Artificial intelligence in business: which rules apply, and who enforces them

A company that introduces a virtual assistant, analyses documents with a generative model, or automates decisions about staff must check several areas of law at the same time. The AI Act regulates AI systems and models; the GDPR protects personal data; other rules govern cybersecurity, employment, financial services, consumers and liability. Compliance with one discipline does not cover the others. ^{[1] [6]}

This guide maps the cross-cutting framework and the main sector-specific rules for businesses. Applicability depends on the activity, the company's role in the supply chain, the data processed and the purpose of the system: not every rule referenced here imposes obligations on every business.

1. AI Act: start by identifying the company's role

The EU reference text is Regulation (EU) 2024/1689, to be read together with the amendments made by Regulation (EU) 2026/1744, the Digital Omnibus on AI. A company may be a provider, a deployer — i.e. a professional user — an importer or a distributor. A business that buys a service for internal use is normally a deployer; a business that develops a system, or has one developed, and places it on the market or puts it into service under its own name may be a provider. Roles must be assessed case by case. ^{[1] [2]}

For high-risk systems, Article 25 also governs the shift to provider obligations: relevant triggers include affixing one's own name or trademark, a substantial modification, or a change of purpose that makes a previously non-high-risk system high-risk. Simply integrating a model via an API is therefore not enough, on its own, to determine the role — the finished product must be examined. ^[3]

Prohibited practices and high-risk systems

Article 5 bans specifically identified practices: manipulation and exploitation of vulnerabilities under the conditions set out in the rule, certain social-scoring systems, the indiscriminate scraping of facial images to build databases, and certain biometric uses. Emotion-recognition in the workplace is prohibited, except for medical or safety purposes. An employee's consent is not a general exemption. The new prohibitions on non-consensual intimate imagery and child sexual abuse material apply from 2 December 2026. ^{[4] [5]}

High-risk status is determined under Article 6 and Annexes I and III. For businesses, the relevant categories include candidate selection and evaluation, worker management, creditworthiness assessment of individuals, and risk assessment or pricing in life and health insurance. Education, biometrics and safety components of critical infrastructure are also relevant. The exception for systems intended to detect financial fraud does not automatically exempt every piece of banking software. ^{[7] [8]}

An assistant that summarises documents is not automatically high-risk. But if the system takes part in the decisions covered by Annex III, the conditions of Article 6 must still be checked: the exclusion for narrow or preparatory tasks is not automatic, and profiling of individuals retains high-risk classification in the cases covered by the provision. ^[7]

Obligations differ for providers and for users

The high-risk framework requires risk management, data quality, technical documentation, event logging, human oversight, accuracy, robustness and cybersecurity. The provider must follow the compliance obligations applicable to its system category. The deployer must use the system according to the instructions, assign oversight to competent people with the necessary authority, monitor operation, and manage risks and incidents. For logs under its control, Article 26 requires at least six months' retention, subject to any other applicable rules. ^{[1] [9]}

The fundamental rights impact assessment, or FRIA, under Article 27, applies to specific categories of deployer, including those identified for credit and for life and health insurance — it is not required of every business across the board. It is distinct from the GDPR's DPIA: the two assessments can be coordinated and share relevant parts, but they rest on different legal triggers. ^[10]

Transparency, chatbots and generated content

Article 50 sets out several distinct obligations. Providers must make interaction with an AI system recognisable when this is not already obvious, and must ensure technically detectable marking of synthetic outputs in the cases specified. Deployers must disclose the artificial nature of deepfakes and, under certain conditions, of text published to inform the public on matters of public interest. An exception applies to such text where there has been human review or editorial control and editorial responsibility by a natural or legal person. There is therefore no identical obligation to label every email or document produced with the help of AI. ^[11]

AI literacy, too, should be read in the updated text: Article 4 requires measures to support the development of skills, appropriate to the people and the context involved; it does not require guaranteeing everyone a specific level of literacy, nor does it introduce a universal "AI licence". ^[12]

General-purpose models and open-source software

For providers of general-purpose AI models (GPAI), Article 53 covers documentation, information for parties integrating the model, a copyright policy, and a public summary of training content. Open source benefits from narrow exceptions: it does not remove every obligation, and it does not exempt a provider in general from copyright, GDPR, or the rules on systems built on top of the model. Additional requirements apply to models with systemic risk. ^{[13] [1]}

2. Deadlines: the calendar was changed in 2026

Entry into force of the AI Act must be distinguished from the date each individual obligation becomes applicable. The original AI Act calendar was updated by the Digital Omnibus. ^{[2] [5]}

Postponing the high-risk requirements does not postpone obligations already applicable under the GDPR, employment law, financial regulation or cybersecurity law. For systems and models already on the market, Article 111 must also be examined: a single general date does not resolve every transitional situation. ^{[5] [6]}

3. GDPR and privacy: what to check when AI processes personal data

Regulation (EU) 2016/679 applies whenever information relating to an identified or identifiable natural person is processed, even if contained in invoices, CVs, corporate documents or conversations. In Italy it operates alongside Legislative Decree 196/2003, the Privacy Code. The AI Act does not replace these rules. ^{[6] [14]}

The review should cover legal basis and purpose, data minimisation, accuracy, retention, privacy notices, the roles of controller and processor, contracts with providers, security, and transfers outside the European Economic Area. The central provisions are Articles 5, 6, 9, 13–14, 25, 28, 30, 32–36 and 44–49. Consent is not the only legal basis and should not be used as an automatic default, particularly in employment relationships. ^[6]

A DPIA is required where processing is likely to result in a high risk to rights and freedoms, under Article 35: using AI calls for an assessment, but does not by itself trigger a universal DPIA obligation. Likewise, appointing a DPO depends on the conditions in Article 37, not on the mere purchase of a chatbot. ^{[6] [15]}

Article 22 protects individuals from decisions based solely on automated processing that produce legal effects or similarly significant consequences. Specific exceptions and safeguards apply: a purely formal human signature does not amount to a genuine review. The issue is particularly relevant for hiring, credit, benefits and the evaluation of individuals. In one case concerning a medical research project, the Italian Data Protection Authority required, as concrete safeguards, transparency about the statistical models used and human involvement in selecting the training data. ^[16]

On training, EDPB Opinion 28/2024 clarifies that the anonymity of a model and the usability of legitimate interest as a legal basis must be assessed concretely. Public availability of data does not mean it can be freely used; a model trained on personal data is not necessarily anonymous. ^[17]

A local model may reduce certain risks of transfer and third-party access, but it does not automatically make processing compliant. As an organisational measure, a company should authorise which tools and data may be used, check retention and reuse for training, and govern access.

4. Italy's AI law and the professions

Law No. 132 of 23 September 2025, in force since 10 October 2025, supplements the EU framework. For businesses, the most relevant provisions are Article 11 on employment, Article 13 on the regulated professions, Article 20 on competent authorities, Article 25 on copyright, and Article 26 on criminal provisions. ^[18]

Article 13 permits the use of AI in the regulated professions for ancillary and support activities, with the intellectual work that is the subject of the professional service remaining predominant; it requires clear, simple and comprehensive disclosure to the client of the systems used. For accountants, lawyers and other professionals, professional judgment and liability therefore remain central. ^[19]

5. Employment: screening, monitoring and discrimination

Alongside the AI Act and the GDPR, Article 4 of Law 300/1970 on remote monitoring and Article 1-bis of Legislative Decree 152/1997 on automated decision-making or monitoring systems also apply. Article 11 of Law 132/2025 refers to the latter rule for informing workers, and requires respect for dignity, confidentiality and non-discrimination. ^[20]

The duty to inform must be distinguished from the preconditions for installing or using monitoring tools: where required, a trade-union agreement or an authorisation is not replaced by individual consent. General anti-discrimination rules also remain applicable, including those in Legislative Decrees 198/2006, 215/2003 and 216/2003, depending on the protected ground concerned. ^{[20] [21]}

Depending on their respective powers, the National Labour Inspectorate, the Italian Data Protection Authority and the courts become involved; equal-opportunity bodies act within their own remits. The postponement of the AI Act's high-risk rules does not authorise unlawful monitoring of employees today.

6. Bank of Italy, Consob, IVASS and DORA

The Bank of Italy is not the general authority for every business that uses AI. Article 74(6) of the AI Act assigns supervision of high-risk systems to the competent national financial authorities when placing on the market, putting into service, or use is directly connected with the provision of regulated financial services. Law 132/2025 expressly preserves the powers of the Bank of Italy, Consob and IVASS. ^{[22] [23]}

An ordinary business that analyses its own cash flow does not thereby become a regulated financial intermediary. A bank that uses AI for credit scoring, or an insurer that uses it for life and health pricing, must instead coordinate the AI Act, privacy law, governance and sector-specific rules.

Regulation (EU) 2022/2554 — DORA, applicable from 17 January 2025, governs digital operational resilience, ICT risk, incidents, testing and the external ICT providers of the financial entities within its scope. Legislative Decree 23/2025 implements it in Italy and identifies the Bank of Italy, Consob, IVASS and COVIP as competent authorities, each according to its own supervisory remit; the ECB retains its role in matters of EU competence. An AI provider may be drawn into contractual ICT safeguards and, if designated critical, into the European oversight framework: this does not mean every provider is subject to the same obligations as the bank itself. ^{[24] [25]}

The Bank of Italy communication of 17 July 2026 highlights the cyber risks of advanced models and requires its addressees to take specific action on governance, inventories, vulnerabilities, monitoring and testing. It expects the board of directors and the board of statutory auditors to examine the communication in a joint session, and requires a report with a work plan to be sent to Supervision by 31 December 2026. This is a supervisory communication addressed to the entities listed in the document, not an obligation extended to every Italian company. ^[26]

7. Cybersecurity: NIS2 and the Cyber Resilience Act

Directive (EU) 2022/2555 — NIS2 is implemented in Italy by Legislative Decree 138/2024. For entities within its scope, it introduces obligations on risk management, supply-chain security and incident notification. The National Cybersecurity Agency (ACN) is the competent national authority and single point of contact; sector authorities contribute to implementation. Using AI does not, on its own, bring a business within the scope of NIS2. ^[27]

Regulation (EU) 2024/2847 — the Cyber Resilience Act (CRA) concerns products with digital elements, within the limits and exceptions of its scope. For a software product that incorporates AI, the security of the life cycle, vulnerabilities and conformity must be assessed. Not every SaaS service is automatically included: the criteria on remote data-processing solutions are also relevant. ^[28]

The CRA's main obligations apply from 11 December 2027, while the reporting obligations under Article 14 apply from 11 September 2026. The reporting chain involves the coordinating CSIRT and ENISA through the single reporting platform; this is distinct from market surveillance and from notifying personal-data breaches to the Data Protection Authority. A single incident can trigger several disciplines and several notifications. ^{[28] [29]}

8. Data, copyright, trade secrets and platforms

Data Act and Data Governance Act

The Data Act — Regulation (EU) 2023/2854, generally applicable from 12 September 2025 with specific later dates for certain provisions, covers access to and use of data from connected products, data sharing, and switching between data-processing service providers. It does not authorise training on the data obtained without restriction, and it leaves privacy rules unaffected. On competent authorities, Article 37 and the related national designation are relevant: the privacy function remains with the data protection authorities. Among the official sources consulted for this guide, no definitive Italian measure has been identified that would allow a single authority to be named without reservation for every aspect of the Data Act. ^[30]

The Data Governance Act — Regulation (EU) 2022/868 has a different scope: reuse of certain protected public-sector data, data-intermediation services, and data altruism. Legislative Decree 144/2024 assigns the relevant national functions to AgID. Not every dataset manager or AI application is a data intermediary under the DGA. ^[31]

Intellectual property and confidential information

Law 633/1941 and Directive (EU) 2019/790 govern copyright and text-and-data mining. The exceptions have precise conditions, including lawful access and, where relevant, a reservation of rights. Law 132/2025 protects works created with the help of AI where they are the result of the author's own intellectual effort, and introduces a reference to the text-and-data-mining rules for AI models and systems. ^[32]

Articles 98–99 of Legislative Decree 30/2005 protect trade secrets under the conditions set out there. Uploading confidential information to an external service therefore requires a contractual and confidentiality assessment, in addition to a privacy assessment. ^[33]

Marketing and electronic communications

When AI is used for promotional campaigns, profiling or interaction analysis, the rules on electronic communications still apply: Directive 2002/58/EC and, in Italy, Articles 122 and 130 of the Privacy Code. Consent, access to the terminal device, marketing and the related exceptions must be assessed separately from the mere availability of an AI system. Oversight here is a matter for the Data Protection Authority. ^[14]

Consumers, digital services and competition

The Italian Consumer Code — Legislative Decree 206/2005 also applies to services marketed as AI. Unfounded accuracy claims, material omissions, or misleading communications can fall within the unfair commercial practices policed by the Italian Competition Authority (AGCM). Algorithms do not exempt a business from competition rules. ^[34]

The Digital Services Act — Regulation (EU) 2022/2065 covers intermediary services and platforms within its scope, not every chatbot. AGCOM is Italy's Digital Services Coordinator; the European Commission exercises the powers provided for very large online platforms and search engines. Automated advertising, recommendation and moderation may therefore encounter further obligations. ^[35]

The Digital Markets Act — Regulation (EU) 2022/1925 is addressed to designated gatekeepers and the core platform services concerned. Integrating AI into such services may bring its rules into play; it is not a general regime imposed on every SME. Enforcement is a matter for the European Commission, within the cooperation framework provided with national authorities. ^[44]

9. Liability, governance and other sectors

Legislative Decree 160 of 9 September 2026, in force since 30 September 2026, contains provisions on the use of AI in policing and on civil and criminal liability. For businesses, the relevant provisions are those on access to evidence about how the system works, a presumption of causal link under the conditions set out there, and the rule that conformity with the AI Act does not by itself rule out liability. Documentation and traceability can therefore also matter in litigation. ^[36]

Directive (EU) 2024/2853 on liability for defective products brings software within the new framework. The EU transposition deadline is 9 December 2026: this must be kept distinct from the national rules already in force, and should not be presented as if all of its effects were already operative as of 6 October. ^[37]

The adequate-organisational-arrangements duty under Article 2086 of the Italian Civil Code offers an organisational reference point for governing AI-related risk and liability. Folding AI risk into these arrangements is an operational consequence to be assessed according to the nature and size of the business; the provision does not by itself mandate a universal "AI committee". Legislative Decree 231/2001 may also be relevant where the use of AI contributes to a predicate offence under the conditions governing corporate liability: not every model error gives rise to liability under that decree. ^[38]

10. Who enforces what: the map of authorities

The DPO, the board of statutory auditors, the supervisory body and internal control functions are not public authorities. They may have essential advisory, oversight or internal-control duties without absorbing those of external authorities. Even a notified body assessing conformity has a role distinct from that of a supervisory authority. ^{[6] [23] [38]}

11. Penalties and documentation to prepare

Article 99 of the AI Act sets out different maximum amounts: up to €35 million or 7% of worldwide turnover for prohibited practices; up to €15 million or 3% for the violations listed in paragraph 4, including transparency; up to €7.5 million or 1% for inaccurate, incomplete or misleading information under paragraph 5. For undertakings, the higher of the two figures normally applies; for SMEs, paragraph 6 applies the lower figure, and the new paragraph 6a extends this favourable treatment to small mid-cap companies, but only for violations of paragraphs 4 and 5: for prohibited practices under paragraph 3, the full €35 million/7% ceiling still applies. These are maximum amounts, not automatic fines. ^[43]

The GDPR keeps its own penalty regime, including the ceilings of up to €20 million or 4% for the infringements covered by Article 83(5). Temporal applicability, the specific conduct involved, and the procedural rules must all be assessed: it is not correct to simply add all the percentages together. ^[6]

As an organisational suggestion, a business can start its review with a file proportionate to its own use of AI:

1. Inventory: tools, purposes, users, data, providers and integrations.
2. Classification: role in the supply chain, AI Act risk level, prohibitions and sector-specific rules.
3. Privacy and contracts: legal bases, privacy notices, roles, transfers, retention and conditions for reusing data.
4. Impact assessments: checking the preconditions for a DPIA and a FRIA, without conflating them.
5. Operational safeguards: access controls, training, output checks, oversight and the ability to stop the system.
6. Traceability: versions, tests, decisions and incidents, retained consistently with applicable obligations.
7. Responsibility and deadlines: who is in charge, periodic review, and which authority to contact for each type of event.

Not all of these documents are formally mandatory for every business. Their value lies in making the risk assessment and the implementation of the obligations that actually apply verifiable. Voluntary certifications, such as those for AI management systems, can support the organisation; on their own, they do not replace regulatory compliance.