ATLAS
← Back to News and Insights

National Assembly Library details EU AI Act high-risk delays to 2027 and 2028

3 min

The 8 July Digital Omnibus keeps the EU’s risk-based model, narrows high-risk scope, expands SMC support and strengthens controls over harmful AI.

LinkedIn Post
National Assembly Library details EU AI Act high-risk delays to 2027 and 2028
South Korea · Photo: Louie Nicolo Nimor / Unsplash

The National Assembly Library published its review on 17 November 2026 of Regulation (EU) 2026/1744, the EU’s Digital Omnibus on AI. Enacted on 8 July, the regulation amended the EU AI Act and postponed the application of high-risk AI obligations to 2 December 2027 for standalone high-risk AI systems and 2 August 2028 for product-related high-risk AI systems.

For firms within the AI Act’s scope, the change resets the implementation timetable rather than removing the high-risk regime. Businesses should revise their classification, documentation, conformity assessment and governance roadmaps against the two new dates, while continuing to track the standards, guidance and supervisory arrangements that prompted the delay.

Two application dates, not one

The Library says the amendment responds to concerns that the high-risk obligations were approaching in August 2026 without the necessary implementation infrastructure. Relevant harmonised standards and guidance had not been completed, while the designation of competent authorities and conformity assessment bodies had been delayed. The resulting compliance costs could increase and innovation could be hindered.

The amendment retains the AI Act’s risk-based structure but gives firms a longer and more predictable preparation period. The Library specifically highlights the use of fixed dates as improving the predictability of compliance planning.

The postponement is divided by category:

- standalone high-risk AI systems face the obligations from 2 December 2027;
- high-risk AI systems linked to regulated products face them from 2 August 2028.

The later date is not a general exemption from the AI Act, and the source does not describe the high-risk requirements as being withdrawn.

Relief is targeted

The amendment excludes AI systems used for ancillary functions unrelated to safety from high-risk classification in principle. It also reduces overlapping obligations where the AI system is already covered by product safety legislation.

Support measures previously limited to small and medium-sized enterprises are extended to small mid-cap companies, or SMCs. These measures include simplified technical documentation and quality management requirements, reduced fines and priority access to regulatory sandboxes.

The AI literacy requirement is also softened. The Library describes the change as moving from an obligation to “ensure” AI literacy to one to “support” it, a translated rendering of the Korean source.

The amendment therefore combines a delayed timetable with narrower scope and targeted relief. It does not treat every smaller business or every AI use case as outside the regime.

Harmful AI controls become stricter

The same amendment strengthens controls in areas involving serious harm. It prohibits the launch and use of AI that generates non-consensual sexual deepfakes or child sexual abuse material.

It also expands the AI Office’s supervisory and enforcement authority over AI systems based on general-purpose AI models, among other systems. The change places greater attention on the relationship between general-purpose models and downstream AI systems, even as some high-risk compliance dates move out.

Kim Seong-wan, head of the National Assembly Library’s Legal Information Division, said South Korea’s AI Basic Act, which took effect in January 2026, is also at an early implementation stage. He identified “the allocation of obligations according to the characteristics of regulated parties” and improved predictability in implementation as important issues, in a translation from Korean.

For compliance planning, the Library’s key dates are 2 December 2027 for standalone high-risk AI systems and 2 August 2028 for product-related high-risk AI systems.

Official source: assembly.go.kr
AIAI FrameworksAI Risk TiersAI Use RestrictionsAI Governance

Permanent link to this Atlas analysis

Continue with coverage connected by market, topic and operating context.

Continue with Atlas

Move from this development into the relevant research, comparison and workflow.