Providers of standalone Annex III high-risk systems

The situation

The intended purpose places the system in an Annex III area: recruitment screening under point 4, admission to education under point 3, creditworthiness evaluation under point 5(b), or the triage of emergency calls under point 5(d). Art. 6(2) classifies such a system as high-risk unless one of the Art. 6(3) conditions applies, and the provider then carries the whole of Chapter III Section 2: a risk management system under Art. 9, data governance under Art. 10, technical documentation under Art. 11 in the Annex IV structure, automatic logging under Art. 12, information to deployers under Art. 13, human oversight design under Art. 14, accuracy, robustness and cybersecurity under Art. 15, and a quality management system under Art. 17.

Most of these requirements state no numeric level. Art. 9(8) leaves the test metrics and probabilistic thresholds to the provider, Art. 15(1) asks for an appropriate level, and Art. 10(3) names representativeness and completeness without a figure. No harmonised standard is cited in the Official Journal of the European Union as of the anchor date, so no presumption of conformity is available and the provider sets each threshold from draft standards and the state of the art, justifies it against the context of use, and records the justification. For Annex III points 2 to 8, Art. 43(2) sets the conformity assessment procedure at internal control under Annex VI, so the same provider-assembled file is what a market surveillance authority reads under Art. 74, and what an assessor checks is the justification and the evidence behind it.

A retrain or a version change raises Art. 43(4): a substantial modification within Art. 3(23) requires a new conformity assessment, and only a change pre-determined at the initial assessment and recorded in the Annex IV section 2(f) description escapes that rule. The regression testing service covers the re-test that follows a change. Regulation (EU) 2026/1744 moved the application date of the whole set and changed none of its content, so the documentation, test and quality-system backlog is the same size as before the amendment. The questions this hub answers:

• Legal has forwarded the Act and expects engineering to say what is tested and what the evidence looks like
• No harmonised standard is cited in the Official Journal yet, only draft standards, so the acceptance thresholds have to be chosen and defended by the provider
• The Annex VI self-assessment route needs an internal file that would survive a market surveillance request
• A model retrain or version change may count as a substantial modification and reopen the assessment
• The Omnibus moved the application date, and the backlog of documentation, tests and QMS work did not shrink

Obligations and dates

The dates below render from the regulation data, and each obligation row links to the page that splits the article into testable, documentation and process elements.

Where the Act sets no number, the threshold comes from a six-level hierarchy: a harmonised standard cited in the Official Journal, of which none exists as of the anchor date, as the standards page records; a JTC 21 draft named with its stage, such as prEN 18229-2Draft for the structure of the accuracy and robustness declaration and prEN 18228Enquiry for the Art. 9(2) risk management cycle; an ISO/IEC standard or technical report; a Commission guideline or code of practice; NIST AI RMF or a sector framework; and the provider's justified state of the art. Each obligation page names the level in use for that article and the protocol step that records the provider's justification.

Two points in this obligation set are unsettled. The first concerns Art. 6(3): what the streamlined registration under Art. 49(2) still requires after a not-high-risk self-assessment, and the working reading is the one the Art. 9 obligation page and the mock assessment follow.

The second concerns Art. 10(5): the necessity threshold for processing special categories of personal data in a bias examination, which decides what legal-basis record an assessor asks for alongside the bias results.

What an assessor or authority will ask for

The list starts with the Annex IV technical file and continues with the evidence its sections cross-reference; the last item is the matrix that ties each requirement to a test and an artifact.

1. Annex IV technical file with every section populated and cross-referenced to test results
2. Risk management file under Article 9 with the residual-risk judgement recorded
3. Dataset documentation and bias examination results under Article 10
4. Accuracy, robustness and cybersecurity test reports with declared metrics, the metric definitions and the rationale for each threshold
5. Logging design and a decision reconstructed from logs under Article 12
6. Human oversight design and override test results under Article 14
7. QMS description and internal audit records under Article 17
8. Traceability matrix from each requirement to the test and the artifact that answers it

Each item on the list maps to an artifact family specified with a synthetic sample on the deliverables page: the accuracy, robustness, bias and override results are test reports with a protocol run record for each execution; the Annex IV file, the risk management file and the QMS records are assembled into an evidence pack; the requirement-to-test-to-artifact chain is a traceability matrix; the dataset documentation is a dataset profile with dataset cards; and gaps found on the way are entered in a findings register.

Scenarios

The scenario pages tagged for Annex III providers each take one situation from the list above to the article it turns on, the evidence it requires and the consequence of inaction.

• Which AI systems require a conformity assessment, and when a notified body is involved
• When do high-risk AI obligations apply: the runway after the Omnibus
• What does the EU AI Act require of human oversight: proving it is real, not symbolic
• Our legal team sent us the AI Act: what do we test
• AI Act investigations: preparing an evidence pack for the first information request
• Our model version changed: what re-testing does the AI Act require
• Documenting a not-high-risk decision under the Article 6(3) exemption
• EU AI Act and recruitment: an HR screening tool as an Annex III system

Guides

The guides tagged for Annex III providers give the commands, configurations and synthetic examples behind the tests in this obligation set.

• Annex IV of the AI Act: assembling the technical documentation
• Guidelines on classification of high-risk AI systems: applying them to one system
• EU AI Act Article 17 quality management system: building the QMS documentation
• AI bias testing for fairness under the EU AI Act: subgroup disaggregation
• Fines under the AI Act: Article 99 and the action that follows each tier
• Formal robustness verification with ISO/IEC 24029-2:2023
• Internal control vs third-party conformity assessment: Annex VI and Annex VII
• ISO/IEC 42001 and the AI Act: what a QMS audit against Article 17 checks
• AI red teaming: running a jailbreak-resistance campaign
• EU AI Act Article 12 logging: verifying log completeness
• Override testing: what AI human oversight under Article 14 looks like in a test
• Provider vs deployer under the AI Act
• AI risk assessment under Article 9: building the risk management file
• Robustness testing with ISO/IEC TR 24029-1

Services

Each service tests against one or more articles of this set and prepares the evidence named in its deliverables, within the limits stated on what we cannot do.

• Robustness and adversarial testingnow, toward the Annex III date2027-12-02Deferred
• Bias and fairness testing under the Omnibus sensitive-data provisiontoward the Annex III date2027-12-02Deferred
• Logging and human-oversight verificationtoward the Annex III date2027-12-02Deferred
• Technical documentation and evidence-pack preparationtoward the Annex III date2027-12-02Deferred
• Pre-conformity mock assessmenttoward the Annex III date2027-12-02Deferred
• Data governance and dataset quality testingtoward the Annex III date2027-12-02Deferred
• QMS gap analysistoward the Annex III date2027-12-02Deferred
• Regression testing after substantial modificationongoing2027-12-02Deferred

Protocols

Every service on this hub runs a published protocol from the methodology, each with numbered steps cited to a standard or an article and a version from data.

• PR-15-ROB v1.0Robustness and adversarial testing protocol
• PR-10-BIAS v1.0Bias and subgroup performance protocol
• PR-12-LOG v1.0Logging completeness and reconstruction protocol
• PR-14-OVR v1.0Human oversight verification protocol
• PR-11-DOC v1.0Annex IV documentation assembly and consistency protocol
• PR-43-MOCK v1.0Pre-conformity mock assessment protocol
• PR-17-QMS v1.0QMS gap analysis protocol

Tools and templates

The tools run in the browser, send nothing to a server, and publish their logic on the tool methodology pages; the document templates for this obligation set are on the templates page.

• Risk and role classifier
• Obligation tracker
• Test checklist generator
• Traceability matrix template

Questions

+−What are the obligations for high-risk systems under the AI Act?

Chapter III Section 2 of Regulation (EU) 2024/1689 sets the requirements: risk management under Art. 9, data governance under Art. 10, technical documentation under Art. 11 with Annex IV, record-keeping under Art. 12, information to deployers under Art. 13, human oversight under Art. 14, and accuracy, robustness and cybersecurity under Art. 15. Art. 16 adds the provider duties: a quality management system under Art. 17, conformity assessment under Art. 43, registration under Art. 49 and post-market monitoring under Art. 72. The table on this page counts the elements per article.

+−What is an AI conformity assessment?

Art. 3(20) of Regulation (EU) 2024/1689 defines conformity assessment as demonstrating whether the Chapter III Section 2 requirements are fulfilled. Art. 43 sets the procedure: for Annex III points 2 to 8, internal control under Annex VI without a notified body; for Annex III point 1 biometric systems, Annex VI or Annex VII where harmonised standards are applied in full, otherwise Annex VII with a notified body. PR-43-MOCK rehearses the Annex VI file review; the assessment itself stays with the provider and, where required, a notified body.

+−How does the EU AI Act affect AI testing requirements?

Art. 9(6) to 9(8) of Regulation (EU) 2024/1689 require testing before placing on the market, against prior defined metrics and probabilistic thresholds appropriate to the intended purpose. Art. 15(1) requires an appropriate level of accuracy, robustness and cybersecurity, Art. 10(2)(f) an examination for bias, and Art. 14(4) oversight measures a person can exercise. The Act sets no numeric threshold for any of these; the provider sets, justifies and documents each one. Art. 13(3)(b)(ii) requires the declared level in the instructions for use, and a test verifies that declared level.

+−Digital Omnibus on AI: what changed?

Regulation (EU) 2026/1744 amended Art. 113(c) of Regulation (EU) 2024/1689 so that Chapter III Sections 1, 2 and 3, except Art. 6(5), apply to systems classified as high-risk under Art. 6(2) and Annex III from the date on the deadlines list of this page, with a separate date for Annex I products. The content of Art. 9 to Art. 17 is unchanged. Art. 50 transparency, the general-purpose AI model obligations and the Art. 5 prohibitions did not move. Every date on this site renders from the regulation data.

Every inquiry receives a reply within 24 hours. Scope and availability are discussed per request.

Ask about scopehello@aiassurancelabs.com