Compliance & audit
EU AI Act — article-by-article mapping
Every article of the EU AI Act that a gateway can help with, mapped to AgentKeyRoute capability and the console screen that produces the evidence.
Timeline
| Date | What applies |
|---|---|
| 2 Feb 2025 | Prohibited practices and AI literacy obligations take effect. |
| 2 Aug 2025 | General-purpose AI model (GPAI) obligations begin. |
| 2 Aug 2026 | Main obligations for high-risk systems enforceable (Articles 6–27, logging and oversight requirements). |
| 2 Aug 2027 | High-risk systems embedded in regulated products under existing EU law. |
Compliance writers say February is the real date
If you want six months of logs on enforcement day, Article 19 means you need the gateway in production by early February 2026 — not August.
Article-by-article
| Article | What the Act requires | How AgentKeyRoute meets it |
|---|---|---|
| Article 12 — Automatic logging | High-risk AI systems must technically allow for the automatic recording of events (logs) over the lifetime of the system, sufficient to identify risks and support post-market monitoring. | Every request, redaction, policy decision, model choice and response is written to the hash-chained request log. AgentKeyRoute Guard decisions in the browser (log / inspect / proxy) write the same shape of record. |
| Article 13 — Transparency | Deployers must know what the system does and the risks of using it, in plain terms. | Each governance profile carries a plain-language "What this profile does" panel shown to admins and included in the evidence pack. |
| Article 14 — Human oversight | High-risk AI must allow natural persons to oversee it, including the ability to stop or override the system. | Workloads can be marked autonomous or assistive. Autonomous high-risk flows can be held for review; kill switches stop a workload, a key or every workload in one click. |
| Article 15 — Accuracy, robustness, cybersecurity | High-risk AI must meet accuracy, robustness and security requirements proportionate to the risk. | Served-model verification refuses answers from a model your policy did not allow. Guardrails screen for prompt injection, unsafe content and denied topics. |
| Article 17 — Quality management system | Providers must have a documented QMS covering risk, testing, change control and incident reporting. | Audit log, incidents, two-admin approval and two-step verification are the enforcement side; the profile library and framework attestations are the documentation side. |
| Article 18 — Documentation retention | Technical documentation and the QMS must be kept for 10 years after a system is placed on the market. | Not covered by the gateway directly — but every artefact the gateway produces (evidence pack, audit log export) is exportable and portable for your own long-term storage. |
| Article 19 — Log retention | Logs under Article 12 must be kept for at least six months unless other law extends that. | Retention is set per profile (90 days to seven years). The default on EU AI Act and HIPAA profiles exceeds six months. |
| Article 20 — Corrective actions | Providers must take prompt corrective actions on non-conforming systems and inform competent authorities. | Kill switches and incident workflow; audit log records both the detection and the action. |
| Article 21 — Cooperation with competent authorities | Providers must give competent authorities the logs and information needed to demonstrate conformity, on request. | The evidence pack exports framework scores, gaps, attestations, retention policy and a signed audit-chain check in one file. Audit log exports to CSV, Parquet or your SIEM. |
| Article 26 — Deployer obligations | Deployers must use the system per instructions, assign human oversight, monitor it, and keep logs if they control them. | Workloads record owner, team, purpose, data classes and EU AI Act tier. Every AI use shows up in the inventory. |
| Article 50 — Transparency obligations | Certain AI must tell people they are interacting with AI; AI-generated content must be marked as such. | Public assistants can be configured to include a disclosure. For AI-generated content marking, use upstream provenance; the gateway records that the content originated from a model and which one. |
Scope: which of your AI uses is “high risk”?
Annex III of the Act lists the high-risk categories: biometric identification, critical infrastructure, education and training, employment and HR, essential private and public services, law enforcement, migration and asylum, administration of justice and democratic processes. Register each AI use as a workload (opens in a new tab) with its tier (minimal, limited, high, prohibited); the gateway then enforces the right controls.
Evidence you can hand to a regulator
- Evidence pack under Frameworks & evidence (opens in a new tab): framework scores, gaps, attestations, retention policy and a signed audit-chain check.
- Audit log export (CSV / Parquet / SIEM): every request and configuration change, hash-chained.
- Workload inventory: every registered AI use with owner, data classes and tier.
- Profile printouts: plain-language description of the rules each app follows.