For most of 2026, compliance teams worked toward a single date: 2 August 2026, when the bulk of the EU AI Act’s high-risk requirements were due to apply. Then, six days before that deadline, the Digital Omnibus on AI entered into force and moved the goalposts. Many headlines summarized it as “the AI Act is delayed.” That framing is dangerous. Some obligations moved by 16 to 24 months, but others became enforceable on schedule, and a few new ones were added.
This article explains what actually changed, what is enforceable today, and how organizations that build, buy or integrate AI should use the extra runway. It is written for CIOs, product leaders, risk owners and engineering managers who need a working plan rather than a legal treatise. It is not legal advice; use it to structure the conversation with your counsel.
What the Digital Omnibus Actually Changed
Regulation (EU) 2026/1744, often called the AI Omnibus, was published in the Official Journal on 24 July 2026 and entered into force on 27 July 2026. It amends the AI Act (Regulation (EU) 2024/1689) rather than replacing it. The main changes, as summarized by the Cloud Security Alliance and the National Law Review, are:
- Stand-alone high-risk systems (Annex III) — AI used in employment, education, credit scoring, essential services, law enforcement and critical infrastructure — now face their core obligations from 2 December 2027 instead of 2 August 2026.
- High-risk AI embedded in regulated products (Annex I), such as medical devices, toys or machinery covered by EU product-safety law, moved to 2 August 2028.
- Two new prohibited practices target systems designed to generate or manipulate realistic non-consensual intimate imagery or child sexual abuse material, with a technical-safeguard grace period until 2 December 2026.
- AI literacy under Article 4 remains, but the final text softens the duty toward “measures supporting” literacy rather than a hard outcome requirement.
- The AI Office gained supervisory reach over certain AI systems built on general-purpose models and over AI in very large online platforms and search engines, including information demands, inspections and periodic penalty payments.
- National regulatory sandboxes must be in place by 2 August 2027, giving companies a formal route to test systems with regulators.
The stated rationale was readiness: the harmonised standards that high-risk providers rely on to demonstrate conformity, being drafted in CEN-CENELEC’s joint technical committee, were not going to be finished in time. Without them, companies would have had to prove compliance against requirements with no agreed technical yardstick.
What Did Not Move: Obligations Enforceable Today
The parts of the Act that were already in force, or due on 2 August 2026 outside the high-risk regime, largely stayed on schedule. That makes the post-Omnibus landscape a layered one.
Prohibited practices (since February 2025)
Article 5 bans — including manipulative techniques that cause significant harm, social scoring, untargeted scraping of facial images, and emotion recognition in workplaces and schools with narrow exceptions — have applied since 2 February 2025. Penalties for violating them reach up to €35 million or 7% of worldwide annual turnover, whichever is higher.
General-purpose AI model obligations (enforceable since August 2026)
Obligations for providers of general-purpose AI (GPAI) models under Articles 53 to 55 have applied since 2 August 2025. What changed on 2 August 2026 is enforcement: the Commission, through the AI Office, can now request documentation, run its own evaluations, require mitigations, and fine providers up to €15 million or 3% of global turnover. Models already on the market before August 2025 have until 2 August 2027 to comply, according to the Commission’s GPAI guidelines.
Transparency obligations under Article 50 (since August 2026)
Article 50 duties apply by function, not by risk tier. If your system interacts with people, it must tell them they are dealing with AI unless that is obvious. Providers of generative systems must mark synthetic audio, image, video and text in a machine-readable way, and deployers must disclose deepfakes and certain AI-generated text published on matters of public interest. Only a narrow four-month grace period was granted for machine-readable marking on generative systems already on the market before 2 August 2026. In practice, that means customer-service chatbots, AI voice agents and content-generation features need disclosures and labeling now.
The Omnibus deferred the high-risk regime; it did not defer the AI Act. Transparency duties, GPAI enforcement and the prohibitions are live, and the extra 16 months for high-risk systems is runway, not a reprieve.
Are You a Provider, a Deployer, or Both?
The Act assigns obligations by role, and many organizations underestimate how often they hold more than one. Getting the role wrong is the most common root cause of misdirected compliance effort.
- Provider: you develop an AI system or model, or have one developed, and place it on the market or put it into service under your own name. A SaaS company shipping an AI feature to EU customers is usually a provider of that system.
- Deployer: you use an AI system under your authority in a professional context. A bank using a vendor’s credit-scoring model is a deployer, with duties around human oversight, logging, input data and, for some public-facing uses, fundamental-rights impact assessments.
- Downstream modifier: you fine-tune or otherwise modify a GPAI model. The Commission’s guidelines indicate that you may become the provider of the modified model if your modification uses more than one third of the original model’s training compute. Most enterprise fine-tuning sits far below that line, but you should document the calculation.
- Importer and distributor: you make a non-EU provider’s system available in the EU, with verification duties of your own.
One more trap: if a deployer substantially modifies a high-risk system, changes its intended purpose, or puts its own brand on it, it can inherit provider obligations. Product teams that “just wrap” a vendor model in a new workflow should check this explicitly.
A Practical Readiness Plan for the Next 12 Months
The extended runway is only valuable if it is used for work that is hard to compress later: inventory, classification, data governance, documentation and post-market monitoring. A sequenced plan looks like this.
- Rebuild the AI inventory against the amended text. List every AI system and GPAI model in use or in development, its owner, intended purpose, users, EU exposure and vendor. Cite Regulation (EU) 2026/1744 in your trackers rather than earlier drafts, and re-check Annex I scoping where product-safety legislation has changed.
- Classify and record the rationale. For each system, decide whether it is prohibited, high-risk under Annex I or III, subject to Article 50 transparency, or minimal risk. Where you rely on the Article 6(3) carve-out (an Annex III use that does not pose a significant risk, for example a narrow procedural task), write down the reasoning; you will need it.
- Close Article 50 gaps immediately. Add AI disclosures to chat and voice interfaces, implement machine-readable marking (for example C2PA content credentials or watermarking supported by your model provider) for generated media, and agree an editorial process for deepfake and AI-generated text disclosures.
- Run GPAI supplier due diligence. For every foundation model you depend on, confirm whether the provider signed the GPAI Code of Practice, request the Article 53 documentation intended for downstream providers, check for an EU authorized representative if the provider is outside the EU, and add documentation and incident-notification clauses to contracts.
- Build the high-risk technical file progressively. For systems likely to be Annex III high-risk, start the risk management system, data governance records, technical documentation, logging design, human-oversight procedures, and accuracy and robustness testing now. These artifacts take quarters, not weeks.
- Design post-market monitoring and incident reporting. Define what a serious incident is for each system, who decides, and how it reaches the people responsible for regulatory reporting.
- Plan for model withdrawal. The AI Office can restrict a GPAI model. Keep an abstraction layer and a tested fallback model so a supplier issue does not take your product down.
Using ISO/IEC 42001 and Standards as Scaffolding
Harmonised standards under the Act are still pending, but you do not need to wait for them to build a credible management system. ISO/IEC 42001, the AI management system standard published in 2023, gives a certifiable structure for policy, roles, risk assessment, impact assessment, lifecycle controls and continual improvement. ISO/IEC 42005:2025, published in May 2025, provides guidance on AI system impact assessments. Neither confers a presumption of conformity with the AI Act, but the records they produce — risk registers, impact assessments, supplier controls, monitoring evidence — map closely onto what the Act asks high-risk providers and deployers to document.
The pragmatic approach is to treat the AI Act as the “what” and a management system as the “how.” Build one control set, map each control to the relevant Act articles and to ISO/IEC 42001 clauses, and generate evidence once. That avoids running parallel compliance programs for regulators, customers and auditors.
Worked Example: A B2B HR-Tech Platform
Consider a mid-sized software company selling a recruiting platform to European employers. It offers three AI features: a candidate-ranking model, a chatbot that answers applicant questions, and a generator that drafts job descriptions using a third-party GPAI model.
- Candidate ranking is an Annex III employment use, so it is high-risk. Before the Omnibus, the company was racing to finish its conformity work by August 2026. Now the hard deadline is 2 December 2027. Instead of pausing, it uses the time to fix training-data documentation, run bias and robustness testing on representative data, design meaningful human review in the recruiter interface, and draft instructions for use that its customers — the deployers — will need for their own obligations.
- The applicant chatbot falls under Article 50. It needed a clear AI disclosure from 2 August 2026, and the company added a hand-off to a human recruiter for contested decisions.
- The job-description generator relies on a GPAI model. The company is a downstream provider of an AI system, not the model provider. It collected its supplier’s Code of Practice status and downstream documentation, confirmed its prompt engineering and light fine-tuning are far below the one-third compute threshold, and added a contractual right to be notified of serious incidents.
The result is a single register, one control framework, and a roadmap that puts the expensive high-risk work on a steady 12-month track rather than a last-minute sprint in late 2027.
Common Pitfalls
- Pausing the program. Teams that stop work after the Omnibus will face the same compressed timeline in 2027, with less institutional memory.
- Ignoring transparency because the system is “low risk.” Article 50 applies regardless of risk tier.
- Treating the vendor as the compliance owner. Deployers have their own duties, and contracts rarely transfer them.
- Classifying once. A feature change, a new customer segment or a new intended purpose can move a system into high-risk scope. Re-classify at every major release.
- Over-documenting the wrong systems. Spending months on minimal-risk tools while high-risk use cases lack data governance is a poor allocation of scarce expertise.
Outlook: What to Watch in 2027
Several developments will shape the next phase. The Commission is expected to publish further guidance on high-risk classification and on the interplay between the Act and sector law. CEN-CENELEC’s harmonised standards will determine how conformity is demonstrated in practice. National market-surveillance authorities are still building capacity, and their early enforcement priorities will signal where scrutiny lands first. Meanwhile, the AI Office’s expanded remit over GPAI-based systems and very large platforms makes it the regulator to watch for generative AI.
Outside the EU, the pattern is similar: binding rules are arriving unevenly, and customers increasingly ask suppliers to evidence governance through certification and audit rather than policy documents. BSI, for example, reported in November 2026 that it had issued its 50th ISO/IEC 42001 certificate, with its 2026 total more than tripling the 11 it issued in 2025.
Conclusion
The Digital Omnibus changed the calendar, not the destination. Prohibitions, GPAI enforcement and transparency duties already apply, and high-risk obligations arrive in December 2027 and August 2028 with standards that are still being written. Organizations that use the next year to inventory, classify, document and monitor their AI systems will reach those dates with evidence in hand and a governance capability that also satisfies customers and auditors. Those that read the Omnibus as a pause will find that the hardest work cannot be done in a quarter.
How Hanoi Technology Can Help
Hanoi Technology helps teams inventory and classify AI systems, build ISO/IEC 42001-aligned controls and engineer the logging, testing and documentation the EU AI Act expects.
Further Reading
- EU AI Act’s High-Risk Deadline: Deferred, Not Cancelled – Cloud Security Alliance
- EU Digital Omnibus on AI Enters into Force – National Law Review
- GPAI Enforcement Is Live: What Security Teams Must Do Now – Cloud Security Alliance
- European Commission Issues Guidelines for Providers of General-Purpose AI Models – WilmerHale
- BSI Awards Its 50th ISO/IEC 42001 Certificate – BSI