ΕλληνικάA deferral is not a pause
On 27 July 2026 the Digital Omnibus on AI, Regulation (EU) 2026/1744, entered into force, six days before the date on which the EU AI Act high-risk obligations were due to apply. For eighteen months organizations had been working towards 2 August 2026. That deadline has moved, and the question every board is now asking is whether the pressure has come off.
It has not. The Omnibus deferred one part of the regime and left the rest untouched. Obligations that apply by reference to what an AI system does, rather than to the risk tier it occupies, took effect on 2 August 2026 exactly as originally scheduled. A second date falls on 2 December 2026. The reputational and legal risk in this period is not that firms miss a distant deadline. It is that they read a headline about delay, stand down a programme, and fail obligations that are already live.
What the Omnibus actually changed
The instrument amends the AI Act, Regulation (EU) 2024/1689, and is the first formal set of amendments since its adoption. It was approved by the European Parliament on 16 June 2026, adopted by the Council on 29 June 2026, signed on 8 July 2026, published in the Official Journal on 24 July 2026 and entered into force on the third day following publication.
The central change is the deferral of the high-risk obligations. For stand-alone high-risk systems classified under Annex III, covering use cases in areas including employment, education, credit scoring, biometrics, critical infrastructure and law enforcement, the application date moves from 2 August 2026 to 2 December 2027. For AI embedded in products already governed by EU sectoral safety legislation under Annex I, including medical devices, machinery and toys, the date moves to 2 August 2028. Providers of high-risk systems intended for use by public authorities have a further extended period under Chapter III.
Two features of the new timetable deserve attention. The dates are now fixed rather than linked to the finalization of harmonized standards, which removes an uncertainty that had made programme planning difficult. And the deferral is expressed as a change of date, not a change of substance: the obligations themselves, covering risk management, data governance, technical documentation, human oversight, conformity assessment and registration, are unchanged.
What applied on 2 August 2026 and did not move
The transparency obligations in Article 50 were not amended by the Omnibus and took effect on schedule. They apply according to the function of the system rather than its risk classification, which means mainstream generative deployments are squarely within scope notwithstanding the Annex III deferral.
From 2 August 2026, providers and deployers must disclose that a person is interacting with an AI system. Deployers must disclose deep-fakes depicting real persons, places or events. Deployers of emotion recognition and biometric categorization systems must inform the individuals exposed to them and must process personal data in accordance with EU data protection law, subject to limited exceptions for the detection, prevention and investigation of criminal offenses.
Two other bodies of obligation also remain on their original timetable. The prohibited practices in Article 5 have applied since February 2025. The obligations on providers of general-purpose AI models, including transparency, documentation and systemic risk requirements for the largest models, have applied since August 2025 and were not touched.
The date that most firms will miss: 2 December 2026
Article 50(2) requires providers of systems generating synthetic audio, image, video or text to mark outputs in a machine-readable and detectable format. The Omnibus granted a limited transitional period, but only for systems already on the market. A generative system placed on the EU market before 2 August 2026 must comply with the marking obligation by 2 December 2026. A system placed on the market on or after 2 August 2026 must comply from that date.
This is a four-month grace period, shorter than the six months originally proposed, and it is the single most commonly misunderstood element of the current position. Firms that read the Omnibus as a general delay will treat watermarking as a 2027 problem. It is not. For legacy generative systems it is a December 2026 problem, and machine-readable marking is an engineering change rather than a policy document.
Two new prohibited practices from 2 December 2026
The Omnibus adds prohibitions directed at AI systems used to create non-consensual intimate imagery and child sexual abuse material. The prohibition reaches the placing of such systems on the EU market for that purpose, and also the placing on the market of systems without reasonable safety measures to prevent such creation. It applies from 2 December 2026.
For providers of general-purpose generative systems this is a design and safeguards question rather than a disclosure question, and the obligation to have taken reasonable preventive measures cannot be satisfied retrospectively after a system is already in the market.
The wider Digital Omnibus Package is not law
The AI Omnibus is the artificial intelligence strand of a broader package published by the Commission on 19 November 2025. The remainder of that package, which proposes amendments to the General Data Protection Regulation, the ePrivacy Directive, the NIS2 Directive and the Data Act, remains subject to negotiation in the Parliament and the Council and has not been adopted.
This distinction matters commercially. Advisers and vendors have begun to describe the package as though it were a single enacted reform. It is not. Only the AI strand is in force. Any compliance decision taken on the assumption that NIS2 or GDPR obligations have been simplified is being taken on a proposal, and the negotiating history of the AI strand shows how much can change between proposal and adoption.
Where this meets DORA, NIS2 and existing supervisory expectations
For regulated financial entities the AI Act does not arrive in isolation. DORA has applied since 17 January 2025 and is now in its first genuine supervisory cycle. NIS2 obligations continue to bite through national transposing law. An AI system that fails inside a regulated firm can engage ICT incident reporting under DORA, cybersecurity incident notification under national NIS2 rules, personal data breach notification under the GDPR and, depending on the system, AI Act obligations at the same time.
Supervisors have already signaled that they expect this to be governed as one problem. In their first annual report on major ICT-related incidents, published on 3 June 2026, the European Supervisory Authorities noted that the evolution of highly capable AI-driven tools should encourage financial entities to strengthen cybersecurity measures. The practical consequence is that AI governance is not a separate workstream that can wait until December 2027. It sits inside the ICT risk, incident and third-party frameworks that regulated firms are already required to operate.
A sensible engagement sequence
The right starting point is a paid legal and regulatory assessment that establishes which systems the organization actually provides and deploys, how each is classified, which obligations are live today, and what falls due on 2 December 2026. Only after that diagnostic can a remediation or implementation programme be scoped and priced responsibly. OIKONOMAKIS LAW performs the legal and regulatory workstream; F SOCIETY provides the operational readiness, governance, controls and evidence workstream, with specialist technical providers used where required.
Deeper analysis and implementation:
Transparency applies by function, not by risk tier
The most consequential structural point in the current position is that Article 50 does not depend on the Annex III classification. An organization can conclude, correctly, that none of its systems is high risk, and still be fully within the transparency regime because it deploys a customer-facing chat interface, generates marketing copy or images with a generative tool, or uses synthetic voice in a service channel.
This is where the deferral does most damage if it is misread. A classification exercise that stops at the high-risk question, concludes that nothing is in scope and closes the file will produce a compliant-looking record and a non-compliant organization. The classification must run twice: once against Annex III, and once against the functional triggers in Article 50.
Article 50 also contains an editorial control element for AI-assisted content that undergoes human review before publication. The scope of that carve-out should be assessed against the actual editorial workflow rather than assumed, because a review process that is nominal will not support it.
The deferral changes the deadline, not the build time
Conformity assessment, technical documentation under Annex IV, risk management, data governance, human oversight and post-market monitoring are not documentation exercises that can be produced in the final quarter before a deadline. They describe how a system is designed, tested, recorded and supervised across its lifecycle. An organization that stands down its high-risk programme in August 2026 and restarts in mid-2027 will have lost the interval in which the underlying capability could have been built, and will be assembling evidence about a period during which nothing was recorded.
Standards and Commission guidance are expected to continue developing through the deferral period. Firms that maintain their programme will be able to align to those instruments as they emerge. Firms that stopped will be reading them for the first time under time pressure.
Governance, evidence and the supervisory record
The organization should be able to produce, on request, an inventory of AI systems provided and deployed, the classification decision for each with the reasoning and the person accountable for it, the transparency measures implemented and the date they took effect, and the change-control process that determines when a system is reassessed. Classification decisions should carry an owner, a date, a source and a statement of what change would require the decision to be revisited.
The same discipline that DORA supervision has begun to apply to the Register of Information will be applied here in due course: a defensible record is one in which an independent reviewer can trace the path from obligation to decision to implementation, and in which the gaps that were identified were owned and managed rather than silently left open.
Primary and official sources for publication verification
Regulation (EU) 2024/1689 (EU Artificial Intelligence Act), consolidated text.
Regulation (EU) 2026/1744 (Digital Omnibus on AI), published in the Official Journal 24 July 2026, in force 27 July 2026.
European Commission, Digital Omnibus Package, published 19 November 2025. Non-AI strands remain in negotiation and are not law.
Regulation (EU) 2022/2554 (DORA); Directive (EU) 2022/2555 (NIS2), as transposed nationally.
European Supervisory Authorities, first annual report on major ICT-related incidents, 3 June 2026, published under Article 22(2) DORA.
Frequently asked questions
Has the EU AI Act been delayed? Only in part. The Digital Omnibus on AI deferred the high-risk obligations for stand-alone Annex III systems to 2 December 2027 and for Annex I embedded systems to 2 August 2028. The transparency obligations in Article 50, the prohibited practices in Article 5 and the general-purpose AI model obligations were not deferred.
What applied from 2 August 2026? Disclosure that a person is interacting with an AI system; disclosure of deep-fakes depicting real persons, places or events; and notification obligations for emotion recognition and biometric categorization systems. These apply by reference to what the system does, not to its risk tier.
When must AI-generated content be watermarked? Systems placed on the EU market on or after 2 August 2026 must comply from that date. Generative systems already on the market before 2 August 2026 have until 2 December 2026. Machine-readable marking is an engineering change and should not be left to the final weeks.
We have concluded that none of our systems is high risk. Are we outside the regime? Not necessarily. Article 50 applies according to system function rather than classification. A classification exercise that tests only against Annex III and closes will miss live obligations. The assessment should run against both.
Do the GDPR and NIS2 changes in the Digital Omnibus apply? No. Only the AI strand has been adopted and is in force. The proposed amendments to the GDPR, the ePrivacy Directive, NIS2 and the Data Act remain in negotiation and are not law.
Should we stand down our high-risk programme until 2027? We would not advise it. The obligations are unchanged; only the date moved. Conformity assessment, technical documentation and post-market monitoring describe how a system is built and supervised over time, and cannot be reconstructed for a period during which nothing was recorded.
This publication provides general legal and regulatory information as at 18 August 2026 and does not constitute legal advice on any individual matter. The AI Act is directly applicable EU law, but national procedure, sectoral regulation and professional-practice requirements remain matters of Member-State law. Classification, applicable obligations and implementing measures should be verified in relation to the specific systems and business model concerned. Nothing here guarantees any regulatory outcome. Assessment of a specific matter requires advice from a qualified lawyer.
Prepared by Christos Oikonomakis, Chairman, OIKONOMAKIS LAW