Picture a team training for months for a match, learning the rules, following them to the letter, only to discover three weeks before kickoff that the game is postponed by a year and the rules loosened. The most conscientious players find themselves not rewarded but disadvantaged: they spent for nothing more than those who dragged their feet. That is, near enough, what has just happened to the firms subject to Europe's law on artificial intelligence. This piece tells how postponing a rule days before it takes effect turns the rule itself into a source of risk.
1 The fact
The rule pushed back three weeks before the deadline.
The announcement
A timetable upended just before the deadline
The most demanding obligations of the AI Act, Europe's law on artificial intelligence, were to apply from 2 August 2026. On 7 May 2026 the Council and the European Parliament agreed on a text, the "Digital Omnibus," that pushes these "high-risk" obligations back by more than a year: late 2027 for most systems, and as far as 2028 for others. The postponement thus comes just weeks before the date it cancels, when the firms concerned had built their entire preparation around that deadline. The finish line, known long in advance, was moved at the last minute.
2 The pioneering law
The world's first comprehensive framework, and the Brussels effect.
The context
The world watched Europe write the rule
The AI Act is no ordinary law: adopted in 2024, it is the world's first comprehensive legal framework dedicated to artificial intelligence. Europe was playing a role it cherishes, that of reference legislator: unable to lead in innovation, it meant to lead in setting the rules, counting on what lawyers call the "Brussels effect," the tendency of European norms to impose themselves well beyond the continent, because it is simpler for a global company to align with the strictest rule. It is precisely this pioneer status that the postponement undermines: a rule its author pushes back inspires less than a rule it enforces.
3 What is postponed
The detail of the delay, and what remains in force.
The mechanics
The "high-risk" part shifted, transparency kept
The delay does not concern the whole law, but its heaviest core: the obligations applying to so-called "high-risk" systems, those touching sensitive areas such as employment, credit or justice. For stand-alone systems listed in Annex III, the deadline moves to 2 December 2027; for AI embedded in already-regulated products (Annex I), to 2 August 2028. Other provisions, notably the Article 50 transparency obligations that require flagging AI-generated content, largely keep their original timeline. Finally, the postponement takes legal effect only once the text is formally adopted and published: until then, the old timeline remains the law, which adds uncertainty to uncertainty.
Stand-alone systems (Annex III)
Dec 2027
instead of 2 August 2026.
AI in regulated products (Annex I)
Aug 2028
the longest postponement.
4 The laggard's bonus
The cost of early compliance, turned into dead loss.
The paradox
Behaving well achieved nothing more than dragging one's feet
Here is the most disturbing effect. To be ready by 2 August 2026, many firms incurred significant spending: audits, documentation, bringing systems into compliance, hiring lawyers and experts. Those costs were paid early, on the strength of an announced timetable. The postponement does not refund them: it is a sunk cost. Meanwhile, the firms that had bet on a loosening, or simply delayed, gain more than a year's reprieve without having spent a thing. The rule, meant to reward good conduct, in effect penalised the most diligent and rewarded the wait-and-see. That is the laggard's bonus: the timetable punished those who took it seriously.
The key idea
When a deadline is pushed back at the last minute, it transfers an advantage: from the one who prepared to the one who waited. Diligence becomes a bet, and waiting a winning strategy.
5 Regulatory risk
When the rule itself becomes a hazard.
The notion
You can prepare for a rule, not for its instability
A company knows how to manage many risks: demand softening, prices moving, a competitor emerging. Regulatory risk is of another kind: it bears on the framework itself, on the possibility that the rule changes, is postponed, tightened or dropped. And it is particularly hard to manage, because it cannot be hedged: no insurance refunds a moved timetable. The AI Act episode is a textbook case. It was not the content of the rule that proved costly, it was its uncertainty. And that uncertainty is all the more corrosive for emanating from the regulator itself, the very party from which stability was expected.
6 Predictability betrayed
A rule is only worth its stability.
The principle
What makes a rule valuable is that one can rely on it
A rule's first use is not to be perfect, but to be known and stable: that is what lets everyone organise their decisions around it. A rule that can be pushed back three weeks before it applies loses part of that use, even when the delay is justified. For it sends a signal: the announced timetable does not really bind. From then on, the next deadline will be taken less seriously, and the incentive to prepare early weakens. A regulator's credibility rests on a fragile capital: the conviction that it will keep what it wrote. Every postponement made without pressing necessity chips at that capital, and the next commitment is worth a little less.
7 Why the delay
The stated reasons, taken seriously.
The motives
Competitiveness, lagging standards, complexity
The delay did not come from nowhere, and its supporters put forward serious arguments. First competitiveness: Europe fears hobbling its AI firms against American and Chinese rivals subject to lighter frameworks, and worries about legislating faster than it innovates. Then the technical side: the detailed norms and standards meant to guide compliance were not all ready, so firms were being asked to comply with requirements whose precise terms were still missing. Finally complexity: applying so vast a text on time looked arduous for administrations and operators alike. Delaying so as not to enforce an unenforceable rule is a defensible choice; it is the timing of the delay, more than its principle, that is the problem.
8 The counterpoint
What this reading should not overstate.
The analysis's honesty
Better an enforceable rule than a rushed one
It would be unfair to see in this delay only a surrender. Force-marching a rule whose technical terms are not ready would also have a cost, perhaps a higher one: legal insecurity, litigation, box-ticking compliance. A legislator that corrects course rather than digging in sometimes shows wisdom. Moreover, the AI Act is neither abandoned nor gutted: simplifying and staggering a timetable is not repealing a law, whose principles and part of the obligations remain. Finally, compliance costs incurred early are not always lost: the documentation and governance put in place keep their value. The reproach, then, is not to revise, but to have done so so late that the most diligent were punished.
9 What it changes
For European credibility and how players behave.
The effects
A new incentive: wait rather than prepare
The episode leaves marks beyond AI. It weakens the Brussels effect: a framework its author pushes back exports less well, since its strength came precisely from its firmness. Above all it installs a new incentive among firms: facing the next European deadline, many will be tempted to wait until the last moment, betting on a delay, rather than complying early. That is the opposite of what a regulator seeks. Finally, it feeds a broader debate on how to regulate a technology that moves faster than the law: fixed rules that time quickly makes obsolete, or more flexible frameworks at the price of less predictability? The AI Act, in searching for itself, illustrates the difficulty of holding both.
10 When the rule moves, trust is what suffers
The conclusion: the real cost is credibility.
The meaning of the episode
Stability is a value in itself, sometimes more than content
The story goes beyond artificial intelligence. It recalls that, for those who must obey it, a rule is worth as much for its stability as for its content: one can adapt to a strict requirement, provided it holds. The AI Act delay may have avoided premature enforcement; it certainly damaged something quieter and more precious, trust in the regulator's word. The real cost of the episode is not read in the audits paid in advance, but in the question firms now ask themselves before each European deadline: "should we really prepare, or wait for the next delay?" A rule that prompts that question has already lost part of its power.
The compass
①
The rule was pushed back just before the deadline. The EU postpones the AI Act's "high-risk" obligations from late 2027 to 2028, when they were to apply on 2 August 2026 (agreement of 7 May 2026).
②
The diligent are penalised. Those who complied on time paid a sunk cost, while the laggards gain more than a year's reprieve without spending a thing: the laggard's bonus.
③
The real cost is predictability. The delay may be justified, but it damages the regulator's credibility and rewards waiting over preparing. This piece frames a debate; it is not advice.
Read alongside: The climate tax shaped like a customs duty (another European regulation with competitive effects, the Brussels effect) and The leaner AI gets, the more it devours (the same technology, energy side). See also The code that attacks the grid writes itself (the AI the rule seeks to frame). Reference: abbreviations & acronyms (AI, EU).