The EU AI Act high-risk deadline moved. Four things did not

What the amendment actually did

Regulation (EU) 2026/1744 was signed on 8 July 2026, published in the Official Journal on 24 July, and entered into force three days after that. It amends the AI Act in several places, but one change accounts for most of the coverage. The obligations for high-risk systems no longer apply from 2 August 2026.

They apply from 2 December 2027 for stand-alone systems classified under Annex III, and from 2 August 2028 for systems built into products covered by Annex I. Sixteen months for the first group, two years for the second.

The Commission proposed this in November 2025 with a condition attached. The dates would move only once the Commission confirmed that the standards and supporting tools were in place. Parliament and Council removed that condition and set fixed calendar dates instead. Nothing now suspends them, and nothing brings them forward.

The reason given for the delay is that the ground was not ready. No harmonised standard under the Act has been cited in the Official Journal, so the presumption of conformity those standards create still does not exist. The European standards bodies are working to an accelerated plan aimed at late 2026.

The amendment does other things. It reinstates the duty on providers to register high-risk systems in the EU database, including those claiming an exemption from the classification. It postpones national regulatory sandboxes to 2 August 2027. It adds two prohibitions.

None of that changed the headline, and the headline is what most compliance plans were built on.

The part that did not move

Article 50 applies from 2 August 2026, exactly as it always did. It is the transparency chapter, and it is the part of the Act that reaches ordinary organisations first. The market surveillance provisions in Articles 73 to 94 arrive on the same date, which is what gives the rest of it teeth.

Three of its paragraphs bind deployers rather than providers. Run a system that recognises emotions or sorts people by biometric data, and you have to tell the people exposed to it. Publish a deepfake, and you have to disclose it at first exposure.

The third is the one that catches communications teams. Text generated by AI and published to inform the public on a matter of public interest has to be labelled, unless a person has reviewed it and someone holds editorial responsibility for it.

No supplier can take over a duty the Act places on the organisation using the system.

That is the point most often missed. An organisation that has decided the Act is a problem for its suppliers has still bought the obligations in Article 50(3) to (5) along with the software, and cannot buy its way back out of them.

One transitional window follows, in a new Article 111(4). Providers whose generative systems reached the market before 2 August 2026 have until 2 December 2026 to bring their output into line with the marking duty in Article 50(2). Anything placed on the market after that date marks from the first day.

Two new prohibitions arrive on 2 December 2026 as well. They cover the generation of non-consensual intimate imagery of identifiable people, and child sexual abuse material. Because they sit in Article 5, they carry the heaviest penalty tier in the Act.

The part that was already due

The original prohibitions in Article 5 have applied since 2 February 2025. So has the obligation in Article 4 covering the competence of the people who operate these systems, although the amendment softened it from ensuring a sufficient level to taking measures that support its development. The new wording also adds that the duty does not require anyone to guarantee a particular level of AI literacy in individual people. The obligation became weaker in substance, not only in tone. The recitals to the amendment say at the same time that AI literacy should be a strategic priority regardless of whether it is a legal obligation and regardless of any penalties. That is unusually direct for a legal text: the case for competence is made on operational grounds, not on enforcement.

The rules for providers of general-purpose AI models have applied since 2 August 2025, together with the governance structure and most of the penalty framework.

The levels are unchanged. The percentages apply to total worldwide annual turnover for the preceding financial year, and the higher of the two figures is the ceiling.

Small and medium enterprises get the lower of the two instead. The amendment extended a similar cap to small mid-caps, meaning firms of 250 to 750 employees, for the middle and lowest bands but not for a breach of the prohibitions.

Whether any of this reaches you

High-risk is a narrower category than the phrase suggests. Which side of the line a system falls on decides what the work has to cover, and the line rarely runs where people assume.

Annex III lists the stand-alone cases. Systems used in recruitment and in decisions about promotion or termination. Systems that decide access to education, or score examinations.

Also on the list: systems that determine eligibility for essential public benefits, for emergency services, for credit, or for the pricing of life and health insurance. Systems used in law enforcement, in migration and border control, and in the administration of justice.

Annex I is the other route, and it covers systems that are safety components of products already regulated at EU level, from machinery and medical devices to vehicles and lifts. Those move to 2 August 2028.

The amendment also narrowed what counts as a safety component. A system used only for non-safety-related aspects of user support, performance optimisation, service efficiency, automation, convenience or quality control is not a safety component, and so does not reach high-risk by that route. That lifts a whole class of systems out of the grey zone.

Public authorities get a transitional rule of their own. Systems intended for use by public authorities that were already on the market before the date of application have until 2 August 2030 to comply. Four years, and the longest period anywhere in the Act.

Most organisations run neither. What they run is a general-purpose assistant, a document tool, a service chatbot. Those usually fall outside the high-risk regime, and did before the amendment too, so the postponement moves nothing for them. What they do not fall outside is Article 50.

The contracts that now point at the wrong date

There is a practical consequence that surfaces in procurement before it surfaces anywhere else.

Supplier agreements signed in the past eighteen months routinely carry a clause requiring compliance with the AI Act by 2 August 2026. That clause now points at a date on which most of the obligations it refers to do not apply.

Whether it still binds the supplier to anything depends on how it was drafted. A clause tied to the Act as it stands from time to time survives the amendment. A clause tied to a fixed date does not, and the drafting was rarely that careful.

The same is true in the other direction. A public buyer who wrote the original date into a framework agreement may find a supplier reading the postponement as relief from a term the buyer thought was fixed. It is cheaper to find out now than at the first delivery review.

Who enforces it, and in Sweden that is still open

The Act applies whether or not a member state has finished appointing anybody. The Swedish arrangement is not finished, and the obligations hold regardless.

A government inquiry reported in October 2025 with a proposed structure. The Post and Telecom Authority would lead market surveillance, alongside the data protection authority, the financial supervisor, the medical products agency and the national accreditation body in their own domains.

It has not been enacted. There is no bill. The penalty framework, meanwhile, has been available since August 2025.

One exception is worth knowing. The amendment gives the AI Office exclusive supervision over systems built on a general-purpose model from the same provider, and over systems that form part of very large online platforms, with its own powers to investigate, inspect and impose penalties. For that part of the market, the Swedish appointments are not the deciding factor.

The practical consequence is small but real. If something goes wrong with a system you deploy this autumn, the question of who asks you about it, and under which statute, has no settled Swedish answer yet. That is a reason to keep your own record, not a reason to wait.

Why sixteen months is not a reprieve

The deadline moved backwards. Deployment did not.

In 2024, 13.5% of EU enterprises with ten or more employees used AI. In 2025 it was 20.0%. Sweden went from 25.1% to 35.0%, and among large EU enterprises the figure reached 55.0%.

An organisation that files the postponement as time won will meet the high-risk regime in December 2027 with a larger estate than it has today, assembled faster, and documented no better.

The inventory does not get easier by being deferred. It gets longer.

There is a second cost, harder to see. Sixteen months is long enough for the people who built the first compliance plan to move on, and for the plan to become a document nobody can explain. The work that survives a postponement is the work that was written down properly.

What the sixteen months are for

Three things are worth doing before the end of this year, and none of them wait on standards that do not yet exist.

Know what you are running. Most organisations cannot produce a list of the AI systems in use across the business. It is the first document all of this requires and the slowest one to assemble, and nothing about December 2027 makes it shorter.

Sort that list against Annex III. The classification decides whether the new dates are a deadline or a non-event for you, and it decides which of the two dates applies. For a system that is a safety component of an already regulated product, that assessment is made together with the notified body.

Then read Article 50. It applies to how you use systems rather than what you build, and it is the part of the Act that most organisations will actually have to do something about.

The full calendar, every date and what applies from when, sits on our EU AI Act page.