TL;DR
- The Act does not name non-human identity. Articles 12, 14, 15, 19, 26 and 72 impose duties that cannot be evidenced without it.
- Article 12 (record-keeping) and Article 14 (human oversight) are where identity evidence lands hardest. Both require attributing actions and demonstrating that oversight was exercised, not merely designed.
- Dates after the Digital Omnibus: Annex III high-risk to 2 December 2027, Annex I to 2 August 2028, Article 50(2) marking to 2 December 2026. Article 5 prohibitions and GPAI obligations did not move.
- A deferral moved the deadline, not the work. The evidence these articles require takes quarters to build.
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.
Current dates
| Obligation | Applies from | Changed? |
|---|---|---|
| Article 5 prohibited practices | 2 February 2025 | In force. Unchanged. |
| Chapter V general-purpose AI model obligations | 2 August 2025 | In force. Unchanged. |
| Article 50(2) marking, systems already on market | 2 December 2026 | Deferred from 2 Aug 2026 |
| New Article 5 prohibition (NCII and CSAM generation) | 2 December 2026 | Added by the Omnibus |
| Chapter III high-risk, Annex III stand-alone systems | 2 December 2027 | Deferred from 2 Aug 2026 |
| High-risk, Annex I systems in regulated products | 2 August 2028 | Deferred from 2 Aug 2027 |
The Digital Omnibus on AI was endorsed by the European Parliament on 16 June 2026 and approved by the Council on 29 June 2026. Verified 5 August 2026. This is a summary for planning, not legal advice; check the consolidated text for your specific obligations.
Article by article
Our practical reading. The Act does not name non-human identity; this describes where identity evidence is required to satisfy an obligation.
| Article | Obligation in substance | Identity evidence required |
|---|---|---|
| Art. 12 Record-keeping | High-risk systems automatically record events over their lifetime, enabling traceability | The core one. Logs that attribute an action to a specific system and, where the system acts for a person, to that person and their consent. A log showing only a service principal does not provide traceability. See delegation chains. |
| Art. 14 Human oversight | High-risk systems are designed so natural persons can oversee them, including intervening or interrupting | Evidence that oversight was exercised: approval records showing what was presented, refusals, and a working interruption mechanism with a measured effect time. See human-in-the-loop patterns. |
| Art. 15 Accuracy, robustness, cybersecurity | Resilience against attempts to alter use or behaviour by exploiting vulnerabilities | Credential scope and lifetime; controls against authority escalation including prompt injection; tested revocation |
| Art. 19 Automatically generated logs | Providers retain logs under their control for an appropriate period | Retention that covers the required period, with the identity attributes intact. Retention shorter than the period is the same failure as in SOC 2. |
| Art. 26 Deployer obligations | Deployers use systems per instructions, assign oversight to competent persons, and keep logs | Named accountable persons per system; the agent inventory that maps systems to those persons; deployer-side logs |
| Art. 72 Post-market monitoring | Providers collect and analyse performance data across the lifetime | Monitoring that distinguishes authorised from unauthorised behaviour, which requires knowing what authority was granted |
What evidence actually looks like
Working backwards from the articles, five artefacts carry most of the weight:
- An inventory of in-scope AI systems and the agents within them, each with a named accountable person. Satisfies Art. 26 and underpins everything else.
- Logs that reconstruct a delegation chain: which system acted, for whom, under what consent. Art. 12 and Art. 72.
- Oversight records including refusals. Art. 14. An approval queue with a 100% approval rate is evidence the control is not operating.
- A measured interruption time. Art. 14 requires the ability to interrupt; a number demonstrates it exists.
- Retention configuration evidence showing logs survive the required period with identity attributes intact. Art. 19.
None of these is producible retrospectively. Logs not written cannot be recovered, and oversight not exercised cannot be evidenced later.
GPAI obligations are already in force
It is easy to read the deferral as meaning nothing applies yet. Chapter V obligations on general-purpose AI models have applied since 2 August 2025, and Article 5 prohibitions since 2 February 2025. Neither moved.
For organisations that build on general-purpose models rather than providing them, the practical consequence is in the documentation and information-flow duties along the value chain: what you receive from your model provider, and what you must pass on. Identity work touches this where you must show which systems consume which models under which terms. Again an inventory question.
Reading the deferral correctly
Runway, not relief
The co-legislators were explicit that preparation should already be under way, and the obligations are unchanged in substance. The evidence they require, inventories, delegation logs, oversight records with refusals, tested interruption, takes quarters rather than weeks to build, and none of it can be produced retrospectively.
The framing that survives a board conversation: we gained runway, and the work that fills the runway is the same work. An organisation that builds the management system now arrives at 2 December 2027 with evidence. One that waits arrives with intentions. See ISO/IEC 42001 for the management system that produces most of this evidence, and the compliance officer guide for how to present it internally.
Frequently asked questions
Which EU AI Act articles require non-human identity controls?
The Act does not name non-human identity, but at least six articles impose obligations that cannot be evidenced without it. Article 12 on record-keeping and Article 14 on human oversight are the densest. Article 15 covers resilience against attempts to alter system behaviour, Article 19 covers log retention, Article 26 covers deployer obligations including assigning oversight to competent persons, and Article 72 covers post-market monitoring.
What are the current EU AI Act deadlines after the Digital Omnibus?
Annex III stand-alone high-risk obligations apply from 2 December 2027, deferred from 2 August 2026. Annex I high-risk systems in regulated products apply from 2 August 2028, deferred from 2 August 2027. Article 50(2) marking for generative systems already on the market applies from 2 December 2026. Article 5 prohibitions have applied since 2 February 2025 and general-purpose AI model obligations since 2 August 2025; neither moved.
What evidence satisfies the Article 12 record-keeping obligation for AI agents?
Logs that attribute an action to a specific system and, where that system acts on behalf of a person, to that person and the consent under which they acted. A log showing only a service principal identifies which component acted but not under whose authority, which does not provide the traceability the article requires. This is the delegation-chain problem, and it cannot be reconstructed retrospectively.
Does Article 14 human oversight require approval on every action?
No. It requires that systems are designed so natural persons can effectively oversee them, including the ability to intervene or interrupt. Gating every action produces approval fatigue and weakens the control. What satisfies the article is oversight that demonstrably operates: approval records showing what was presented to the approver, refusals as evidence the control is real, and a working interruption mechanism with a measured effect time.
Does the deferral mean we can pause EU AI Act preparation?
The co-legislators were explicit that preparation should already be under way, and the obligations are unchanged in substance. The evidence they require. Inventories, delegation logs, oversight records including refusals, and tested interruption. Takes quarters to build and none of it is producible retrospectively. The accurate framing is that you gained runway, and the work that fills the runway is the same work.
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.
Mapping NHI controls to a framework you are held to?
HumanAudit runs framework-mapping and readiness work where non-human and agent identity is in scope, and tests what you could actually produce on request rather than what the policy says.