By HumanAudit Inc. editorial teamLast reviewed 5 August 2026
VerifiedLast reviewed 5 August 2026 by the HumanAudit Inc. editorial team.Corrections logEditorial policy
On this page
  1. AI Act in 60 seconds
  2. Risk-based scope
  3. High-risk obligations
  4. NHI control mapping
  5. Agentic AI specifics
  6. General-purpose AI models
  7. Sister site

Timeline changed. The August 2026 high-risk deadline was deferred

The Digital Omnibus on AI amends Regulation (EU) 2024/1689. The European Parliament endorsed the final text on 16 June 2026 and the Council gave final approval on 29 June 2026.

  • Annex III stand-alone high-risk systems: 2 December 2027 (was 2 August 2026)
  • Annex I embedded high-risk systems: 2 August 2028 (was 2 August 2027)
  • Article 50(2) marking for systems already on the market: 2 December 2026
  • Unchanged: Article 5 prohibitions (2 February 2025) and GPAI obligations (2 August 2025) are in force now

A deferral is not a reprieve. The co-legislators explicitly stated that preparation should already be under way, and the evidence obligations are unchanged in substance. Reviewed 5 August 2026.

Live now, and widely misread

The deferral covered high-risk obligations only. Four obligation sets are in force today and carry penalties:

  • Article 50 transparency applies from 2 August 2026 and did not move. If a person in the EU can talk to your AI or see content it generated, this reaches you regardless of risk tier. The maximum is €15M or 3% under Article 99(4), not the €35M figure most summaries quote.
  • Article 49 registration and national market surveillance enforcement powers also apply from 2 August 2026.
  • Article 4 AI literacy has applied since 2 February 2025 to every provider and deployer at every risk tier, including minimal risk. It is the most commonly undocumented live obligation.
  • Article 5 prohibitions since 2 February 2025 and GPAI obligations, Articles 51 to 56, since 2 August 2025.

For the article-by-article position on the Regulation itself, our sister publication EU AI Act Checklist tracks the consolidated text with a review date on every page.

Not legal advice

This page is a practitioner-oriented cross-mapping. It is not legal advice. The AI Act interacts with the GDPR, product-safety law, sectoral regulations, and national implementing measures. Qualified counsel and your regulatory affairs team are the right advisors on AI Act compliance.

AI Act in 60 seconds

The EU AI Act is the first broad horizontal regulation of AI systems. Key milestones:

  • Entered into force 1 August 2024.
  • Prohibitions (Chapter II) and AI literacy (Article 4) applicable from 2 February 2025.
  • General-purpose AI model obligations (Chapter V) applicable from 2 August 2025.
  • Main obligations for high-risk systems (Chapter III), Annex III stand-alone systems: originally 2 August 2026, deferred to 2 December 2027 by the Digital Omnibus on AI.
  • High-risk obligations for Annex I systems (safety components of regulated products): originally 2 August 2027, deferred to 2 August 2028.
  • Article 50(2) transparency and marking obligations for generative systems already on the market at 2 August 2026: 2 December 2026.
  • A new Article 5 prohibition covering AI systems for creating non-consensual intimate imagery and CSAM: 2 December 2026.

The regulation uses a risk-based approach: prohibited practices, high-risk systems, limited-risk systems with transparency duties, and minimal-risk systems (most AI). The heaviest obligations attach to high-risk systems.

Risk-based scope

  • Prohibited practices (Article 5): social scoring by public authorities, real-time biometric identification in public spaces (with exceptions), and several others. These are out.
  • High-risk systems (Annex III and Annex I): AI systems used in biometrics, critical infrastructure, education, employment, access to essential services, law enforcement, migration, justice, and democratic processes (Annex III); and AI systems that are safety components of regulated products (Annex I).
  • Limited-risk (transparency obligations, Article 50): chatbots, deepfakes, emotion recognition, etc. Disclosure requirements.
  • General-purpose AI models (Chapter V): separate obligations for foundation-model providers.

Key high-risk obligations (Chapter III, Section 2)

Providers of high-risk AI systems must implement (among others):

  • Risk management system (Article 9)
  • Data governance (Article 10), training, validation, testing data quality
  • Technical documentation (Article 11)
  • Record-keeping / logging (Article 12), automatically generated logs, retained for traceability
  • Transparency and information to deployers (Article 13)
  • Human oversight (Article 14)
  • Accuracy, robustness, and cybersecurity (Article 15)

Mapping to NHI controls

Several of the Article 9 to 15 obligations touch on NHI governance directly, especially for AI systems that interact with other systems autonomously (i.e. agentic AI):

AI Act articleNHI control interpretation
Art. 9, Risk managementRisk register covers NHI-specific risks for AI systems: agent credential compromise, prompt-injection-driven authority escalation, delegation-chain failure, over-privileged agent identities.
Art. 10, Data governanceNHI credentials used to access training data and pipelines are themselves governance-relevant. Inventory credentials that can access sensitive data flows.
Art. 11, Technical documentationDocument the identity architecture of the AI system: how the agent authenticates to downstream systems, what permissions it holds, how credentials are managed.
Art. 12, LoggingNHI audit logs (which agent authenticated, to which system, for what action) feed the automatically generated logs required for traceability.
Art. 14, Human oversightFor agentic systems, human-oversight design often includes approval gates on privileged actions, implementable via just-in-time elevation of the agent's NHI privileges.
Art. 15, CybersecurityThe broadest hook: NHI hygiene (rotation, least privilege, leak prevention) is a cybersecurity measure for the AI system. Breach of an agent's NHI credential would be a cybersecurity incident under Article 15.

Agentic AI specifics

Agentic AI, language models acting with tools, memory, and goals, is where the AI Act meets NHI governance most sharply. Key implications:

  • Agent identity. Each agent instance needs its own identifiable NHI, auditable to the agent's actions.
  • Delegation chains. Where an agent acts on behalf of a user, the chain should be traceable, user → agent credential → downstream system credential. Article 12 logging expectations align with this.
  • Human-in-the-loop gates. Privileged actions should require human approval. Implementation: the agent's NHI holds narrow default scope, with JIT elevation requiring human authorisation.
  • Prompt-injection resistance. Article 15 cybersecurity obligations likely extend to designing agents so that prompt injection cannot escalate authority beyond the agent's intended scope.

Our AI agent identity hub covers agentic AI NHI governance in depth.

General-purpose AI model obligations

Providers of general-purpose AI models (Chapter V) have a separate set of obligations around technical documentation, copyright compliance, and summary of training content. For providers of models with systemic risk (Article 51), additional obligations include systemic-risk assessment and cybersecurity. NHI governance is relevant here too, a GPAI provider whose model-serving infrastructure is compromised through NHI paths has a cybersecurity incident at systemic-risk scale.

Sister site

HumanAudit Inc. Where this page gives the NHI-specific angle, the sister site covers the Act more broadly (scope determination, high-risk classification, technical documentation templates, conformity assessment). Cross-linked frequently.

Primary sources

Turning this into evidence

The obligations above resolve into artefacts: an inventory of in-scope systems with named owners, logs that reconstruct who authorised an agent action, oversight records including refusals, and a measured interruption time. Those are produced by a management system, not by a checklist.

ISO/IEC 42001 documentation toolkits for the artefacts themselves, or a 20-minute call with HumanAudit if you need the scope decided first.

Need this mapped to an audit you actually face?

Frameworks tell you what good looks like. Evidence is what an auditor asks for. HumanAudit runs ISO/IEC 42001 and NHI readiness reviews that test whether your evidence would hold, including agent delegation-chain testing.