Guide for Auditors of AI Systems Under EU Rules
Verification required. EU claims on this page were compiled on 7 October 2026 from secondary sources and from memory of the legal text. The Official Journal text was not read, and each claim is tagged [Reported], [Recalled], [Conflict] or [Unverified] as explained on the section home page. Check the primary text on EUR-Lex before relying on any of them. This is not legal advice.
For internal audit, external assurance and readiness reviewers. EU claims carry the tags used in the EU AI Regulatory Hub. A readiness audit is not a conformity assessment by a notified body.
1. Set the audit criteria honestly
| Criteria | Examples | Caution |
|---|---|---|
| Binding law | AI Act, GDPR, NIS2, DORA, CRA | Read the Official Journal text; it was not read for this section |
| Guidance | Commission guidelines, EDPB opinions | Some are drafts only [Reported] |
| Standards | EN 18286:2026 (OJ citation pending) and drafts [Reported]; ISO/IEC 42001 | A standard gives a presumption of conformity only after OJ citation |
| Internal policy | Client AI policy | Always in scope |
State in the engagement letter which criteria are law and which are reference. Because the high-risk dates have moved [Reported], also state the assessment date logic: "tested against requirements as if applicable" or "tested against requirements currently applicable".
2. Scoping
- Obtain the inventory (T09). If none exists, report that.
- Check the classification and role decisions (T01) for each system. Retest a sample independently. The common failure is classifying a system into a lower tier than it belongs in.
- Identify the jurisdictions and competent authorities (T11).
- Pick the sample: all high-risk systems and all systems close to a prohibited practice, plus a risk-weighted selection of the others.
- Confirm the evidence period.
3. Control families and tests
| Family | Wiki controls | Test | Evidence |
|---|---|---|---|
| Governance | AI-CTRL-011, AI-CTRL-012, AI-CTRL-037 | Policy, roles, board reporting, AI literacy | Policy, minutes, T07 |
| Classification | AI-CTRL-011 | Each system has a role and a tier with reasoning; Art. 6(3) cases documented | T01 |
| Prohibited practices | AI-CTRL-012 | Screening against Art. 5 is documented | T01 |
| Risk management | AI-CTRL-037 | Lifecycle risk process; residual risk accepted by a named owner | Risk records |
| Data governance | AI-CTRL-013, AI-CTRL-041 | Provenance, bias examination, special category necessity | T10 |
| Technical documentation | AI-CTRL-038 | Annex IV content complete and current | T10 |
| Logging | AI-CTRL-008, AI-CTRL-038 | Logs reconstruct a decision; retention set | Log samples |
| Human oversight | AI-CTRL-023, AI-CTRL-025 | Oversight design; reviewers trained; overrides evidenced; kill switch tested | Review logs |
| Accuracy and cyber | AI-CTRL-005, AI-CTRL-034, AI-CTRL-039 | Declared metrics tested; adversarial testing | Test reports |
| Transparency | AI-CTRL-039 | Disclosures, marking and notices work (T06) | Screenshots, tests |
| Privacy | AI-CTRL-013, AI-CTRL-007, AI-CTRL-017 | DPIA, lawful basis, rights, transfers (T02, T04) | Records |
| Supplier | AI-CTRL-014 | Due diligence, contract, sub-processors (T03) | Records |
| Incident | AI-CTRL-035 | Playbook with AI Act, GDPR, NIS2 and DORA paths; exercise (T05) | Records |
| Post-market | AI-CTRL-035 | Monitoring plan and results | Reports |
| Deployer duties | AI-CTRL-023, AI-CTRL-038 | Use per instructions; log retention; worker information | T02, T08 |
Use T08 for line-by-line testing and the EU Evidence Register to track requests.
4. Sampling
These are suggested defaults and have no regulatory status.
- All high-risk systems.
- Of the lower tiers, at least 25 percent or five systems, whichever is greater.
- Ten decisions per high-risk system, to trace explanation and oversight.
- All incidents in the period.
5. Techniques
- Re-performance of bias and accuracy tests.
- Configuration inspection: hosting region, retention, model version.
- Walkthrough of one decision from end to end.
- Interviews.
- Document inspection, for example that approval predates go-live.
- Log reconstruction test: pick a decision and rebuild its inputs and model version from the logs.
6. Common findings
- High-risk systems classified as minimal without documented reasoning.
- An organisation treated as a deployer only, although it fine-tunes the model.
- Technical file started after launch.
- Logs missing the model version or the input reference.
- Oversight reviewers who approve without time to review.
- No FRIA where the deployer category needs one [Recalled; Art. 27 scope to verify].
- AI literacy limited to a slide deck with no role-based content.
- Transfers to non-EU AI APIs that were never assessed.
- Sub-processors unknown.
7. Rating and wording
Rate each test Effective, Partially effective, Ineffective or Not tested. Write each finding as criteria, condition, cause, effect and recommendation.
Limitation sentence: "Criteria were taken from the client's regulatory register dated [date], compiled from secondary sources and not verified against the Official Journal. This report is not a legal opinion or a conformity assessment."
8. Report checklist
- Scope, period, criteria and their verification status
- Systems and sample
- Limitations
- Findings with references
- Management responses with dates
- Follow-up
9. Handling
Keep evidence that contains personal data under production controls. Do not paste it into external AI tools.