A narrow, dated note on EU instruments relevant to AI-agent oversight. The EU regime is staged: different articles of the same regulation apply from different dates, and a 2026 amendment package (now in force) has moved some of them. Each entry names the instrument's exact status at the review date.
These pages describe the product and cite public instruments. They are not legal advice, not a certification, and not a statement that any instrument applies to your organisation. The official source is authoritative; applicability questions belong to your advisers.
Instruments, by status
| Instrument | Status | What it addresses |
|---|---|---|
| AI Act, Regulation (EU) 2024/1689 | Binding law, staged application. In force 1 August 2024. Prohibitions and AI-literacy duties applied 2 February 2025; general-purpose-AI obligations 2 August 2025 (subject to the Act's transitional provisions, including for pre-existing models); general application from 2 August 2026 (including Article 50 transparency). The Digital Omnibus, Regulation (EU) 2026/1744, in force 27 July 2026, then deferred the high-risk obligations: Annex III standalone high-risk duties now apply from 2 December 2027, and Annex I product-related high-risk obligations from 2 August 2028; two new Article 5 prohibitions apply from 2 December 2026. | Providers and deployers of AI systems, including systems used in the EU by organisations established elsewhere. Human oversight (Article 14), record-keeping, and deployer duties for high-risk systems are the provisions action-layer records speak to. AAES is not a conformity-assessment tool. |
| GDPR, Article 22 | Binding law. | Decisions based solely on automated processing with legal or similarly significant effect, subject to exceptions and safeguards. GDPR roles are controller and processor; meaningful human involvement routes, safeguards and information duties sit with the controller. |
| DORA, Regulation (EU) 2022/2554 | Binding law; applies since 17 January 2025. | ICT risk management, incident reporting, and third-party oversight for financial entities. AI agents acting on financial systems sit inside the entity's ICT risk picture. The entity answers to its supervisor; designated critical ICT third-party providers can also face direct oversight. |
What AAES records may contribute
| Expectation | AAES contribution | Capability status |
|---|---|---|
| Named accountability for AI-supported work (deployer governance and internal accountability) | Every registered agent has a named accountable manager, a configured association rather than a statutory appointment; approvals are attributable to a named person; agent identities cannot act as approvers through the supported AAES approval path. | Implemented. |
| Human oversight of consequential actions (AI Act Article 14 pattern; GDPR Article 22 safeguards) | Actions the operator registers as irreversible (a registration classification, not an independent assessment of effects) require approval by an authorized person, bound to one intent and the exact payload, with expiry. An approval record does not by itself show the approver had adequate information, time or practical power to intervene. | Implemented. |
| Records supporting internal review and supervisory evidence (AI Act record-keeping pattern; DORA ICT documentation) | Sealed, hash-chained decision records with offline integrity verification by the client's own reviewer. | Implemented. An independently controlled witness can be configured separately; that is not an external assessment, and the default deployment has none. |
AAES decision records may support review of access and approval controls. They do not establish conformity with the AI Act, lawful processing under the GDPR, DORA compliance, or compliance with any instrument as a whole.
Within the configured trust model, integrity verification checks the exported records; it does not establish that every action was recorded, that an external action succeeded, or that recorded assertions are true. Approval records do not by themselves establish meaningful human oversight. Enforcement depends on control of the credential path, and actions bypassing AAES are not recorded.
Residual responsibility and explicit gaps
- The client remains the regulated entity. Whether you are a provider or deployer, and which annex applies, is your adviser's question, not this page's.
- Work outside AAES is invisible to AAES. Records cover actions routed through AAES; they say nothing about actions on bypass paths.
- No independent certification or assessment of AAES exists. No SOC 2 report, no penetration test, and AAES performs no conformity assessment of your AI system.
- Status changes. The application dates shown here have already moved once (see the AI Act entry above), and further amendment proposals are in the pipeline. Re-check the Official Journal before relying on the dates shown here.
Bring the instruments your risk team cites to a scoped evaluation.
Related pages:
This note: https://aaes.ai/library/jurisdictions/european-union.html
Hong Kong note: https://aaes.ai/library/jurisdictions/hong-kong.html
Singapore note: https://aaes.ai/library/jurisdictions/singapore.html
Trust model: https://aaes.ai/library/trust-model.html
Scope an evaluation: https://aaes.ai/contact.html?ref=library-evaluation