EU AI Act (high-risk)#
The EU AI Act's high-risk provisions take effect 2 August 2026. Penalties up to €35M or 7% of global turnover. If your AI system falls under Annex III, you need the eu_ai_act_high_risk compliance profile.
Is your use case "high risk"?#
Annex III — High-risk AI use cases:
- Biometric identification + categorization
- Critical infrastructure management
- Education + vocational training (admissions, scoring)
- Employment + HR (CV sifting, promotion, performance)
- Essential public + private services:
- 5.a — Essential public services / benefits
- 5.b — Creditworthiness evaluation ← BFSI credit scoring
- 5.c — Emergency response dispatching
- 5.d — Health + life insurance risk assessment
- Law enforcement
- Migration, asylum, border control
- Administration of justice
- Democratic processes
If your model does any of the above for the EU market, you are a high-risk AI provider under the Act.
swarm's eu_ai_act_high_risk profile maps to 8 Articles#
Article 9 — Risk management system#
Must be established, implemented, documented, maintained.
| swarm feature | Evidence |
|---|---|
risk_assessment artefact (auto-generated) |
Pre-deployment risk register; per-pipeline |
| Continuous drift + fairness monitoring | Nightly cron-scheduled checks |
| Updating risk register on retrain | Each new pipeline version appends to the risk history |
Article 10 — Data and data governance#
Training, validation, test datasets must meet quality criteria; data governance practices documented.
| swarm feature | Evidence |
|---|---|
data_profiler + data_validator agents run every pipeline |
Profile + validity checks documented |
reports/data_profile.json mandatory artefact |
Statistical summary, missing-ness, outlier counts |
Data lineage in run_events |
Source + transformation + loader all tracked |
feature_engineer records transformations |
Reversible + reviewable |
Article 11 — Technical documentation (Annex IV)#
A specific technical file must exist — required before market deployment under the Act.
| swarm feature | Evidence |
|---|---|
generate_annex_iv tool (shipped in eu_ai_act_high_risk profile) |
Auto-populates the required Annex IV template |
annex_iv.pdf artefact |
Dedicated output file; follows Commission template |
| Sections covered | 1. General description; 2. Detailed description; 3. Monitoring + test; 4. Risk management; 5. Post-market monitoring plan |
Article 12 — Record-keeping (logs)#
Automatic logs of high-risk AI use. Retention: at least as long as appropriate for intended purpose; min 6 months.
| swarm feature | Evidence |
|---|---|
run_events journal |
Append-only; default 10-year retention for EU AI Act profile |
permission_denials table |
Every denied action preserved |
| Conversation journals | Agent reasoning preserved |
| Retention enforcement | SWARM_RETENTION_RUN_EVENTS_DAYS=3650 (10 years) for EU profile |
Article 13 — Transparency + provision of information to deployers#
Model card + user instructions in plain language.
| swarm feature | Evidence |
|---|---|
reports/model_card.md |
Human-readable card for deployer |
deployer_instructions.md |
Auto-generated deployment guide including input expectations, performance bounds, intended use |
| System limitations stated | Mandatory "Limitations" section in every model card |
Article 14 — Human oversight#
Design for human oversight; operator intervention must be possible.
| swarm feature | Evidence |
|---|---|
require_approval HITL gates |
6 gate types; configurable per pipeline |
| Permission engine ASK tier | Any tool can be gated by YAML policy |
| Dashboard approval queue | Operator visibility + action |
| "Override" logging | Every override + rationale recorded |
Article 15 — Accuracy, robustness, cybersecurity#
Technical characteristics documented; resilience + security assessed.
| swarm feature | Evidence |
|---|---|
model_evaluator + cross_validate |
Reported accuracy, confidence intervals |
smoke_test agent |
Pre-deployment adversarial + edge-case testing |
detect_drift cron |
Ongoing robustness monitoring |
| Cybersecurity: SOC 2 controls, BYOK, audit logs | Covered by platform security posture |
Article 72 — Post-market monitoring plan#
Monitoring plan documented; operated throughout lifecycle.
| swarm feature | Evidence |
|---|---|
post_market_monitoring_plan.md mandatory artefact |
Template pre-filled with drift thresholds, incident process |
| Cron-scheduled monitoring | Drift + fairness + performance checks |
Incident reporting via POST /api/v1/audit/incidents |
Structured format; forwardable to notified body |
Conformity assessment#
swarm produces the evidence bundle for the conformity assessment — it does not perform the assessment. A notified body or your internal quality management system does that.
For Annex III high-risk systems:#
- Third-party assessment required for biometric ID systems (Art. 43.1)
- Internal quality management system sufficient for most Annex III use cases (Art. 43.2), provided the quality management system meets Article 17 + harmonised standards
Budget + timeline#
- Notified body audit: €20K–€50K (varies by complexity + country)
- Lead time to book: 2-3 months in 2026 (supply constrained)
- Re-assessment: every 3 years or on substantial modification
CE marking#
Once conformity assessed + declaration-of-conformity signed:
- CE mark applied to the AI system (via deployment metadata)
- Technical documentation retained 10 years post-last unit placed on market
- Notified body informed of substantial modifications
swarm supports these via the audit-PDF rollup mechanism. CE metadata is a field in the model card.
Post-market surveillance#
Article 72 requires a plan; deployment. swarm's default covers:
- Drift detection (Article 72.3 — trend monitoring)
- Incident log (Article 73 — serious incident notification within 15 days)
- Feedback loop from deployers to provider (via
POST /api/v1/deployer-feedback) - Corrective action tracking
GPAI (General Purpose AI) note#
If you use an LLM in your pipeline (Anthropic, OpenAI, etc.), the GPAI provider obligations apply to them — not you as a deployer. Your relationship is as deployer of the provider's GPAI model.
Your obligations (Article 25): make sure the GPAI provider has complied; maintain documentation of your use. swarm's audit trail suffices.
Serious incident reporting (Article 73)#
If a deployed model causes: - Death / serious harm to person - Serious damage to property / environment - Breach of EU law (fundamental rights)
You (the provider) must notify market surveillance authority within 15 days (immediately for fatal). Template for the report:
swarm audit incident-report \
--model <name> \
--version <v> \
--incident-date <YYYY-MM-DD> \
--harm-description "..." \
--corrective-action "..." \
--output incident_<date>.pdf
Output is formatted per Commission template.
Data residency + CLOUD Act#
See Deployment: Data residency for the EU-region + non-US-HQ considerations. The EU AI Act itself doesn't mandate EU-region data; GDPR does for personal data. But the combination often demands EU-region hosting.
Limitations / honest disclosure#
- The Act's secondary legislation is still landing. Commission implementing acts + harmonised standards are being published through 2026. Some details below may firm up as standards land.
- swarm is not the notified body. We are the evidence-production platform.
- GPAI model providers (Anthropic, OpenAI, Microsoft) shoulder their own obligations. If they comply, your deployer obligations (Article 25) are lighter. Verify their status.
Next#
- Reading the audit PDF — EU AI Act flavour of the audit
- Concepts: Compliance profiles
- Deployment: Data residency — EU-region + non-US-HQ considerations