. Using a generative assistant for preliminary research, document synthesis or preparing a draft does not remove the need to verify sources, facts and the relevance of the argument. The result may be linguistically convincing and legally inadequate. The competence required therefore includes the ability to identify what the tool has not established: the authenticity of a reference, whether a rule remains current, and whether a cited judgment supports the proposition attributed to it.

The connection with oversight of high-risk systems reinforces this interpretation without conflating the different commencement dates. Article 26, paragraph 2, links training with the competence and authority of the person entrusted with oversight. Knowledge without power to intervene is insufficient; power without the necessary preparation risks becoming merely apparent approval. As a criterion for organisational design, a relationship can be drawn between training, allocation of tasks and availability of information: none of the three can permanently compensate for the absence of the others.

A proportionate programme should therefore start from authorised uses and duties. Senior governing bodies need to understand the investment, the allocation of risks and the information needed to decide; procurement needs to assess the provider’s documentation and commitments; operational functions need to recognise errors, limitations and escalation needs; and professionals need to retain control over sources, confidentiality and intellectual work. This differentiation is a methodological proposal, not the attribution to Article 4 of a mandatory curriculum that it does not contain.

Evidence of the activities undertaken must also remain true to their substance. A register of training initiatives is useful but does not replace assessment of their relevance; a certificate does not turn improper use into lawful use. For businesses, public authorities and professional organisations, the value of training lies in its potential to translate into better-informed decisions: knowing when to use the system, when to seek further examination and when to stop automation. It is this ability to preserve judgement as the tool changes that makes training a component of the quality of the activity and the trust of those who rely on it.

References [5] [7]

12. The professions, training and fair remuneration

Within the decree’s framework, training in artificial intelligence becomes a structural component of professional competence. Article 47 includes literacy and training programmes in the initial and continuing courses provided by professional bodies. The content cannot be reduced to learning commands or familiarity with an interface: it encompasses how systems work, their potential and limitations, the legal rules and professional ethical responsibilities, duties to inform clients and the human-centred principle.

The integration of these levels matters. Exclusively technical knowledge may leave the question of permissible use unresolved; exclusively legal knowledge may overlook the ways in which error occurs. The need for training emerges most clearly when a result appears plausible and well expressed. At that point, the professional must be able to distinguish presentational quality from reliability, recognise the need for external verification and decide what information may be entered into the system without compromising the interests entrusted to their care.

The decree provides for a minimum number of hours for each continuing professional development assessment period, but does not determine a uniform number of hours or credits for all professional categories. National councils, professional colleges and competent bodies must amend their regulations or organisational instruments within six months of entry into force, following sectoral procedures. For unregulated professions, the provision addresses programmes promoted by umbrella bodies of professional associations where AI use is relevant to the activity. Presenting an identical numerical obligation for every professional today would therefore be a simplification unsupported by the text.

Training does not replace responsibility for the engagement. Attendance at a course attests to one component of preparation; it does not establish that the system is suitable for a particular activity or that an individual result has been checked. Professional verification retains a concrete subject: the relevance of the data, the accuracy of references, the adequacy of the argument and the compatibility of the use with the instructions received. From this perspective, literacy should increase the ability to ask questions of the system and, above all, to exercise judgement over its answers.

The client relationship therefore assumes an importance preceding the choice of tool. It is necessary to understand which activities can be assisted, which require particularly intensive personal verification and what conditions of confidentiality are needed. An internal policy can make these safeguards consistent: permitted tools, access arrangements, review criteria, error management and data retention rules. These are organisational choices to be proportionate to the profession and engagements, not a standard formula capable of replacing individual judgement.

Article 49 addresses fair remuneration by providing that AI use may lead to its adjustment under Law No. 49/2023, according to parameters taking account of the system’s risk classification. The decrees setting those parameters are to be supplemented within twelve months of entry into force. The rule does not establish an automatic discount for AI-assisted services and does not permit the inference that time saved corresponds entirely to a reduction in professional value.

The criterion chosen merits critical reflection. The system’s risk class follows a regulatory logic concerning the possible effects of its use; the value of the service also depends on the complexity of the problem, the responsibility assumed and the verification work. These are not interchangeable measures. Future parameters must be interpreted and applied without turning technological classification into an exhaustive measure of the intellectual quality of the work. Automation may reduce certain routine activities while simultaneously making competent checking of a result produced with less practical effort more important.

A consequence follows for businesses, organisations and professionals defining engagements: clarity about the scope of the service acquires further value. Identifying expected outcomes, required checks, responsibilities and the manner in which tools are used allows remuneration to be discussed on the basis of the actual service. The central issue remains the quality of the professional decision and protection of the interest entrusted to the professional, even where technology transforms the process leading to that decision.

References [1]

13. Administrative decisions and the legality of fact-finding

The use of artificial intelligence in public administration raises a question preceding technological classification: how can the exercise of a public function remain identifiable when part of the fact-finding is entrusted to a system capable of selecting information, assigning priorities and formulating proposals? Procedural speed is an important interest; the ability to understand and challenge the reasons is a condition of the procedure’s legitimacy.

Article 14 of Law No. 132 of 2025 connects AI use with efficiency and service quality, requiring its operation to be ascertainable and its use traceable. Algorithmic support does not absorb the responsible person’s decision-making autonomy. This provision offers a precise interpretative key: the presence of an official in the organisational chart or a signature on the decision is insufficient to establish that the decision has remained effectively human.

Consider an application that reorganises observations submitted in an authorisation procedure. If the reorganisation systematically omits minority representations, the problem arises before the final decision. Incomplete fact-finding may steer discretion without the system having formally adopted any decision. Review must therefore also cover what the output does not show: excluded documents, aggregated information, relevance criteria and margins of uncertainty. Automation of fact-finding deserves attention precisely because it may appear neutral while determining the framework within which the decision-maker exercises their power.

As a matter of interpretation, a distinction follows between a technical explanation and the reasons for an administrative decision. The former concerns how the tool works; the latter must make intelligible the relationship between established facts, interests considered, the provision applied and the choice made. A reliability percentage or ranking of alternatives does not replace that relationship. On the other hand, the reasons need not necessarily reproduce every computational step: they must enable an understanding of the weight given to the algorithmic contribution and why the authority considered it usable in the particular case.

The case file should therefore allow at least the purpose of use, the documents actually considered, the system version used, checks on responses and corrective interventions to be reconstructed. This is an organisational proposal to be calibrated to the procedure, avoiding both indiscriminate recording of data and loss of decisive evidence. Document management becomes part of the design of the decision: when a challenge arises, it is not sufficient that the system produces a plausible result today; it must be possible to understand how the choice adopted at the time was formed.

Legislative Decree No. 179 of 2026, entering into force on 23 October, addresses this foundation through Article 40. Public-sector staff training is linked to organisational innovation, a minimum competency framework to be defined by guidelines and the institutional contribution of SNA and Formez PA. The decree does not make procedural rules redundant: it strengthens the professional conditions needed to apply them in a different technological environment.

For a public authority, creating these conditions means assigning verifiable tasks to the officer responsible for the procedure, technical departments, the digital function and those responsible for data protection. For the business subject to the exercise of public power, it means being able to identify the stage at which an inaccurate algorithmic representation affected the fact-finding and prepare a relevant challenge. Legal advice thus performs a preventive and, where necessary, contentious role: making the procedure open to scrutiny before technology turns a gap in documentation into a problem of legal protection.

References [13] [1]

14. Public procurement and technology governance

In AI procurement, the subject matter of the contract is not exhausted by access to an application. The public authority purchases a processing capability that will be incorporated into its processes, affecting the allocation of responsibilities and preservation of evidence. It is therefore necessary to describe the expected functional result together with the conditions under which that result may be used.

Article 30 of Legislative Decree No. 36 of 2023 places automation within the public contract lifecycle and connects it with principles of ascertainability, comprehensibility, the non-exclusive nature of algorithmic decision-making and non-discrimination. When purchasing or developing solutions for this area, contracting authorities and concession-granting entities must ensure the availability of source code and documentation useful for understanding its logic, and include the assistance and maintenance clauses needed to correct errors and unwanted effects. The requirement concerns automation of the contract lifecycle: it does not amount to a general obligation to publish the code of any technology purchased by public authorities. Its rationale is not equivalent to a generic clause declaring the product compliant with legislation.

An operational consequence follows from this approach: the competitive comparison must also address the ability to govern the system after award. A solution that is inexpensive to purchase may become costly if it prevents data export, requires continual additional services to document a result or does not allow a function to be suspended without interrupting the entire service. The legal cost of technological dependence should form part of procurement design through requirements relevant and proportionate to its subject matter.

The specifications may define permitted and excluded uses; information to be made available to users; timeframes for reporting anomalies; acceptance criteria and periodic checks; conditions for updating the model; and arrangements for exporting data and records. They may also regulate the involvement of subcontractors, incident management and assistance in responding to requests from authorities. These are design recommendations: their formulation must suit the system and the contract, without turning every procurement into an indiscriminate obligation to hand over all of the provider’s technical assets.

Subsequent modifications are particularly important. An update may improve average performance while changing the system’s behaviour in a specific procedure. It is necessary to establish who authorises it, what tests precede it, which documents are updated and how to revert to the previous configuration if the result is unacceptable. Maintenance is not a minor detail of the relationship: it may alter the technical premise on which the lawfulness of use was assessed.

Acceptance testing should therefore assess situations relevant to the service, including difficult ones: incomplete documents, contradictory inputs, out-of-scope requests and unreliable information. A positive test on ordinary cases proves little if the principal risk lies in handling exceptions. Human intervention must also be testable: who can correct the response, within what timeframe and with what information? A clause providing for oversight has value only if it corresponds to a power that can actually be exercised.

Classification under the AI Act remains a separate matter. Public procurement does not automatically make a system high-risk: its actual function and the regulation’s conditions are relevant. In the meantime, deferral of specific European obligations does not suspend the rules already in force governing contracts and procedure. An immediate field of work emerges: designing procurements in which performance, checks, documentation and remedies are consistent, so as to protect service continuity and the authority’s ability to answer for its decisions.

References [14] [6] [4]

15. Healthcare and the conditions for clinical autonomy

AI in healthcare makes particularly clear the limits of a purely formal conception of human oversight. A doctor may retain the power to confirm or reject a recommendation without having the time, information or tools needed to assess it. In that case, the issue does not concern individual conduct alone: it also encompasses the conditions in which the healthcare organisation introduced the system.

Article 7 of Law No. 132 of 2025 reserves decisions in prevention, diagnosis, care and treatment choices to the medical profession, requires information about AI use and attention to the reliability and updating of systems and data. It also prohibits the introduction of these tools from determining access to services according to discriminatory criteria. These principles must be read together, because the appropriateness of an individual choice also depends on the fairness of the pathway that allowed the patient to reach it.

The organisational question can be framed as follows: who verifies that the purchased system is suitable for the clinical context in which it is used? The performance stated by the provider must be considered in relation to the patient population, available data, internal procedures and the role assigned to the output. A model intended to support an assessment should not be transformed, through organisational practice, into a filter that automatically excludes further investigation.

It is advisable to distinguish at least three functional responsibilities. Management defines the conditions for introduction and the resources for checks; technical departments oversee integration, continuity, security and the quality of data flows; and the professional interprets the result in the context of the person receiving care. This allocation is a methodological proposal, not a framework capable of determining responsibility for every injury in the abstract. Healthcare liability requires the conduct, organisation, causal link and rules applicable to the case to be established.

Decree No. 179 expressly recognises the organisational level in Article 51, concerning continuing medical education (ECM) and healthcare management training. The programmes include technical, legal and professional ethical aspects; the proportion of the three-year training requirement and the management programmes are to be defined by subsequent measures, within six months of entry into force, by the National Commission for Continuing Education and AGENAS respectively. It would therefore be incorrect to present as already determined credits or percentages that the decree leaves to those measures.

Useful training should enable uncertainty to be recognised, situations outside the intended use to be identified, disagreement with the output to be documented and a review to be initiated. Its value does not lie in obtaining generalised trust in the tool. It lies in making trust in the individual use reasonable. Patients too must be able to understand the role of technology in their care pathway, without being overwhelmed by a computing explanation or given reassuring but empty formulas.

The healthcare organisation’s perspective is therefore integrated: purchase, validation in the context of use, training, information, monitoring and management of adverse events. When these stages are treated separately, gaps tend to emerge only after a challenge. Advice accompanying the organisational decision can instead make the technology contract, internal protocols and clinical documentation consistent. This is where innovation meets the protection of health and the legal sustainability of the organisation.

References [13] [1] [18]

16. Five applications in regulated and strategic sectors

The following cases are hypothetical and do not describe professional engagements. They demonstrate why the rules cannot be applied through a simple sector label. Falling within the scope of NIS2, public control of a company and the strategic importance of an activity do not, on their own, coincide with the classification of AI as high-risk. That requires assessment of the system’s function and the conditions in Article 6, Annex I and Annex III, with attention to the application timetable.

Energy. An operator introduces one tool to forecast consumption and another to support network safety. The two applications must not be conflated. For the second, the category of safety components in infrastructure management may be relevant, while the first requires a separate assessment. The project must distinguish commercial forecasting, operational decisions and the protective function: a statistically useful result is not necessarily suitable for supporting a command affecting continuity of supply. The contract should reflect this differing intensity of risk, particularly in relation to errors, unavailability and updates.

Communications. A business uses an assistant to respond to users and a system to analyse anomalies in digital infrastructure. Analysis of the intended purpose is decisive: the regulation covers certain safety functions of critical digital infrastructure, not every software application used by a telecommunications operator. Customer service, in turn, raises different issues: accuracy of contractual information, the recognisability of automated interaction, complaints handling and the ability to reach a human operator. The coexistence of these tools requires different rules, even where the technology provider is the same.

Waste management and publicly owned enterprises. A company with public-sector shareholding adopts AI to optimise collection routes and allocate staff shifts. The first use does not become high-risk merely because the service concerns the environment; the second may require specific examination of the categories relating to worker management. Compliance with service standards, employment law, safety and data protection obligations remain distinct. Optimisation might reduce mileage but impose inappropriate burdens on some employees or make the service less accessible in certain areas. The project must state which objectives may be optimised and which limits are non-negotiable.

Healthcare. An organisation uses one system to organise inventory and another to support triage. The organisation’s identity does not remove the difference between the uses. A system for selecting patients in emergency healthcare encounters a specific provision in Annex III; an administrative application requires a different analysis. Classification as a medical device also follows its own conditions: it does not derive merely from use in a hospital. Separating the applications allows checks and expertise to be directed where an error may have the greatest impact on personal rights.

Defence and national security. The exclusion in Article 2 of the AI Act concerns systems insofar as they are used exclusively for the purposes covered by the provision; it does not grant a general exemption to every business in the defence supply chain. A product also intended for civilian uses therefore requires analysis of its actual intended purposes and the associated responsibilities. For an industrial project, this suggests separating configurations, users, access and changes of intended purpose contractually. The system’s classification must follow its evolution: a solution initially developed for a limited use may require different legal safeguards when it enters other markets.

In each example, the law contributes by asking questions that change organisational and contractual choices: what result is sought, who can intervene, what error is intolerable and what evidence will be needed if the decision is challenged. This work is best undertaken while the project can still be modified.

References [2] [6] [18] [4]

17. The connection with cybersecurity and the NIS2 framework

An altered result may arise from a model error, inadequate data quality or hostile intervention. From the perspective of a business decision, the consequences may be similar; legally, the causes, obligations and responsibilities must be established. AI governance must therefore communicate with cybersecurity safeguards without conflating frameworks whose personal and material scopes differ.

Falling within the scope of NIS2 does not mean that every AI tool in the organisation is high-risk under the AI Act. Conversely, using a system classified as high-risk does not, on its own, subject the entire business to Legislative Decree No. 138 of 2024. The assessments must be undertaken separately and then coordinated, taking account of the sector, size, any designations and the characteristics of the use.

For entities to which Decree No. 138 applies, Article 23 assigns administrative and management bodies tasks relating to approval of arrangements for implementing risk-management measures, supervision and training. The CISO and technical departments contribute to their implementation in accordance with their respective functions. Introducing AI may affect providers, external dependencies, access authorisations and service continuity: relevant information must reach the appropriate decision-making level.

The organisation can coordinate checks while avoiding duplication. A shared inventory of services and dependencies can inform both security assessments and the classification of AI uses. A single change-management process can identify events requiring a technical assessment, a data protection assessment or a legal review. The conditions and deadlines for notifications under the different frameworks remain distinct: sending a communication to one authority does not automatically satisfy every other obligation.

Supply chain security deserves particular attention in essential services. The issue may concern a plant maintenance contractor, a component supplier or the operator of an outsourced process using AI, as well as the IT provider. This does not mean indiscriminately imposing every NIS2 obligation on every counterparty. It means assessing relevant dependencies and vulnerabilities and translating them into proportionate measures, verifiable responsibilities and contractual conditions consistent with the risk.

For a healthcare operator or environmental services provider, advice can therefore address coordination between senior management, service managers, the CISO and the legal function. The objective is to identify who decides, who checks and who must be informed when the system changes its behaviour. Cooperation between areas of expertise becomes effective when each function knows what information it must produce and which decisions fall to it.

References [15]

18. Regulatory experimentation and the authorities

Articles 26–31 of Decree No. 179 govern the Italian AI regulatory sandbox and its coordination with other experimentation environments. Joint management by AgID and ACN reflects the attempt to connect technological development with regulatory assessment. The interesting feature is the express recognition of regulatory learning: experience from projects must also contribute to the authorities’ knowledge and the assessment of the rules.

Access should not be presented as a possibility already available at any time for any project. Operating arrangements, eligibility criteria and operational matters require the measures provided for by the decree. When an application is submitted, the measures actually adopted will need to be checked. The text identifies providers and prospective providers as the intended participants in the controlled environment and provides for attention to access for SMEs and start-ups within the relevant European framework.

For a business, legal preparation for experimentation begins with the question to be put to the authority. A project promising, in general terms, AI useful to healthcare or public services offers little basis for assessment. The function, context, people involved, necessary data and risks to be measured must be defined. The plan must make it possible to understand when the experiment has produced sufficient information and in which circumstances it must be suspended or modified.

Article 28 assigns the operational group responsibility for examining, selecting, approving and monitoring plans, also with reference to records, audits and the traceability of changes. Where the project processes personal data, the prescribed coordination with the Data Protection Authority applies; universities and research centres may be involved for specific needs. These connections make it inappropriate to regard experimentation as mere permission to try a technology without responsibility.

The European framework retains civil liability for damage and defines the conditions governing penalties during experimentation. Any benefit available to those who comply with the plan and the authority’s guidance does not amount to a universal exemption. Nor should exiting the sandbox be presented as indiscriminate certification of every future use of the system. Findings must be related to the project actually examined and the relevant conformity assessment procedures.

Article 30 also addresses dual uses, judicial activities, regional and local sandboxes and coordination with financial experimentation. This reveals an advisory role that does not consist in promising privileged access to institutions. It consists in making the project intelligible, identifying the appropriate route, preparing a coherent supporting assessment and translating the guidance received into practicable changes. The quality of engagement depends on the quality of the dossier.

References [1] [17]

19. Civil liability and access to evidence

The framework established by Legislative Decree No. 179/2026 must be read alongside Legislative Decree No. 160 of 9 September 2026, already in force from 30 September, which devotes Articles 16–20 to procedural tools for compensation for damage caused in the use of AI. The two frameworks operate on distinct but connected levels: the organisation of supervision does not exhaust protection for the injured person; administrative compliance does not subsume the assessment of responsibility for damage. It would therefore be incorrect to attribute to Decree No. 179 the new civil law rules introduced by the earlier measure.

The first problem is information asymmetry. A person affected by an AI-assisted decision may know its effects but lack the material needed to reconstruct the system’s operation and human intervention. Article 17 allows the court, at the injured person’s request, to order the opposing party or a third party to disclose specifically relevant evidence. Facts and material making the claim plausible are required, including in relation to the connection between the system’s result and the alleged damage. This is therefore not indiscriminate access to another party’s technological assets: the application must establish an identifiable evidential issue.

The provision covers logs, risk-management documentation, relevant technical information and information about the parameters and arrangements for human oversight. It is significant that the evidence may concern both the machine and the organisation using it. In practice, reconstructing the system version, the incident reported and the decision to accept or correct a result may be more useful than abstract access to the code. This is an operational consequence of the rules, not a uniform requirement to retain every item of data produced by any application.

Failure to cooperate entails specific consequences. Unjustified non-compliance by a party, even partial, allows the court to draw evidential inferences; where it concerns the documentation specified in paragraph 2, the court, having assessed all other material, treats the facts alleged by the applicant as admitted. A third party failing to comply without justification may be subject to a financial penalty. Actual unavailability of documents and refusal to produce them are not equivalent situations; establishing either nevertheless requires verifiable material. A business purchasing a service without providing for access, export and evidential assistance risks discovering the contract’s limitations only when it needs to defend itself.

Article 18 expressly introduces a rebuttable presumption of causation for damage resulting from breach of AI Act obligations. Its scope must be coordinated with Article 16, paragraph 2: not every instance of damage occurring in the presence of AI benefits indiscriminately from this rule. The presumption concerns the link between breach and damage; it does not permit automatic presumption of the damage, the existence of every breach or the liability of any person in the supply chain. The wording will need to be applied by identifying the obligation actually relevant, the person to whom it is addressed and the harm alleged. The presumption also requires an obligation that is actually applicable: it does not bring forward European provisions whose application has been deferred. These stages will be the focus of the interplay between allegations, evidence and rebuttal evidence.

Conversely, Article 19 clarifies that even certified compliance does not, on its own, exclude liability. Certification retains its significance, but does not necessarily answer the different question of whether that system was used correctly in that context and in relation to that person. For legal assistance, the point is decisive: documentation must enable the organisational choice and actual conduct to be explained, not merely certificates to be produced.

Protection of trade secrets remains part of the proceedings. The order must be necessary and proportionate; the court protects confidential information, including through Article 121-ter of the Industrial Property Code. Secrecy therefore does not confer evidential immunity, nor does a need to defend a claim authorise generalised disclosure. Both an application for disclosure and opposition to it require precision: which information is needed, to prove which fact, with what restrictions on access and use.

The framework is completed by the direct action against the insurer under Article 20, within the limits of cover, with contractual defences arising before the insured event available and the necessary participation of the person alleged to be responsible. This provision does not introduce a general insurance obligation. In organisational terms, it does establish the value of checking, before the system is used, whether the cover addresses the actual risks. Mapping obligations, contracts, availability of evidence and insurance thus belong to the same preparation for the decision: advice also prevents litigation by enabling, where necessary, a defence grounded in facts.

References [16] [8] [12]

20. A compliance programme grounded in actual decisions

Compliance work can begin with a focused question: which decisions in the organisation currently involve an AI-generated result? The answers may reveal uses other than those purchased centrally. A department uses an assistant to draft customer responses; a provider uses a model to identify anomalies; the human resources directorate receives a ranking from an external platform. The review must encompass actual uses and those embedded in services, while respecting workers’ rights and the rules applicable to monitoring.

The next stage is classification. For each use, the function, context, persons concerned, organisation’s role and applicable rules must be identified. The conclusion may be that a function does not constitute an AI system within the regulatory definition, falls within a prohibition, requires transparency or belongs to a high-risk area with deferred obligations. Reasons for the classification are useful even where they lead to exclusion of an obligation: they allow the choice to be reconsidered if circumstances change.

The action plan must reflect these differences. From 23 October 2026, measures necessary for an employment decision governed by Article 41 cannot be postponed by invoking the European high-risk deadlines in general terms. Conversely, provisions requiring implementing regulations must not be turned into invented quantitative requirements. The organisation may adopt preparatory safeguards where useful, distinguishing them from requirements already mandatory and identifying the conditions that will require them to be updated.

A governance dossier for each use can bring together a description of the function, legal assessment, relevant technical checks, agreements with providers, instructions for users and arrangements for managing events. It is not a new document named in legislation and universally required: it is a methodological proposal to be adapted to the size and complexity of the activity. Its value lies in making accessible what is needed to decide, verify and, if necessary, defend the choice.

Training must be connected with authorised uses. Staff need to know what data they may enter, when a result requires corroboration, how to report an error and who is responsible for the final decision. Senior management must be able to assess the project’s risks; operational managers must be able to recognise deviations. A mere certificate of attendance at a course does not establish that the organisation knows how to govern its system.

Finally, review is necessary. A significant update, change of intended purpose, new category of persons concerned or a challenge may require further assessment. Providing for this possibility in the contract and organisational arrangements avoids treating as definitive an assessment made under conditions that no longer apply. The cost of compliance thus becomes a component that can be planned within the project, instead of emerging suddenly when the system is already indispensable.

References [1] [4] [5] [6]

21. Legal advice and responsibility for choices

Decree No. 179 makes the relationship between technological capability and legal responsibility more visible. The system of authorities and penalties represents its public dimension; employment relationships, training and protection of information assets reveal its consequences within organisations. The European and national rules must be read together, preserving the precision of their respective scopes and timetables.

I believe the deeper issue remains the possibility of making a decision that is truly one’s own. A signature is insufficient where the decision-maker lacks the information needed to disagree. A declaration of compliance does not resolve a use different from the one assessed. A confidentiality clause does not remove the right to understand a decision affecting one’s position. These problems require legal interpretation and organisational design, as well as knowledge of technical operation.

Advice can create value when it enters the project early enough to make constraints negotiable. This applies to a business choosing a provider, a public body preparing a procurement, a healthcare organisation introducing clinical support or an employer reorganising a process. The same knowledge of the project then supports engagement with authorities and the defence of a dispute, because it allows the facts, responsibilities and reasons for choices to be reconstructed.

The Firm assists businesses, organisations and professionals with the legal assessment of AI uses, the preparation of contracts and organisational safeguards, training on regulatory matters and the management of challenges. The work develops through engagement with technical functions and process owners, according to the needs of the individual engagement. A confidential meeting with the Firm can be arranged for an initial assessment.

Professional interest in this subject arises precisely from its purpose: helping those who decide to understand which tools they may use, under what conditions and with what responsibilities. Innovation acquires stability when the organisation can explain its choices and has the means to correct them. In my view, it is by this capability that the legal soundness of businesses and public authorities using artificial intelligence is measured.

References [1] [2] [16]