The Calendar Behind the Calendar

The amended Article 113 of the EU AI Act is the focus here: it assigns different application dates to different provisions and high-risk classification routes. The practical goal is not to memorize a list of dates, but to learn how to read each date as a transition for a rule, a system, an actor, or a system already on the market. That distinction is what turns the Digital Omnibus from a confusing delay story into a usable legal calendar.

Six milestones, several legal clocks

The first clock belongs to the amending regulation itself. Regulation (EU) 2026/1744 was published in the Official Journal on 24 July 2026 and entered into force on 27 July, the third day after publication. This tells us when the new legal instrument became part of the EU legal order. It does not mean that every amended obligation began on that day.

The next clock belongs to provisions that the AI Act had already made applicable. The amended text keeps Chapters I and II applicable from 2 February 2025, with a specific exception for the new prohibited-practice provisions. That baseline matters because it rules out a common but consequential reading: the high-risk delay is not a pause on every AI Act duty. Rules that already applied do not disappear merely because another set of obligations now has a later date.

From that baseline, the calendar adds several milestones:

The table is deliberately phrased as “what the date attaches to.” In legal implementation, the subject of the sentence is often more important than the date itself. A deadline without its scope is not yet an answer.

The 27 July 2026 milestone has two layers. It is the entry-into-force date of the Digital Omnibus, and it is also the application date for Articles 102 to 110 of the AI Act. Those provisions are therefore a good example of why the two concepts must be kept separate: sometimes entry into force and application happen together, but only because the text says so for that particular set of rules.

The 2 December 2026 milestone has a different character. It activates the new Article 5 points covering systems that generate or manipulate non-consensual intimate material and child sexual abuse material, along with the related provider and deployer rules in Article 5(1a) and (1b). Those provisions are substantive boundaries on harmful practices, not merely preparation duties for future high-risk assessment. The date also appears in a transition for certain synthetic-content systems that were placed on the market before 2 August 2026: those providers must take the necessary steps to comply with the Article 50(2) marking obligation by 2 December 2026.

The two headline high-risk dates apply to the same part of the AI Act but not to the same route. Chapter III, Sections 1, 2, and 3 contain the high-risk requirements whose application is being sequenced. The amended text places systems classified under Article 6(2) and Annex III on 2 December 2027, and systems classified under Article 6(1) and Annex I on 2 August 2028. Article 6(5) is expressly excluded from that wording. That exception is one reason a headline such as “high-risk rules start in 2027” remains too broad without checking the provision.

Finally, 2 August 2030 belongs to a legacy-system condition involving high-risk AI systems intended for public authorities. It is not a fifth general high-risk commencement date. It tells providers and deployers in that specific situation when they must have taken the necessary steps to comply. The distinction between a new system entering a future regime and an existing system receiving a transition is essential to understanding what can persist and what can be lost.

The date that most invites confusion is 2 August 2026. The original AI Act used it as its general application date, and the Digital Omnibus refers back to it in transition language. Providers of synthetic audio, image, video, or text systems that were placed on the market before that date receive until 2 December 2026 to take the steps needed for the Article 50(2) marking obligation. That makes 2 August 2026 a historical boundary for a particular transition, not a current master switch that turns the whole AI Act on or off.

The same logic applies to a system that was already on the market when a future high-risk date arrives. The relevant question is not simply whether a company began selling “an AI product” before the deadline. The amended Article 111 focuses on a type and model, and the legislative explanation says that one lawfully placed unit can matter for other units of the same unchanged type and model. A significant design change can alter that position and trigger the applicable high-risk obligations, including conformity assessment. In other words, a date can mark a change in legal treatment, while a later design decision can move the system across that boundary again.

This is why a useful implementation record has at least three columns. The first names the rule: for example, a prohibited-practice provision, an institutional provision, or a Chapter III high-risk requirement. The second names the system or practice to which the rule attaches: a listed use case, a product-related safety component, or a system generating specified material. The third records the relevant lifecycle and actor: new or existing, unchanged or significantly changed, provider or deployer, and—where relevant—intended use by a public authority. Only after those columns are filled does the date become meaningful. A calendar entry is therefore the end of a scope analysis, not its beginning.

Application is cumulative, not an on/off switch

A regulation can add a new application state without turning off the states that came before it. That is the best way to interpret this calendar. On 27 July 2026, the Digital Omnibus was in force and Articles 102 to 110 applied. On 2 December 2026, the specified Article 5 provisions join the rules already applicable. On 2 December 2027, the relevant Chapter III obligations join them for one high-risk route. On 2 August 2028, they join them for another route. The later dates add obligations for their defined subjects; they do not retroactively make earlier rules optional.

The diagram is not a complete inventory of the AI Act. It shows the structure of the change: several branches grow from an existing legal base. The branches also explain why “the AI Act is in force” and “the AI Act applies from 2028” can both be misleading. The first statement may refer to the regulation as a legal instrument; the second may refer to one category of high-risk obligation. Neither is enough on its own.

Consider a provider of a system used in employment. The fact that the system may fall within an Annex III high-risk use area points toward the 2 December 2027 application date for the specified Chapter III provisions. It does not erase duties that apply earlier, such as the general framework already in operation or the new Article 5 provisions if the system is also used for a prohibited purpose. Nor does the label “employment AI” answer every question about the provider’s obligations. The system’s actual function, role, market status, and relevant provisions still matter.

Now consider an AI component integrated into a lift. The Commission’s political-agreement press release used lifts and toys as examples of product-related high-risk systems whose rules would apply on 2 August 2028. That date is useful orientation, but the legal answer still depends on the Article 6(1) and Annex I route in the final text and on the product-law relationship. A product example cannot be used as a shortcut around classification.

A third example involves a generative system already placed on the market before 2 August 2026. Its provider may face the 2 December 2026 deadline for the steps needed to comply with Article 50(2). That is a transition tied to the system’s market history and a transparency obligation. It is not evidence that the provider’s other AI Act responsibilities all begin on that day or that the system has entered the high-risk regime.

These examples show a general rule about legal states: a system can occupy more than one relevant state at once. It can be subject to an already-applicable general provision, approach a future high-risk date, and also have a transition tied to when it was placed on the market. The states are not mutually exclusive simply because the dates are different.

The 2027/2028 split is a classification question

The two high-risk dates are not arbitrary postponements separated by eight months. They follow two routes through the AI Act. The 2 December 2027 route concerns systems classified as high-risk under Article 6(2) and Annex III. The 2 August 2028 route concerns systems classified under Article 6(1) and Annex I. The next chapter will examine the legal mechanics of those routes; for the calendar, the important point is that the classification route is an input to the date.

This also explains why the Commission’s public examples cannot be converted into a universal sector table. Biometrics, critical infrastructure, education, employment, migration, asylum, and border control illustrate the Annex III side of the schedule. Lifts and toys illustrate the product-related side. The examples clarify the architecture, but they do not eliminate the need to check the final text and the facts of the system.

The distinction between application and transition becomes sharper when a system changes over time. The amended Article 111 explains that, for certain high-risk systems already placed on the market or put into service, the relevant grace period can follow the same type and model if at least one unit was lawfully placed on the Union market before the relevant date and the design remains unchanged. A significant design change can trigger full compliance with the applicable high-risk provisions, including conformity assessment. The date is therefore not the only time variable. The transition trigger is also a change in the system.

The practical consequence is a record-keeping problem before it is a calendar problem. An operator needs to know what the system does, which legal route may classify it, who provides or deploys it, when the relevant type and model first entered the market, whether its design changed significantly, and whether a public authority is involved. Without those facts, selecting a date from the table is guesswork dressed as precision.

The most reliable question is not “When does the AI Act start?” It is a sequence of narrower questions:

1. Which provision or group of provisions is at issue?
2. Is the system already within an applicable general rule, a new prohibited-practice rule, one of the two high-risk routes, or a transition?
3. What is the relevant actor’s role—provider, deployer, or both within the same undertaking?
4. Is the system new, already placed on the market, unchanged, significantly changed, or intended for a public authority?
5. What date does the final text attach to that combination?

That sequence does not replace legal analysis, and it is not individualized legal advice. It is a way to avoid a known category error: treating a date as if it were self-explanatory.

The calendar’s deepest message is therefore about accumulation. The Digital Omnibus did not move the entire AI Act from one day to another. It kept an existing legal base, activated some institutional and rights provisions in 2026, separated two high-risk routes across 2027 and 2028, and preserved a narrower legacy transition into 2030. To understand why the routes receive different dates—and why product law, safety functions, and conformity assessment sit behind the split—we now have to examine what makes a system enter each route in the first place.

References

• Regulation (EU) 2026/1744 — official English text
• Regulation (EU) 2026/1744 — Publications Office record
• Commission press release IP/26/1024