The map of regulations and where to start

EU Regulation 2024/1689 introduces a risk-based approach: AI systems are classified from minimal to unacceptable risk, with proportionate obligations that include impact assessments, system registers, transparency toward users, effective human oversight and a continuous, documented risk management system. The obligations do not fall only on developers: deployers, the organizations that use AI systems in their own processes, also carry direct responsibilities, in particular for high-risk systems. The application calendar was rewritten by Regulation (EU) 2026/1744, published in the Official Journal on 24 July 2026: the transparency obligations of Article 50 and the penalty regime apply from 2 August 2026, the high-risk requirements of Annex III from 2 December 2027, and the requirements for systems embedded in products already regulated under Annex I from 2 August 2028.

Since 14 October 2026 the AI Act has its first harmonised standard. Commission Implementing Decision (EU) 2026/2274 published in the Official Journal the reference to EN 18286:2026, which covers the quality management system that Article 17 requires of providers of high-risk systems. Applying it gives a presumption of conformity only for the requirements the standard covers: Article 17(1) and the first sentence of Article 11(1), on drawing up the technical documentation. The other requirements, from risk management to cybersecurity, await their own harmonised standards, still being developed by CEN and CENELEC.

In the AI Rating model every key requirement of the regulation is mapped to an assessment dimension, so the assessment works as a preliminary gap analysis for compliance: each gap found already comes linked to the relevant regulatory reference and the corrective action.

ISO/IEC 42001:2023 is the first certifiable international standard for AI management systems. It defines the system through which an organization governs AI on an ongoing basis, from policies to operational controls, with the same logic ISO/IEC 27001 uses to govern information security. It is entering supplier qualifications as third-party verified proof that the stated governance actually exists. It does not give a presumption of conformity with the AI Act: for the Article 17 quality management system the Commission found it not aligned and requested a dedicated European standard, EN 18286. Organizations that already run an ISO/IEC 42001 or ISO 9001 system find a clause-by-clause correspondence in the annexes of EN 18286 and can focus on the differences. From 13 March 2026 the standard is also a European norm: CEN adopted it unchanged as EN ISO/IEC 42001:2026, with national standardization bodies expected to transpose it by September 2026.

The ZeroFive.AI framework is built in line with the standard: implementing the recommendations that come out of the rating means, in substance, walking the path toward its requirements, with the advantage of knowing in advance where you are behind and how much it costs to close the gap.

EU Directive 2022/2555, transposed in Italy through Legislative Decree 138/2024, imposes security obligations on essential and important entities that extend to the supply chain. A company within the NIS2 perimeter that integrates AI vendors into its processes must evaluate, contract and monitor them as part of its own risk perimeter, and those who sell AI solutions to those entities should expect to be assessed by that same yardstick.

AI governance and cyber security meet here: the assessment engages with the controls the CISO has already built, without duplicating them, and gives the risk committee a single view of where AI systems expose the organization.

Regulation (EU) 2024/2847 imposes cybersecurity requirements on anyone placing products with digital elements on the European market, a category that includes software distributed as a product and not only connected devices. The perimeter differs from NIS2, which targets essential and important entities: here the criterion is placing the product on the market, and this brings within scope many software houses and vendors that fall outside the NIS2 perimeter. The calendar is staggered: the regulation has been in force since 10 December 2024, Chapter IV on market surveillance authorities applies from 11 June 2026, the reporting obligations of Article 14 from 11 September 2026, and the essential requirements of Annex I from 11 December 2027.

The part already in force changes teams' daily work. Anyone who becomes aware of an actively exploited vulnerability must submit an initial alert within twenty-four hours, a full notification within seventy-two hours and a final report within fourteen days of a corrective measure becoming available; for a severe incident the initial sequence is identical and the final report is due within one month. The channel is the Single Reporting Platform managed by ENISA, with a single submission addressed to the CSIRT designated as coordinator, which in Italy operates within the National Cybersecurity Agency. The obligation also applies to products already made available on the market before 11 December 2027.

Full application of the CRA falls nine days after the AI Act obligations for the high-risk systems of Annex III. For those developing or integrating software with AI components intended for the European market, this is effectively a single preparation window, with evidence that overlaps substantially: system registers, risk management, version traceability, supplier requirements. In the rating this evidence is assessed once, without opening two parallel workstreams that ask the same documents of two different parts of the company.

On the proposal front, the cyber package presented by the Commission on 20 January 2026 would replace the 2019 Cybersecurity Act with a new regulation (COM(2026) 11) and would amend NIS2 in a targeted way (COM(2026) 13). The block that most directly affects technology buyers concerns the ICT supply chain: the Commission could designate third countries as raising cybersecurity concerns, classify vendors controlled by them as high risk and impose equipment phase-out periods. The text is under negotiation, and in the progress report of 22 May 2026 the Council asked for clarifications on the identification methodology, so it should be treated as a signal of direction rather than an imminent obligation. Supply decisions taken in the coming months will still play out within that framework.

Every AI system that processes personal data remains subject to the GDPR, from data minimization principles to privacy impact assessments. Around the regulation sit the security standards procurement has requested for years: ISO/IEC 27001 for the information security management system, ISO/IEC 27017 for cloud-specific controls, ISO/IEC 27018 for the protection of personal data in the public cloud. Since 14 October 2026 they are joined by ISO/IEC 27090, guidance on the threats specific to AI systems, such as training data poisoning and model theft, which complements 27001 and 27002 across the system lifecycle. It is neither certifiable nor harmonised, and gives the security team a shared catalogue of threats to address. A model trained or fed with personal data inherits all the protection obligations attached to that data, and the two compliance tracks, privacy and AI, need to be read together.

ZeroFive.AI does not issue certifications, and we say this clearly: we measure the distance from requirements, prepare the organization for engagement with certification bodies and client audits, and build the evidence those paths require.

Generative AI has brought intellectual property into AI compliance. The Copyright Directive (EU 2019/790) governs text and data mining and rightsholders' opt-out right, while the AI Act requires providers of general-purpose models to have a policy on copyright compliance and a sufficiently detailed summary of the training content. For those adopting these models, the issue moves downstream: what contractual guarantees the provider offers, who is liable for an output that reproduces protected works, and what internal rules are needed for the use of generated content.

These are decisions that precede implementation. We help formulate them, from the questions to ask suppliers to internal usage policies, working alongside the company's legal counsel on matters of legal interpretation.