Übernimmt die KI alles? Unser CTO über Banken, EU AI Act und deutsche Qualität | Umair Zaffar
Umair Zaffar October 29, 2026 0 Comments

EU AI Act Article 11: Technical Documentation, Implemented in Practice

EU AI Act Article 11 requires a high-risk AI system to have technical documentation under Annex IV before it goes live, and to keep it up to date. Our answer: the documentation is a build artefact that the release pipeline assembles at every release from the registry record, model card, data sheets, test reports, risk register and change log.

Article 11 is the article most teams treat as paperwork and dispatch with a Word document that is three releases behind the system. It addresses the provider, via Art. 16(d). The documentation is where every other obligation in Chapter III Section 2 is evidenced, from risk management under Article 9 to cybersecurity under Article 15. Take Article 11 seriously and you see early what is missing in the others.

The EU AI Act has been in force since 1 August 2024. The prohibitions and the AI literacy duty have applied since February 2025, the obligations for GPAI models since August 2025, the transparency obligations under Article 50 since August 2026. The obligations for high-risk systems under Annex III apply, after the postponement by the amending Regulation (EU) 2026/1744 (Digital Omnibus), from 2 December 2027, and for Annex I products from 2 August 2028. Article 11 belongs to these high-risk obligations. The Omnibus moved deadlines; it removed no obligation.

What EU AI Act Article 11 requires

Three paragraphs. The substance sits in Annex IV, which paragraph 1 points to.

  • Art. 11(1), first subparagraph: The provider draws up the technical documentation before the system is placed on the market or put into service, and keeps it up to date. It has to demonstrate that the system meets the requirements of Chapter III Section 2 and give competent authorities and notified bodies, in a clear and comprehensive form, the information for that assessment. The minimum content is the set of elements in Annex IV.
  • Art. 11(1), second subparagraph: SMEs including start-ups, and after the Omnibus amendment also small mid-cap enterprises, may provide the Annex IV elements in a simplified manner. The Commission establishes a simplified form for small and microenterprises, and notified bodies have to accept it for the conformity assessment.
  • Art. 11(2): Where the system is part of a product covered by the harmonisation legislation in Annex I Section A, a single set of technical documentation is drawn up for both legal acts.
  • Art. 11(3): The Commission may adapt Annex IV to the state of the art by delegated acts under Article 97.
  • Annex IV, nine points: (1) general description: intended purpose, provider, version and earlier versions, interaction with other hardware and software, form of supply, hardware, interface, instructions for use; (2) elements and development process: methods including third-party pre-trained systems, design decisions with assumptions and trade-offs, architecture, data sheets, assessment of the oversight under Article 14, predetermined changes, tests with dated and signed logs, cybersecurity; (3) capabilities, limitations, accuracy including for groups of persons, unintended outcomes; (4) suitability of the performance metrics; (5) risk management under Article 9; (6) changes over the lifecycle; (7) harmonised standards applied, or the solutions chosen instead; (8) a copy of the declaration of conformity under Article 47; (9) post-market monitoring under Article 72 with its plan.
  • The connecting articles: Under Art. 18(1)(a) the provider keeps the documentation for ten years. For most Annex III systems the conformity assessment under Article 43 is the internal control of Annex VI based on exactly this documentation. Market surveillance can request it under Article 74.

The technical example

A public authority receives a few thousand documents a day: applications, supporting evidence, objections, complaints, scanned, via portal, by email. A system classifies each document by type, procedure, urgency and missing attachments and routes it to the right case and the right case worker. It decides nothing. It sorts.

The data flow: the document goes through text recognition. The business application calls the policy gateway with a use-case ID, never the model. The gateway resolves the ID in the registry, checks the data zone and replaces names, addresses and file numbers with pseudonyms; the documents sit in the gray zone. A classification model, fine-tuned on historical documents labelled by case workers, returns class and confidence; a rules engine holding the procedural logic returns the routing proposal. Below the confidence threshold the document goes to a clearing desk staffed by people. Every assignment lands in the evidence log, and every rerouting by a case worker flows back as a signal.

Where does Article 11 bite? First in the classification, and here I am honest: pure inbox sorting is, on my reading, not a high-risk system under Annex III. But the line is close. As soon as the classification sets deadlines, prioritises applications for social benefits or rejects them as incomplete, Annex III point 5(a) comes into view, and point 8 for courts. In projects this shifts quietly, because the next department wants the system for the next process. That is why we generate the Annex IV documentation for every registry record, regardless of risk class. And: if the authority develops the system itself and puts it into service for its own use, it is a provider under Art. 3(3), not only a deployer.

Second, it bites at every change. New text recognition, an additional document class, a new vendor model, a new batch of labels: the documentation has to describe what runs. A Word document from go-live describes, after the third release, a system that no longer exists.

How we implement it

At SIFAMO the technical documentation is not a document someone writes but an artefact the release pipeline produces. It belongs to the governance layer, next to the registry, the policy gateway, the evidence log and the quality gates.

Every point in Annex IV has a source

We keep a mapping table: Annex IV point, source system, artefact. The registry record supplies point 1, its history point 6. The model card supplies points 2(a) and 2(b) and point 3: method, pre-trained base, accuracy per document class and scan quality. The data sheets, one per source, supply point 2(d): provenance, legal basis, labelling procedure, proof of pseudonymisation. The architecture diagram is rendered from the deployment manifests, point 2(c). The evaluation reports from the quality gates, with a result per case, a date and three signatures, supply points 2(g), 3 and 4. Oversight configuration, gateway policy, risk register and monitoring configuration supply points 2(e) and 2(h), 5 and 9.

What has no source, a human writes: the rationale for design decisions, why the metrics fit, the list of standards under point 7, the risk narratives. These texts live versioned in the same repository as the system and belong to the business owner.

The dossier build runs at every release

The pipeline collects the artefacts, checks the mapping table and builds the dossier in the structure of Annex IV: a PDF for people, a manifest with hashes for machines. An empty cell breaks the build. Not applicable is allowed, but only with a reason and a signature. Simplified:

{
  "dossier_id": "DE-OEFF-2026-0021/r7",
  "annex_iv_schema": "2024/1689-AnnexIV",
  "registry_version": 7,
  "pinned": {
    "model": "doc-cls-ft-2026-10-r3",
    "ocr": "ocr-engine 5.2.1",
    "rules": "case-rules-2026-11-03",
    "label_set": "labels-2026Q3 sha256:4a1f...",
    "eval_set": "inbox-eval-2026Q3 sha256:b77d...",
    "gateway_policy": "gw-policy-2026-10-14"
  },
  "sections": {
    "1":  {"source": "registry", "artefacts": ["record.json", "instructions-for-use.pdf"]},
    "1f": {"status": "not_applicable", "reason": "not a product component", "signed_by": "business"},
    "2c": {"source": "deploy-manifests", "artefacts": ["architecture-r7.svg"]},
    "2d": {"source": "datasheets", "artefacts": ["ds-inbox-2019-2025.md", "ds-labels-2026Q3.md"]},
    "2g": {"source": "quality-gates", "artefacts": ["eval-r7.pdf", "bias-r7.pdf"],
           "signed_by": ["business", "risk", "it"], "signed_at": "2026-11-18"},
    "5":  {"source": "risk-register", "artefacts": ["risks-r7.md"]},
    "6":  {"source": "registry-history", "artefacts": ["changelog-r1-r7.md"]},
    "9":  {"source": "monitoring-config", "artefacts": ["pmm-plan-r7.md"]}
  },
  "bundle_hash": "sha256:e3b0...",
  "built": "2026-11-19T07:40:12Z"
}

Version pinning: one dossier describes exactly one state

The registry record pins everything that determines behaviour: model, text recognition, rules, label and evaluation set with hash, gateway policy and the hash of the dossier. The gateway reads the record on every call; a component in a version that is not pinned gets no route. So the system cannot drift away from the dossier, and a release without a complete dossier does not pass the registration gate. For hosted models whose vendor swaps versions silently, we pin the version the vendor names and run the regression tests daily. A deviation is, for us, a change under Annex IV point 6 and triggers a dossier build.

The simplified form for SMEs

For the second subparagraph of paragraph 1 the generator has a profile: same structure, same artefacts, shorter narratives, evidence as an annex. The Commission form becomes the target format as soon as it is available in a project. I tell start-ups two things. The simplification concerns the form, not the evidence. And the authority, as deployer, needs the instructions for use and the statements on accuracy and limitations for Articles 26 and 27, however small its supplier is.

What the auditor and the notified body receive

A read-only portal, not email attachments. Per registry record and version: the dossier as PDF, the manifest, the mapping table with hashes, the diff to every earlier version and the jump from the dossier into the evidence log, that is from intended behaviour to the proof for one specific document. For an Annex III system like the inbox sorter the conformity assessment is the internal control under Annex VI. For Annex I products and biometric systems the notified body under Annex VII receives the same package plus the quality management documentation under Article 17. Every dossier is stored write-once, for ten years under Article 18, with the hash signed at risk.

Where Annex IV point 6 asks only for relevant changes, we document all of them, because we do not know in advance which one will become relevant. More under AI Governance.

Typical mistakes

  • The documentation describes the model, not the system. Annex IV means the whole: text recognition, pseudonymisation, model, rules, oversight. The vendor model card covers one point of nine.
  • One document at go-live, then never again. After three releases it describes a different system. Keeping it up to date is part of the obligation in paragraph 1.
  • Test reports without date and signature. Annex IV point 2(g) asks for dated, signed logs. A dashboard screenshot is not one.
  • No version pinning. The documentation names one model version, production runs another. Without pinning nobody notices.

Article 11 checklist

  • Is there a mapping table from Annex IV point to source to artefact, with no empty cells and every not-applicable justified?
  • Does the dossier exist for the version running today, and did it exist before that version went live?
  • Does the registry record pin model, pre-processing, rules, data sets, policy and dossier hash, and does the gateway check this on every call?
  • Do the evaluation reports carry a date, a result per case and the signatures of business, risk and IT?
  • Is there a data sheet per data source with provenance, legal basis, labelling procedure and proof of pseudonymisation?
  • Is retention for ten years under Article 18 in place, write-once, with a signed hash outside engineering?

Open questions

  • Harmonised standards. The standards on technical documentation are still in progress. Until then every provider documents its own solutions under point 7, and two auditors can judge the same state differently.
  • Third-party models. Point 2(a) asks for information on pre-trained third-party systems. What GPAI providers supply under Article 53 and Annex XII is often not enough for that.
  • What is a relevant change? Point 6 does not say, and the boundary to a substantial modification under Art. 3(23) is blurry. We document everything and let risk classify.

Frequently asked questions

Does this apply to us as a deployer?

Article 11 addresses the provider. As an authority that buys a system, you are covered by Article 26, the fundamental rights impact assessment under Article 27 and the registration under Art. 49(3). If you develop the system yourself or modify it substantially, you become a provider under Art. 3(3) or Article 25, and Article 11 applies directly.

Is the model provider’s model card enough?

No. It describes a model; Annex IV asks for a description of your system: purpose, data, architecture, tests, oversight, risks, changes. The model card is input for point 2(a), and what you cannot verify you mark as a third-party statement.

How often do we have to update the documentation?

At every change that touches behaviour, because paragraph 1 demands the current state. Nobody keeps up by hand, so in our setup the pipeline builds the dossier, and a release without a dossier does not exist.

This post belongs to my series EU AI Act in practice. Here you find all articles at a glance.

Related: EU AI Act · All articles · AI Governance · AI Consultant

Leave Comment