Annex IV technical documentation, section by section
Annex IV is the table of contents for the technical file. It is boring, specific and impossible to fake: nine sections, each answering a question a notified body will ask. Here is what belongs in each one.
Article 11 requires technical documentation for high-risk systems, and Annex IV defines its content. The document has two audiences: the conformity assessment body, which checks it, and the deployer's technical team, which has to run the system. Write it for the second and it will satisfy the first.
1. General description
- Intended purpose, in the words of Article 3(12): the use for which the system is intended, including specific context and conditions.
- Provider identity, commercial name and version.
- How the system interacts with hardware or software outside itself, including other AI systems.
- Forms in which it is placed on the market: API, on-premise, embedded.
- Hardware requirements and, where relevant, photographs or diagrams.
- Description of the user interface as the deployer sees it.
- Instructions for installation and configuration.
Auditors check one thing here: whether the intended purpose is specific enough to bound the risk assessment. "AI platform for HR" is not an intended purpose. "Ranks applicants for warehouse roles against a fixed job description" is.
2. Detailed description of elements and development
- Methods and steps used to build the system, including third-party tools, pretrained models and datasets.
- Design specifications: architecture, components, the role of each model.
- Computational resources used for training, validation and inference.
- Data requirements: datasheets, provenance, scope, labelling rules, cleaning steps, known gaps.
- The assessment of human oversight measures required by Article 14.
- Pre-determined changes and the conditions under which the system may be updated without a new assessment.
- Validation and testing: procedures, datasets, metrics, test logs, results per subgroup.
- Cybersecurity measures.
Pre-determined changes. If the model, prompt or threshold can change without a new conformity assessment, Annex IV expects you to say which changes are pre-approved and how they are controlled. Silent model upgrades are a finding.
3. Monitoring, functioning and control
- Capabilities and limitations of the system, in operational language.
- Accuracy, robustness and cybersecurity metrics, with the datasets and conditions they were measured under.
- Foreseeable unintended outcomes and the mitigations in place.
- Human oversight measures, including the interface features that make intervention possible.
- Specifications for input data: formats, ranges, quality requirements, rejected inputs.
4. Risk management system
A description of the Article 9 process, not just a register: how risks were identified, how they were evaluated against acceptance criteria, which measures were adopted, what residual risk remains and who accepted it. The register is an annex; the description is the section.
5. Post-market monitoring plan
- What is monitored and how often (accuracy, drift, override rate, complaints, incidents).
- Data sources: logs, user feedback, dealer reports, support tickets.
- Thresholds that trigger investigation.
- Who reviews, who decides, and what happens with the finding.
- How the plan links to Article 72 reporting and Article 73 incident reporting.
6–9. The short sections
| Section | Content | Why it matters |
|---|---|---|
| 6. Changes over the lifecycle | Version history with dates and what changed | Proves the documented system is the shipped system |
| 7. Harmonised standards applied | List with versions, or justification if none | The standardisation gap is why deadlines moved; say what you used |
| 8. EU declaration of conformity | Signed copy, with the assessment procedure referenced | The legal claim that the system conforms |
| 9. Post-market evaluation | How the deployed system's performance is evaluated | Closes the loop with section 5 |
How to keep it alive
The technical file is a living document and the most common audit failure is a file that describes version 1 while version 4 is in production. Three habits prevent that: version the document in the same repository as the code, require a documentation change in the pull request that changes behaviour, and generate the parts that can be generated — model versions, test results, control status — instead of retyping them.
That last point is where an evidence pack earns its place. The pack is not the technical file, but it carries the parts that change constantly: model versions in use, the control status on a date, approvals, and the log chain. The technical file explains the design; the pack proves the operation.
Classification, control status, model versions, oversight records and the hash chain — one click, PDF and JSON, sealed.
Start with the free checkFrequently asked
Who writes the technical documentation?
The provider. Deployers receive it through the instructions for use and must keep their own operational records. If you fine-tune or rebrand a vendor model, check whether you have become a provider for that configuration.
Does it have to be a PDF?
No. Annex IV defines content, not format. What matters is that it is versioned, complete and matches the system in production. A generated document is usually more accurate than a manually maintained one.
How long does it take to produce?
For an existing system with decent engineering records, two to four weeks of focused work. The long pole is validation results per subgroup and the data governance section, not the writing.
What if we have no harmonised standards to cite?
Say so, and describe the alternative technical solutions you used to meet the requirements. The AI Digital Omnibus acknowledged the standards gap; a documented alternative is defensible, an empty section is not.