Last week, a client who makes export equipment asked one question over the phone: Wasn't that EU AI Act delayed? Can we put our part on hold for now?
Behind this question lies a very common understanding: an extension equals loosening restrictions. But if you lay out the timetable for this round of revisions in 2026, you will find that what it extends and what it does not extend are two different categories of things.
What has been postponed is the application date for the high-risk obligations—from August 2, 2026 to December 2, 2027, a full sixteen months. What has not been postponed are the two actions on December 2, 2026. Today is November 17, with fifteen days left until that date.
1. First, line up the five dates
The EU Artificial Intelligence Act is Regulation (EU) 2024/1689, which entered into force on 1 August 2024 and is itself a law that applies in stages. The 2026 Digital Omnibus on AI (Regulation (EU) 2026/1744) is its first substantive amendment: it was published in the Official Journal of the European Union on 24 July 2026 and entered into force on 27 July.
| Date | What happened | meaning for enterprises |
|---|---|---|
| 2026-08-02 | The main provisions of the bill apply; the transparency obligation under Article 50 takes effect | When interacting with people, disclose that it is AI; deepfakes must be disclosed |
| 2026-12-02 | Two additional prohibitive provisions come into force; the transition period for labeling generative systems already on the market before 2026-08-02 expires | Fifteen days later, two things expire at the same time |
| 2027-08-02 | General-purpose models already on the market before 2025-08-02 must comply. | Model-side inventory cleanup |
| 2027-12-02 | Annex III High-risk Obligations for Stand-alone Deployment (originally 2026-08-02, postponed by sixteen months) | Recruitment, education, credit, law enforcement, critical infrastructure |
| 2028-08-02 | Annex I Systems incorporating regulated products (medical devices, machinery, toys, etc.) | Hardware product line with AI features |
The easiest place to misread this table is in the fourth and fifth rows. They look like "there are still one or two years, no rush," but note that the nature of the two dates before and after them is completely different: the December 2 entry governs artifacts and functions, while the two 2027 entries govern processes and obligations.
| Date | What does it govern? |
|---|---|
| 2026-12-02 | It governs “whether what you produce has ever been labeled” and “whether you have built this type of feature.” What it checks is the existing stock, not the process. |
| 2027-12-02 | What it regulates is "whether your high-risk system has technical documentation, data governance, human oversight, and post-market monitoring." What it examines is the system. |
2027-12-02
What it regulates is "whether your high-risk system has technical documentation, data governance, human oversight, and post-market monitoring." What it examines is the system.
Systems can be remedied, but outputs cannot. For a piece of synthetic content that has already been generated and published, adding a label after the fact is meaningless.
II. What was due on December 2 was actually two very small matters
The first is the addition of two new prohibitive clauses to Article 5: generating or manipulating realistic intimate imagery and child sexual abuse material. The logic of such clauses is not "rectify and continue using," but that in principle they must not be developed, provided, or deployed. For the vast majority of enterprises doing ERP, manufacturing, or enterprise management software, this clause will not be encountered. But for product lines doing content tools, marketing material generation, or image processing, it is necessary to check whether there are any borderline switches in the feature list.
The second point deserves more attention from most enterprises: for generative systems already placed on the market before August 2, 2026, the machine-readable marking obligation under Article 50(2) runs through a transitional period until December 2. This transitional period is set out in the amended Article 111(4), with a grace period of four months.
Three things often misread as "postponed together":
First, the grace period only covers the marking obligation in Article 50(2) and cannot cover the rest of Article 50. The requirement that chatbots must state “you are interacting with AI” has already applied since August 2, 2026.
Second, the timetable for obligations of general-purpose AI models has not changed, including the group on transparency, documentation, and systemic risks.
Third, the penalty provisions were also left unchanged; instead, they became more detailed—to be elaborated in the next section.
There is also a technical detail worth thinking through in advance: Article 50(2) requires the marking to be "machine-readable and detectable as AI-generated or manipulated"; the technical solution is not restricted, and both watermarking and metadata count as options. But it has an exception—if the system only performs auxiliary functions for standard editing, or does not substantially change the input data provided by the deployer or its semantics, it is outside the scope.
In practice, a much easier approach is the reverse: label everything that can be labeled, use exceptions only where truly necessary, and write a short internal memo for each feature that relies on an exception, explaining why it applies. When regulators or clients come asking, a memo is far more useful than a verbal explanation.
III. What this round of amendments truly changes is the penalty provisions and jurisdiction
The postponement is easy to talk about in the news, but the part of the amendment that is more practically binding on companies is the change in the enforcement mechanism.
Penalties are divided into three tiers: for violating prohibited provisions, a maximum of €35 million or 7% of global annual turnover, whichever is higher; for violating wide-ranging obligations for operators, deployers, and notified bodies as well as Article 50 obligations, a maximum of €15 million or 3%; for providing incorrect information to regulators, a maximum of €7.5 million or 1%. For SMEs and startups, the lower of the two applies.
What is more worth watching is the newly added recurring penalties: after the 2026 law amendment, the EU AI Office can impose daily charges for continuing violations, capped at 5% of average daily turnover. One-off fines can be amortized as project costs, but daily penalties cannot — they turn 'knowing but not fixing it yet' into something that costs money every day.
The maximum daily penalty that the AI Office may impose for persistent non-compliance, based on average daily turnover
Meanwhile, the AI Office has obtained exclusive regulatory authority over two types of systems: AI systems built on general-purpose models, and AI systems embedded in very large online platforms and search engines under the Digital Services Act. The toolbox includes investigations, information requests, on-site inspections, binding commitments, and periodic fines.
So why delay it further? The official reason given is quite simple: the harmonised technical standards that companies use to demonstrate compliance have still not come out, and the development of competent authorities and conformity assessment bodies in various countries is also slow. DIGITALEUROPE estimates that compliance costs across Europe are about €3.3 billion per year, and the initial cost for a company of around fifty people is between €320,000 and €600,000; 47 associations, including MedTech Europe, DigitalEurope, and COCIR, jointly support the postponement.
There is an inference that is not often pointed out: the reason for the delay is precisely the “absence of standards”. In the stage before standards are issued, so-called presumption of compliance can only be built on the company’s own records—what classification you did, what the basis was, who reviewed it, and when it was changed. This is not paperwork; this is the only way to provide evidence during the period when standards are absent.
There is another piece of news that is easily treated as a talisman: the European Commission's implementing regulation on AI regulatory sandboxes (EU) 2026/2487 was adopted on November 10, 2026, published on November 11, and entered into force on December 1, 2026. But the official text clearly states: participating in a sandbox does not automatically exempt obligations.
The signals on China's side are consistent in direction
On November 14, Wu Qing, Chairman of the China Securities Regulatory Commission, delivered remarks at the First China Capital Market Academic Annual Conference, listing "accelerating research on responsibility determination, data security, customer protection, and systemic risk, effectively preventing algorithm black boxes and AI cyberattacks, and enhancing the resilience of financial infrastructure" as key research issues for the next stage.
The wording on both sides differs, yet the point they land on is much the same: neither is in a hurry to add new penalties; both first demand that "when something goes wrong, it can be explained clearly." For a company that builds ERP and business systems, translated plainly, that means: logs that can be retained, a decision chain that can be replayed, and a trace of whatever was changed.
IV. Three-Month Action Plan for the CIO
If you compress all of the above into actions that can be scheduled on a calendar, there are actually only four. None of them require changing the product architecture right now, but all of them need to have a clear answer before the Spring Festival.
| Action | What specifically to do | Deliverables |
|---|---|---|
| Inventory List | Create a table by system and functional module: whether it is placed on the EU market, intended use, deployment scenario, whether the enterprise is the provider or deployer, preliminary risk assessment and basis | AI system checklist, readable on one page |
| Supplementary identifier | Add machine-readable identifiers to generative outputs; write internal memos one by one for features that rely on exceptions | Identification Scheme + Exception Explanation Memo |
| Amend the contract | Clarify with model vendors and cloud service providers: who is responsible for labeling, who provides technical documentation, how soon serious incidents must be notified, and whether they can cooperate with regulatory inquiries | Responsibility Allocation Table for Contract Annex |
| Keep evidence | Make 'explainability' an institutional system: classification basis, reviewers, version changes, log retention period | Deliverable regulatory dossier package |
The three most common pitfalls:
First, treat a postponement as an exemption from inspection. A postponement only changes the effective date of the obligation, not the starting point for evidence; for a system that is already running in 2025, what must be submitted in 2027 should start being accumulated now.
Second, treating the sandbox as an exemption. The original text of the regulations clearly states that it does not automatically grant an exemption; treating it as proof of compliance will cause trouble.
Third, treating 'we are not in Europe' as a territorial exemption. The hook in Article 2 of the bill is 'placing on the market' and 'using the output in the EU'; it has nothing to do with where the company is registered. A model built by a Shenzhen team, with its API sold to a call center in Poland, whose output acts on people within the EU, is within scope.
Back to that client's question. They make packaging equipment with visual inspection, sold to Italy and Poland. What really needs answering isn't "should we remediate now?" but three more specific questions: Is this equipment's AI functionality embedded in a product governed by the Machinery Directive (under Annex I, August 2028); are we a provider or a deployer; and does the inspection report generated by the equipment contain any synthetic content that needs to be labeled.
After answering the three questions, you will most likely find that there is only one small thing that needs to be done before next month, but this thing must be done in the system and cannot be made up for with documentation.
Closing in one sentence:The extension changes "when compliance is required," not "when to start keeping evidence." The former can be scheduled; the latter has been ticking since the day the system went live.
References: Regulation (EU) 2024/1689 (AI Act) and Regulation (EU) 2026/1744 (Digital Omnibus on AI, published on 2026-07-24, effective 2026-07-27), including their provisions and Article 111(4) and the transitional arrangements in Article 113; Commission Implementing Regulation (EU) 2026/2487 (AI regulatory sandbox, adopted on 2026-11-10, published on 11-11, effective on 2026-12-01); EU AI governance dossier (entry in the European Commission’s digital rules progress report of 2026-11-04); MAAK Law, “EU AI Act Article 50: transparency duties explained”; DIGITALEUROPE compliance cost estimates and the joint signing by 47 industry associations (as cited in public reports from November 2026); Sina Finance, “Medical AI Going Global to the EU: How to Respond to the ‘Artificial Intelligence Act’” (2026-11-16); the speech by Wu Qing, Chairman of the China Securities Regulatory Commission, at the opening ceremony of the First China Capital Market Academic Annual Conference (2026-11-14, csrc.gov.cn), and same-day reports by Securities Times and China Securities Network. The client scenarios in this article are examples synthesizing multiple projects and do not refer to a single client.
Zhongdi Information · Enterprise Management Digitalization Observations · Regulatory milestones are subject to the provisions of the Official Journal of the European Union and guidance from competent authorities; this article does not constitute legal advice.
