Manufacturers embedding AI in Annex I products

A manufacturer whose AI system is a safety component of, or is itself, a product listed in Annex I of Regulation (EU) 2024/1689 is the provider of a high-risk AI system under Art. 6(1) and Art. 25(3). Art. 9 to Art. 15 and Art. 17 apply from the Annex I application date 2028-08-02Deferred, assessed inside the sectoral conformity route under Art. 43(3). Regulation (EU) 2026/1744 moved that application date and left the obligations unchanged in content. This hub gives the obligation set with dates, the evidence an assessor asks for, scenarios, guides, services, protocols and tools.

The situation

The manufacturer already keeps a technical file, a quality system and a notified body relationship under the sectoral law that lists the product in Annex I. Regulation (EU) 2024/1689 builds on that position: Art. 8(2) allows the provider to integrate the testing, reporting and documentation the Act requires into the documentation and procedures that already exist under the sectoral law; Art. 11(2) requires one single set of technical documentation that holds the Art. 11(1) Annex IV content and the information the sectoral law requires; Art. 17(3) allows the aspects of the quality management system in Art. 17(1) to sit inside the quality system the sectoral law already imposes. The practical task is to add AI-specific evidence to the existing file, mark what the two sets of requirements share, and show the notified body the mapping.

The assessor is a product-safety assessor. Under Art. 43(3), an Annex I Section A system follows the conformity assessment procedure of the sectoral law, with the Chapter III Section 2 requirements and Annex VII points 4.3 to 4.5 applied within that procedure, and the sectoral notified body checks the AI requirements when it meets the Art. 31 conditions that Art. 43(3) names. Test results the machine-learning team produces therefore have to read as product-safety evidence: the metric in the product's own performance terms, the test set described, the threshold justified, and the run reproducible. For machinery the regulation data records the Omnibus change 2026-07-27In force: the machinery directive left Annex I Section A and the Machinery Regulation entered Section B, so an AI safety component in machinery follows the sectoral route, with the AI-specific requirements arriving through delegated acts under that regulation. For a Section B product, Art. 2(2) limits the direct application of the Act to Art. 6(1), Art. 102 to Art. 109 and Art. 112.

A retrained model inside a product that has passed its sectoral assessment raises the question in Art. 3(23) and Art. 43(4): a change that affects the conformity of the system with the Chapter III Section 2 requirements, or that changes the intended purpose, is a substantial modification and reopens the assessment. For a system that continues to learn after it is placed on the market, Art. 43(4) exempts a change that the provider pre-determined at the initial assessment and recorded in the Annex IV technical documentation. A change-control record that states the pre-determined change envelope and the regression result for each update, under the regression testing service, is the evidence that decides whether an update is a substantial modification.

The questions this hub answers:

The existing notified body will ask for AI Act evidence inside the sectoral assessment, and the technical file has no section for it yet
Overlap between the sectoral requirements and Articles 9 to 15 is unclear, and double documentation costs time in the assessment
Machinery, medical device and lift teams each read the Act differently, and the quality system has to hold one approach
Test evidence produced by the ML team is not in a form a product-safety assessor recognises
A retrained model inside a product that has passed its sectoral conformity assessment may reopen that assessment

Obligations and dates

Two regulation entries carry the dates for this obligation set: the Annex I high-risk entry for Section A products and the machinery entry that records the Omnibus change for Section B.

2028-08-02DeferredAnnex I product-embedded high-risk obligations
2026-07-27In forceMachinery Regulation (EU) 2023/1230 moved from Annex I Section A to Section B

Annex I Section A products (medical devices, lifts, and the other listed product laws) follow the annex-i-high-risk entry; the AI Act requirements enter the sectoral conformity assessment carried out by the product's notified body. Machinery moved to Annex I Section B under the Omnibus, where only Art. 6(1), Art. 60a and Art. 102 to 112 apply and the AI requirements reach the product through the Machinery Regulation; the machinery-regulation-annex-i-b entry records that change.

Obligations that apply to Annex I manufacturers
Obligation	Applies from	Testable	Documentation	Process	Services
Art. 9Risk management system	2028-08-02Deferred	6	11	6	Robustness and adversarial testing Pre-conformity mock assessment QMS gap analysis
Art. 10Data and data governance	2028-08-02Deferred	8	11	6	Pre-conformity mock assessment Data governance and dataset quality testing
Art. 11Technical documentation	2028-08-02Deferred	12	19	4	Technical documentation and evidence-pack preparation Pre-conformity mock assessment
Art. 12Record-keeping	2028-08-02Deferred	10	6	4	Logging and human-oversight verification Pre-conformity mock assessment
Art. 13Transparency and provision of information to deployers	2028-08-02Deferred	5	14	3	Technical documentation and evidence-pack preparation Pre-conformity mock assessment
Art. 14Human oversight	2028-08-02Deferred	7	8	4	Logging and human-oversight verification Pre-conformity mock assessment
Art. 15Accuracy, robustness and cybersecurity	2028-08-02Deferred	7	6	4	Robustness and adversarial testing Pre-conformity mock assessment Regression testing after substantial modification
Art. 17Quality management system	2028-08-02Deferred	8	17	12	Pre-conformity mock assessment QMS gap analysis

Also relevant, without an obligation page:

Art. 43
Annex I
Annex IV
Art. 6

None of the requirements in Art. 9 to Art. 15 carries a numeric threshold in the Act: Art. 15(1) requires an appropriate level of accuracy, robustness and cybersecurity, and Art. 9(8) leaves the prior defined metrics and probabilistic thresholds to the provider. No harmonised standard is cited in the Official Journal as of the anchor date, so no presumption of conformity is available and the first level of the source hierarchy is empty. Thresholds come from the next levels: a JTC 21 draft such as prEN 18229-2Draft for the accuracy and robustness metric categories and the declaration structure, an ISO/IEC standard or technical report at its stage, a Commission guideline or code of practice, a national or sector framework, and last the state of the art that the provider justifies against the context of use and documents. For an Annex I product the sectoral performance requirement sits beside those levels, so the justification states each metric in the product's own performance terms, and the quality system records structured after prEN 18286Published, a JTC 21 draft, hold that justification with the test set and the measurement conditions.

The Art. 6(3) filter removes an Annex III classification only; a system classified under Art. 6(1) with Annex I has no Art. 6(3) exit. The registration point below is shown because the obligation record for Art. 9 carries it; it applies to a manufacturer that also places an Annex III system on the market.

Where a dataset used to detect or correct bias under Art. 10(5) contains special categories of personal data, the necessity threshold is an unsettled point with two readings.

What an assessor or authority will ask for

The sectoral notified body and, under Art. 74, the market surveillance authority ask for the following items.

Mapping table from each sectoral requirement to the corresponding AI Act requirement, with the shared evidence marked
AI-specific annexes to the existing technical file that follow the Annex IV headings
Data governance and dataset documentation for the training, validation and test sets
Accuracy and robustness results expressed in the product's own performance terms, with thresholds justified
Logging and human oversight design integrated with the product's safety functions
Change-control records that decide whether a model update is a substantial modification

Each item on that list maps to an artifact family specified with a synthetic sample on the deliverables page: the mapping table and the change-control records follow the traceability matrix and the register or checklist specifications; the AI-specific annexes and the evidence index follow the evidence pack specification; accuracy, robustness, logging and oversight results follow the test report and the protocol run record specifications; dataset documentation follows the dataset profile and dataset cards specification; and the gaps a dry run finds follow the findings register specification.

Scenarios

Scenario pages tagged with this hub list here as they are published; the seeds are the overlap between the medical device rules and the Act, AI in machinery after the Omnibus Section B change, and adding AI evidence to an existing technical file.

Guides

Guide pages tagged with this hub list here as they are published, covering the test methods and documentation structures behind the evidence list.

Services

Seven services test against the obligation set within the limits on what we cannot do: AIAssuranceLabs prepares evidence for the sectoral conformity assessment and does not perform it.

Robustness and adversarial testingnow, toward 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

Each service runs a published protocol from the methodology, with numbered steps, a citation on every step and a version.

PR-15-ROB v1.0Robustness and adversarial testing 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 are on the templates page.

Questions

What is Annex I of the EU AI Act?

Annex I of Regulation (EU) 2024/1689 lists the Union harmonisation laws whose products bring an AI system into the high-risk class under Art. 6(1). Section A lists New Legislative Framework laws, among them those on medical devices and lifts. Section B lists other Union laws, among them those on aviation, vehicles and, after Regulation (EU) 2026/1744, machinery. An AI system is high-risk under Art. 6(1) when it is a safety component of, or is itself, a product covered by a listed law that requires third-party conformity assessment.

How does the AI Act apply to medical devices?

A medical device with an AI safety component, or a device that is itself an AI system, is high-risk under Art. 6(1) when the device requires third-party conformity assessment under the sectoral law listed in Annex I Section A. The device manufacturer is the provider under Art. 25(3) when the system is placed on the market under its name. Art. 43(3) places the Chapter III Section 2 requirements inside the device's own conformity assessment, and Art. 11(2) requires one technical documentation set for the Act and the sectoral law.

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

Ask about scope hello@aiassurancelabs.com