AI Act Technical Documentation: The Annex IV Contents Explained Section by Section
Annex IV lists what must sit inside the technical file for a high-risk AI system. This section-by-section walkthrough covers all nine requirements, the evidence behind each, and who has to produce it.
Annex IV specifies what must sit inside the technical file for a high risk AI system: nine numbered requirements, from a general description to the post-market monitoring plan. Article 11(1) makes it mandatory before a high-risk system is placed on the market or put into service; Article 18 keeps it available to authorities for ten years afterwards.
Three searches bring most readers here: annex iv technical documentation ai act, ai act technical documentation requirements, technical file for high risk ai system. All three describe one artefact. How to write ai technical documentation comes down to the annex iv structure explained below.
What Article 11 requires — and who it lands on
Three obligations. Drawn up before the system is placed on the market or put into service, kept up to date, and containing at a minimum the elements set out in Annex IV: a floor, not a target.
Two jobs. The file must demonstrate compliance with the Chapter III, Section 2 requirements and give authorities and notified bodies what they need in clear, comprehensive form.
The ten-year clock. Article 18 requires the provider to keep that file, the Article 17 quality-management documentation, notified-body decisions and the Article 47 declaration available to authorities until ten years after the system was placed on the market.
Relief is narrow. Article 11 lets SMEs provide the Annex IV elements in a simplified manner — formatting, not content. Annex III duties apply from 2 December 2027 and Article 6(1) safety components from 2 August 2028, but documentation lead time starts when procurement does.
The map: nine numbered requirements, eight working blocks
Annex IV numbers its requirements 1 to 9; section 4 sits with section 3’s monitoring material, so most teams work in eight blocks. Write every row as ai act documentation for auditors — a reader with no context and a deadline — then map it to evidence you already hold: most of Annex IV already exists in a mid-market company.
| Annex IV section | Demands | Where the evidence already lives |
|---|---|---|
| 1 — General description | Purpose, version, interfaces, UI | AI register, vendor instructions |
| 2 — Elements and development | Design, data, validation, security | DPIA, datasheets, test reports |
| 3 — Monitoring and control | Limits, subgroup accuracy | DPIA limitations, evaluation reports |
| 4 — Performance metrics | Why these numbers fit | QMS metric definitions |
| 5 — Risk management (Art. 9) | Risk process across the lifecycle | ISO 42001 risk register, DPIA |
| 6 — Lifecycle changes | Every relevant change | Change-control records, register dates |
| 7 — Harmonised standards | Standards, or solutions used instead | Statement of Applicability |
| 8 — EU declaration (Art. 47) | Signed declaration of conformity | Vendor conformity pack, or your own |
| 9 — Post-market monitoring (Art. 72) | Field evaluation system | QMS monitoring procedures, incident log |
Annex IV section 1: the general description
What it asks for. One page: intended purpose with the context and conditions of use stated in the instructions for use; the provider’s name; and the version and its relation to previous versions. Then the plumbing — how it interacts with other hardware or software, including other AI systems; firmware versions and update requirements; and the forms in which it is supplied, such as embedded packages, downloads or APIs.
What teams miss. Where the system is a component of a product, section 1(f) wants photographs showing external features, marking and internal layout — plus a description of the deployer’s user interface, the documents your users actually hold.
Annex IV section 2: elements and development process
The heaviest block — the one that earns “design documentation”:
- Development route. Methods and steps performed, including any pre-trained systems or third-party tools and how they were used or modified.
- Design specifications. General logic and algorithms; key design choices with rationale and assumptions about the persons the system is used on; what it optimises for.
- Architecture and data. How components feed into each other and the resources used to develop, train, test and validate; datasheets covering training datasets — provenance, how data was obtained and selected, labelling.
- Oversight assessment. Measures needed under Article 14, including the aids that help deployers interpret outputs under Article 13(3)(d).
- Validation and testing. Test data, metrics for accuracy, robustness, compliance and discriminatory impacts, and dated, signed test reports.
- Cybersecurity measures. What you put in place, not a link to a policy.
Mid-market example. A manufacturer fine-tuning a vision model on its own defect images files the recipe, labelling rules and the signed false-negative report.
Annex IV sections 3 and 4: monitoring, functioning and control
Section 3 wants field behaviour: capabilities and limitations in performance, accuracy for specific persons or groups beside overall expected accuracy; foreseeable unintended outcomes and risks to health, safety, fundamental rights and discrimination; oversight measures under Article 14; and specifications for input data.
Section 4 is one sentence: the appropriateness of the performance metrics for the specific system — why these numbers are the right yardstick here. A broker whose claims-triage model reports 91 per cent overall accuracy must show what that looks like by language and claim type — disclosed in section 3, treated in section 5.
Annex IV section 5: the risk management system
Section 5 asks for a description of the risk management system under Article 9 — not a fresh risk assessment pasted into the file, but how risks are identified, evaluated against intended use and misuse, fed back from post-market data and reviewed.
The shortcut that works. If you run an ISO 42001 risk register or a DPIA programme, section 5 describes that machinery plus this system’s outputs: risk entry, treatment, residual position, owner, review date. Vagueness is what sinks it: “risks are reviewed regularly” is not a description.
Annex IV sections 6 to 8: changes, standards and the declaration
Section 6 — lifecycle changes. Retraining runs, pipeline changes, version bumps, threshold re-tuning, dependency upgrades. Evidence sits in change-control records, release notes, vendor notifications and register review dates.
Section 7 — harmonised standards. List those applied in full or part with references published in the Official Journal. Where none applies, describe the solutions adopted — for most mid-market providers, ISO/IEC 42001, ISO/IEC 27001 and ISO/IEC 23894. A certificate number is not a description of a solution.
Section 8 — the declaration. A copy of the EU declaration of conformity under Article 47: system identified, assessment route, standards applied, notified body where involved — even under internal control.
Annex IV section 9: the post-market monitoring plan
Section 9 asks for a description of the system that evaluates performance in the post-market phase under Article 72, including the plan in Article 72(3). Search ai act design documentation post market monitoring and you get two halves of this annex: section 2 is the design documentation, section 9 tests whether those assumptions held.
What the plan says. What is monitored — accuracy, drift, complaint volume, override rates, subgroup performance; where the data comes from — Article 12 logs, tickets, Article 73 incident reports; thresholds that trigger investigation; who reviews and when; how findings route back into risk management and the change log. Monthly drift reports, a quarterly review with a named owner, an off-cycle pass on incidents.
When you become the provider without meaning to
Most mid-market companies start as deployers, and Article 25 keeps them there — unless. A deployer becomes a provider when it substantially modifies a high-risk system, when it changes the intended purpose so it becomes high-risk, or when it supplies it under its own name: fine-tuning a CV-ranking model on your recruitment data, wrapping a general-purpose model into an in-house claims workflow, white-labelling a vendor’s engine.
Then Article 11 applies in full: you draw up the file, run the assessment, sign the declaration, and Article 18’s ten-year clock starts on your desk. Your register catches it early — a classification change from deployer to provider.
How to check your own position
Ten minutes, your register open:
- Count your high-risk systems. A row with no classification rationale means nothing below can be answered.
- Name the provider for each. Vendor, or you. If you cannot tell, you cannot tell whose Article 11 duty it is.
- Ask for the vendor pack. Annex IV documentation, declaration, monitoring plan, change notice.
- Run the table above against your file room. Nine rows, nine owners, nine yes or no answers.
Frequently asked questions
Who has to produce the Annex IV technical documentation?
The provider — whoever places the system on the market under its own name: the developer, a party that substantially modifies a third-party system, or a deployer that changes the intended purpose or rebrands it (Article 25). Deployers owe Article 26 duties instead: logging, oversight, input-data quality.
Do we need a notified body to review our technical documentation?
Not automatically. Many Annex III systems are assessed through internal control, where the provider documents the evidence itself; a notified body is involved only for particular categories or certification. Either way, authorities read the same Annex IV sections.
How long do we have to keep the technical documentation?
Article 18: ten years after the system was placed on the market or put into service, with the documentation, QMS records, notified-body decisions and EU declaration available to authorities throughout. Member states set conditions for insolvency or succession.
Can we satisfy Annex IV with our ISO 42001 or ISO 27001 documentation?
Partly. Certificates, the statement of applicability, risk registers, DPIAs and change-control records supply much of sections 5 to 7 and the input-data material for section 3 — but not the system-specific narrative: intended purpose, version lineage, design rationale, architecture, validation results.
What if our vendor will not share their Annex IV documentation?
You cannot compel it by asking. Make it contractual: access to the documentation, declaration of conformity, monitoring plan and change notice, with audit rights or a termination trigger. Record the answer either way — under Article 21 an authority may still request yours, and “our supplier refused” is context, not a defence.
Put this guide to work on your own estate.
Run the free exposure scan to see which of these articles actually apply to your company — 12 questions, four minutes, no card. Confirm your email and the full report unlocks.
- Free forever plan
- No credit card
- EU-hosted · GDPR-native
Build a live AI register in minutes — upgrade only when you need the audit pack.