AI Act Omnibus: Political Agreement Reached, Article 4 Holds at August 2

The European co-legislators reached a political agreement on the Digital Omnibus on AI overnight on 6–7 May 2026, eight days after the 28 April trilogue collapsed and a week before the rescheduled 13 May session. The amending regulation still requires legal review and Official Journal publication before 2 August 2026 to take effect, but the substantive shape of the deal is now public.

For organisations preparing under the AI Act, the deal is most usefully read by what it changes and what it leaves intact.

What the deal postpones

• Annex III high-risk obligations (Articles 9–15, including risk management, data governance, technical documentation, human oversight under Article 14, and accuracy and robustness) move to 2 December 2027.
• Annex I high-risk obligations (AI embedded in products covered by EU sectoral safety law — machinery, MDR, IVDR, type-approved vehicles) move to 2 August 2028. A new mechanism gives the Commission implementing-act power to resolve overlaps between AI Act requirements and sectoral law.
• Article 50(2) content-marking obligations for providers of generative AI move to 2 December 2026.
• New Article 5 prohibitions on AI-generated non-consensual intimate imagery and CSAM apply from 2 December 2026.
• Member State AI regulatory sandboxes must be operational by 2 August 2027, a one-year extension. A separate EU-level sandbox operated by the AI Office is added.
• Small mid-cap enterprises (under 750 employees, turnover up to €150M) gain access to SME-style relief on documentation, quality management, and proportionate penalty caps.

What the deal does not change

• Article 4 (AI literacy): the obligation has applied since 2 February 2025; enforcement by national market surveillance authorities begins 2 August 2026. Official communications from the institutions do not indicate any amendment to Article 4 obligations.
• Most Article 50 transparency obligations continue to apply from 2 August 2026.
• AI Office enforcement powers activate as scheduled, with strengthened central competence over AI systems built on a single provider's GPAI model and over AI systems integrated into very large online platforms or search engines under the DSA.
• Article 99 penalties for the prohibitions and GPAI obligations remain live.

The TwinLadder reading

The asymmetry of the deal validates the priority sequence in the seven-pillar framework. The two pillars most directly engaged by the Act — Deployment Competence (Pillar 1, driven by Article 4) and Authority Delegation (Pillar 7, driven by Article 14 and the wider Articles 9–15) — moved in opposite directions. Authority Delegation gained sixteen to twenty-four months of breathing room. Deployment Competence gained none.

This is consistent with what we have argued across enterprise AI governance work. The workforce competence baseline is the binding constraint. Authority cannot be delegated to humans who do not understand the systems they are supervising. Postponing the human-oversight design problem does not postpone the literacy prerequisite that any oversight design depends on.

The legislator just spent six months negotiating a postponement of the work that requires redrawing authority lines, redefining scope of responsibility, and building governance for human–AI decision boundaries — Article 14 work. They left untouched the prerequisite: the obligation to ensure that the people inside those authority lines actually understand the systems they are governing.

This inverts how most enterprises have approached the Act. Article 4 has been treated as a tickbox training exercise. Articles 9–15 have been treated as the "real" compliance work. The 7 May deal flips that for the next sixteen months: Article 4 is the binding constraint with an enforceable August deadline, while Article 14 has become a deferred design problem.

The trap that becomes visible in late 2027

Companies treating the 7 May deal as sixteen months of relief are reading the wrong pillar.

The authority-delegation work that Article 14 will eventually require — deciding which decisions a human keeps, which an AI takes, which require co-signature, which require post-hoc review — only functions if the humans in the loop have the literacy to exercise that authority. Deferring the literacy work in parallel means arriving at December 2027 with both layers unbuilt. The technical documentation, post-market monitoring, and human-oversight regimes that Articles 9–15 will eventually require cannot be retrofitted onto a workforce that has not been brought to literacy first.

Practical implications before 2 August 2026

Organisations that paused Article 4 evidence collection in expectation of an Omnibus delay should restart this week. The literacy obligation is the only part of the AI Act with an August 2 enforcement date that the deal did not move. National market surveillance authorities will begin supervision on that date. Documentation of role-mapped competence, training delivery, and ongoing literacy maintenance — the foundations of any later high-risk governance regime — should be in place.

For organisations preparing for Articles 9–15 obligations under the new December 2027 timeline, the additional sixteen months are most usefully treated as design runway for authority-delegation architecture, not as deferral of the underlying work. The post-market monitoring, technical documentation, and human-oversight regimes that those Articles will eventually require cannot be retrofitted onto a workforce that has not been brought to literacy first.

Caveat

The 7 May agreement is political. Formal adoption and Official Journal publication are still required before 2 August 2026 for the postponements to take legal effect. Until publication, the original AI Act timelines remain the legally binding reference point.

Where to start

For organisations sizing the gap before 2 August 2026:

• Free company Quick Scan (organisational AI readiness across the seven pillars): twinladder.ai/en/assess/quick-scan — no registration, ten minutes, baseline reading.
• AI Power Test (individual scenario-based competence assessment, the measurement instrument behind Article 4 evidence): twinladder.ai/en/assess/ai-judgment — no registration, five to eight minutes.
• Article 4 hub: twinladder.ai/en/article-4 — what the obligation requires, who supervises, and how to evidence it.
• The Authority Gap white paper — the diagnostic argument behind why workforce competence becomes load-bearing the moment AI starts working well: twinladder.ai/en/research/the-authority-gap.