In the early hours of May 7, 2026, negotiators for the Council of the EU and the European Parliament reached a provisional agreement on the Digital Omnibus on AI, the first set of amendments to the AI Act since it was adopted. The headline change: obligations for high-risk AI systems, due on August 2, 2026, now apply from December 2, 2027 for stand-alone systems and from August 2, 2028 for AI embedded in regulated products.
The deal is not law yet. It still needs formal endorsement by both institutions, legal-linguistic revision and publication in the Official Journal. According to IAPP, the Council expects formal adoption “in the coming weeks” and Parliament intends to finish before August 2.
/ 01What was agreed
Based on the summaries published by IAPP, PPC Land and Abreu Advogados, the main points are:
- New high-risk dates. December 2, 2027 for Annex III systems, which cover areas such as biometrics, critical infrastructure, education, employment, law enforcement and border management. August 2, 2028 for AI embedded in products covered by Annex I, such as medical devices. The deal replaces the Commission’s proposed trigger, tied to the availability of standards, with fixed dates.
- Registration returns. Providers who consider their system exempt from high-risk classification must still register it in the EU database.
- New prohibition. AI systems that generate non-consensual intimate content or child sexual abuse material are banned, with compliance from December 2, 2026.
- Shorter grace period for content marking. Generative systems already on the market before August 2, 2026 have until December 2, 2026 to meet marking requirements, three months rather than six.
- Bias detection. Processing special categories of personal data to detect and correct bias is allowed only where strictly necessary, with safeguards.
- Sector overlaps. AI-enabled machinery is handled through the Machinery Regulation, and the Commission will address overlaps with other sectoral laws, including medical devices, connected cars and toys.
- Other changes. National regulatory sandboxes are due by August 2, 2027; some SME privileges extend to small mid-cap companies; and the AI literacy duty shifts from providers and deployers to the Commission and member states.
/ 02Reactions
The reactions IAPP collected split predictably. Cyprus’s deputy minister for European affairs, Marilena Raouna, said the deal “supports our companies by reducing recurring administrative costs.” Consumer group BEUC said the omnibus “rolls back key consumer protections which had barely been adopted a year ago,” while welcoming the shorter transparency deadline and the return of the registry. Industry groups were unimpressed in the other direction: CCIA Europe called the high-risk postponement the “bare minimum,” and DIGITALEUROPE said the lack of medtech relief was a missed opportunity. Kai Zenner, an adviser to MEP Axel Voss, said the changes were positive but “minor steps will not solve the structural problems.”
/ 03Why the extra time is not a reprieve
Sixteen more months for Annex III systems is welcome for organisations that were not going to make August. But the substance of the high-risk regime did not change. Providers still need risk management, data governance, technical documentation, logging, human oversight and accuracy controls. The registration obligation is, if anything, broader than the Commission proposed.
Those obligations share one requirement that is easy to underestimate: you have to be able to show your work. An auditor or market surveillance authority will want to know what data a system used, what rules it applied, how a given output was produced, and who reviewed it. For systems built on large language models, that is precisely the part that is hardest to retrofit. A model’s answer can be logged; the reasoning behind it usually cannot be reconstructed from the weights.
This is where structured knowledge earns its place. When the domain facts, policies and decision rules a system relies on live in an explicit, versioned knowledge graph, each output can be tied to the specific nodes, sources and rule versions that produced it. That turns documentation from a narrative written after the fact into a by-product of how the system runs. It also makes the bias-detection carve-out easier to work within, because the data a system touches is explicit and can be scoped to what is strictly necessary.
The deadline moved. The burden of proof did not.
/ 04What to do now
- Track adoption. Until the amendments are published in the Official Journal, the original August 2 timetable is still the text in force. Plan for both outcomes until then.
- Inventory and classify. Use the time to settle which systems are Annex III, which are Annex I, and which you consider exempt — and remember the exempt ones still need registering.
- Build evidence into the architecture. Prefer designs where every output can be traced to its inputs and rules. That is cheaper than reconstructing evidence for a regulator later.
- Check the December 2026 dates. Content marking and the new prohibition arrive a year before the high-risk rules.
An architecture that records the evidence chain behind each output does not make a system compliant on its own, but it gives compliance teams something concrete to point to when the regulator asks how an answer was reached.
The omnibus buys time, not relief. The high-risk duties still require a provider to show what data a system used, which rule it applied and who reviewed the output, and that is an architecture decision, not a documentation sprint before December 2027.
Take creditworthiness scoring, one of the Annex III uses. In a governed graph each decision links to the applicant data it read, the credit policy version in force that day, the model version and the reviewer who signed off. The Act’s obligations become nodes too, mapped to the controls that meet them. A compliance officer answering a market surveillance query can then hand over a path through the graph instead of a narrative reconstructed from logs. Deterministic inference is what makes that path credible: replaying last March’s decision against last March’s graph gives the same result with the same evidence. A self-declared exemption fits the same pattern, recorded with the reasons that justify it before it goes into the EU database.
How we’d model it
Credit decision—appliedRule→Policy versionCredit decision—reviewedBy→Human overseerPolicy version—satisfies→AI Act obligationAI system—registeredIn→EU database
- EU agrees to amend AI Act, clarifies overlap with machinery rules — IAPP iapp.org ↗
- EU AI Act gets its first real haircut - high-risk deadlines pushed to 2027 — PPC Land ppc.land ↗
- AI OMNIBUS: Council and European Parliament Reach Provisional Agreement — Abreu Advogados abreuadvogados.com ↗
See how Aktiver turns raw data into knowledge graphs your AI can reason over — deterministic, traceable, behind your firewall.