The AI Act is delayed. Accountability is not.

For months, 2 August 2026 was presented as the date when requirements for high-risk systems would become applicable. Regulation (EU) 2026/1744, the AI Omnibus, entered into force on 27 July 2026 and amended that timetable. Some companies read the news as a general reprieve. It is not.

The postponement concerns the main requirements in Chapter III. High-risk systems classified under Annex III, such as those used in recruitment, education or access to essential services, now have a new date of 2 December 2027. For AI embedded in products covered by EU legislation listed in Annex I — for example medical devices, toys or machinery — the date is 2 August 2028. These are two different legal categories with different timelines.

Meanwhile, provisions that have already entered into force continue to apply. Prohibitions and the duty to take measures supporting AI literacy apply from 2 February 2025; obligations for providers of general-purpose AI models, or GPAI, from 2 August 2025 (with a transition until 2 August 2027 for models already on the market before the first date); and Article 50 transparency rules from 2 August 2026. The Omnibus also introduces, from 2 December 2026, two new prohibitions on systems that generate non-consensual intimate or sexually explicit images and child sexual abuse material.

Table 1. EU timeline as at 15 October 2026. Dates depend on the system category and the organisation’s role.

Transparency does not mean “label everything in the same way”. Article 50 sets out different duties: providers of systems intended to interact directly with people must ensure users are informed, unless this is obvious from the context; providers of generative systems must make outputs detectable through machine-readable markings in line with the state of the art; deployers of deepfakes and certain public-interest texts must inform the public, subject to exceptions, including human editorial review for texts. For generative systems already on the market before 2 August 2026, the amended regulation sets a specific transition until 2 December 2026 for Article 50(2) markings.

Saying that the AI Act “regulates uses, not technology” is too simple. The regulation classifies many systems according to their context of use, but also imposes direct duties on providers of general-purpose models. The same model might summarise internal documents or be embedded in a system that screens job applicants: the function, risk and relevant obligations change.

Table 2. Summary for orientation; the legal classification of each system still requires a case-specific assessment.

The regulation can also reach beyond EU borders: it may apply to providers or deployers established outside the EU when a system is placed on the EU market or put into service there, or when its output is used in the Union. For businesses active in several markets, the EU map does not replace local legal analysis.

Responsibility is not decided once for the whole company. It must be assessed for each system and the way it is provided and used.

A **provider** develops a system or places it on the market or puts it into service under its own name or trademark. A **deployer** uses a system under its authority in the course of its activities. Using a third-party commercial model or connecting one to an internal workflow does not automatically make a company a provider.

The assessment may change if a company places a high-risk system on the market under its own name, substantially modifies a high-risk system, or changes its intended purpose so that it becomes high-risk. These are specific conditions in the regulation, not a general rule for every customisation. In uncertain cases, document the technical changes and obtain a legal assessment before release.

For a deployer of high-risk systems, the regulation requires, among other things, use in accordance with instructions, competent human oversight, monitoring and retention of logs under its control for the specified period. Where systems concern workers, the organisation must also inform worker representatives and affected people as required by the applicable provisions. Fundamental-rights impact assessments apply to specified deployers and systems; they are not a blanket duty for every company.

For a luxury house, the distinction is practical. A product-question assistant, a clienteling support tool, automated candidate screening and an AI system embedded in a device do not belong in the same category merely because they use “AI”. Purpose, effect on people, data, human control and the maison’s role in the chain all matter.

Italian Law No. 132 of 23 September 2025, in force since 10 October 2025, sets out national principles and delegated powers on artificial intelligence and must be read consistently with the AI Act. Article 20 designates AgID and ACN as Italy’s national AI authorities, while preserving the supervisory powers of the Bank of Italy, CONSOB and IVASS in their respective fields.

An implementing legislative decree, published on 8 October 2026, will enter into force on 23 October. It defines the framework in more detail: AgID acts as the notifying authority; ACN as the market surveillance authority and single EU contact point. For luxury businesses, the practical rule is not to stop at the label “Italian law”: identify the authority relevant to the company’s role, sector and type of system.

The Italian framework does not create an alternative compliance path to the AI Act. Other laws may apply to particular cases, including the GDPR, employment law, consumer protection and, for entities within scope, NIS2. The assessment depends on the activity and data involved.

In many organisations, compliance sits with legal, security with IT and use cases with commercial teams. If those groups work separately, the inventory may be incomplete and controls may arrive only after a system has been connected to real data or processes.

IBM and the Ponemon Institute’s *Cost of a Data Breach Report 2026* examined 602 organisations that had experienced a breach. In that sample, 21% reported an incident involving an AI model or application, up from 13% in the previous year. Among organisations that had experienced an AI-related incident, 92% lacked the AI access controls that the study considered appropriate. These are findings from the IBM/Ponemon sample, not an estimate of the share of all businesses exposed.

The finding is useful because it shifts attention from the model alone to its integration: identities, permissions, APIs, repositories, connectors and configurations. OWASP lists prompt injection, sensitive-information disclosure and excessive agency among risks specific to applications built with language models.

Table 3. Baseline controls for AI systems connected to company data and tools.

An AI incident may simultaneously be a personal-data breach, a significant cybersecurity incident and, in future, a serious incident involving a high-risk AI system. Each regime has a different trigger, recipient and timeline. An incident plan should therefore assign in advance who assesses the event and who starts the notifications.

Table 4. Indicative deadlines: check scope and national procedures before making a notification.

Not every chatbot error is a reportable incident. But a data exposure, an agent taking an unauthorised action or an output causing harm may require several parallel assessments. An annual exercise — for example, an assistant forwarding documents after reading a manipulated message — can test whether legal, security, privacy and business teams know how to coordinate.

A maison does not need a monumental programme to get started. It needs a reliable view of systems in use and a proportionate process for deciding what to approve, control or stop.

**First 30 days: make the landscape visible.** Name an executive sponsor and a small cross-functional group; list purchased platforms, AI features embedded in software, agents and tools used by teams. For each, record its purpose, supplier, data, users, internal owner and the maison’s role. Look for undeclared use as well, while respecting employment and privacy rules.

**Days 31–60: classify and set rules.** Assess each use against prohibitions, high-risk categories, transparency and sector laws. Define approved tools, data that must not be entered, who reviews outputs and which actions need approval. Contracts should clarify data handling, incident notices, model changes, documentation and audit rights.

**Days 61–90: test the controls.** Test at least one real use case and one adversarial scenario; check permissions, logs, error handling and the ability to stop the system. Assign an operational owner to every project and measure an outcome that matters to the work — time saved, quality, errors or complaints — rather than counting licences and prompts.

After the first 90 days, the steering group should revisit the inventory for every new initiative and material system change. High-risk classifications should be documented and updated when purpose, data, suppliers or use conditions change. Standards such as ISO/IEC 42001 can provide a management structure, but they do not replace the conformity assessment required by law.

The urgent decision is not to buy another compliance tool. It is to decide who leads the work, complete a credible inventory and connect legal, security, privacy and process owners. By 2 December 2026, organisations should also check the new prohibitions and the transition for markings on generative systems already on the market.

The delay gives companies time to prepare high-risk systems, but it does not defer responsibility for decisions already being made. A maison should be able to answer a simple question: if an AI system makes a mistake, exposes data or acts on someone’s behalf, who will notice, who can stop it and who is accountable for the process? The shared-responsibility model across leadership, platform and brand is explored in Organising the Maison for AI.

*This article provides general information and does not replace a legal assessment of a specific system or use case. Regulatory position checked as at 15 October 2026.*