AI Act Annex IV Technical Documentation: What It Asks — AI Automation article by Rui Codex

The EU AI Act technical documentation for a high-risk system must cover a general description, development and data details, monitoring and control measures, risk management, change control, applied standards, the EU declaration of conformity and a post-market monitoring plan. It must stay current throughout the lifecycle.

Key takeaways

  • Annex IV is a content checklist, and each of its nine areas needs a named owner and a version identifier.
  • Documentation must exist before a high-risk system is placed on the market or put into service.
  • Records written during development are far more reliable than documents reconstructed after launch.
  • Post-market monitoring and change control are part of the documentation, not separate projects.

What does AI Act Annex IV technical documentation require?

Annex IV sets the minimum content of the technical documentation that providers of high-risk AI systems must prepare. The regulation itself is published as Regulation (EU) 2024/1689 on EUR-Lex, and the annex is the part engineers and compliance leads should read line by line. It is not a template. It is a set of content requirements that your engineering records either satisfy or do not.

In practice, the annex groups its requirements into nine areas. Each area should map to a named owner, a file or wiki location and a version identifier:

  1. A general description of the system, including its intended purpose, provider, version, the hardware and software it runs on, and the instructions for use.
  2. A detailed description of the design, architecture, development methods, data requirements, and validation and testing procedures.
  3. Information on monitoring, functioning and control, including capabilities, limitations, accuracy levels, foreseeable unintended outcomes and human oversight measures.
  4. A description of why the chosen performance metrics are appropriate.
  5. A description of the risk management system.
  6. A description of the changes made to the system through its lifecycle.
  7. A list of the harmonised standards applied, or a description of the solutions adopted where no such standards were applied.
  8. A copy of the EU declaration of conformity.
  9. A description of the post-market performance evaluation system.

The wording matters. Reviewers compare your documents against the text of the annex, so a summary that paraphrases it is weaker than a document that addresses each point directly.

Who must prepare it, and when does it apply?

The provider of a high-risk AI system prepares the documentation before the system is placed on the market or put into service, and keeps it up to date afterwards. Obligations for high-risk systems listed in Annex III apply from 2 December 2027 under Article 113 of Regulation (EU) 2024/1689, as amended by Regulation (EU) 2026/1744, while systems covered through Annex I product legislation follow from 2 August 2028.

The provider role is defined by the regulation, not by who wrote the code. A company that buys a model from a vendor and then places it on the market under its own name, or makes a substantial modification, can take on provider duties. Check that role before you assume the vendor's file covers your deployment.

Article 11 of Regulation (EU) 2024/1689 also provides for a simplified form of technical documentation aimed at small enterprises and start-ups. Check whether that form has been adopted before you choose a format, and cite the official text rather than secondary summaries when you record that decision.

This article is general information and not legal or tax advice.

How should you structure the documentation?

Start from the nine Annex IV areas, assign an owner to each, and tie every section to a specific system version rather than to a project phase. Documents that follow the release cycle are easier to keep accurate than one file written at launch.

Describe the system and how it was built

The general description covers intended purpose, provider, version, hardware, software and instructions for use. The development section asks how the system was built, including design specifications, architecture and the data used. Write these from the repository and architecture diagrams, not from memory.

Record data, testing and metrics

For machine learning components, record where training, validation and test data came from, how they were split, what preprocessing was applied, and which metrics were chosen and why. A metric list without a stated reason does not answer the requirement on metric appropriateness. Keep dataset identifiers in version control so a reviewer can trace a result back to its data.

Link risk management and change control

The risk management system described in Article 9 of Regulation (EU) 2024/1689 should feed directly into the technical file, and each change should update the affected sections. Without a change log tied to model and data versions, the documentation describes a system that no longer exists.

  • Each Annex IV area has a named owner and a known storage location.
  • Every document states the system version it describes.
  • Training, validation and test datasets have stable identifiers.
  • Each chosen metric comes with a written reason.
  • Every change triggers a documentation review before release.
Tip: Keep the documentation in the same repository as the code and the model registry, so a release cannot ship without a matching document version.

Teams that need logging, model registries and ticketing connected often treat this as an integration problem. Rui Codex offers technical implementation support for such measures: see the AI Act technical measures service page for scope, and the system integration service page for connecting the tools that produce the evidence.

Hypothetical examples of documentation gaps

Here are two hypothetical scenarios:

In the first, a hiring-screening tool is retrained every quarter by an internal team. The launch documentation describes the original model well, but nobody recorded which dataset version produced the current one. The gap sits in the change management and data sections, and it only becomes visible when a reviewer asks for a trace.

In the second, an operations team adapts a vendor's classification model to its own document flow. The vendor's pack covers the base model, but the fine-tuning data, labelling rules and the team's own evaluation results exist only in a shared spreadsheet. The team assumed the vendor's pack was complete, which is an assumption worth testing early.

Where does the documentation approach fail?

The approach fails most often when documents are written after deployment, when they describe a version that is no longer running, and when third-party components are treated as someone else's responsibility. Each failure looks minor until a review asks for evidence.

Writing documentation during development costs time in every sprint, but it keeps the record accurate. Retrofitting it before launch is quicker at the start, yet it usually produces generic text that does not match the built system. Neither option removes the need for independent review.

ApproachMain benefitTypical failure
Written during developmentStays close to the code and dataSlows early sprints if owners are not named
Retrofitted before launchQuick to startDescribes intentions rather than the built system
Vendor documentation pack onlyCovers the base modelMisses fine-tuning, local data and deployment settings
Living documentation in the repositoryTied to each releaseNeeds review steps built into the pipeline to stay current

No single approach suits every team. A simple internal model may work with a short, well-maintained file, while a system with several components usually needs modular documents linked to versions. State the weak parts openly in the file, because a stated limitation is easier to defend than an unsupported claim.

How does post-market monitoring fit in?

Annex IV asks for a description of the post-market performance evaluation system, so the monitoring plan has to exist before release rather than after the first incident. A plan written later tends to measure whatever is easy to collect.

Article 72 of Regulation (EU) 2024/1689 requires providers to set up a system that actively collects and analyses relevant performance data over the system's lifetime. The plan should name what is measured, who reviews it, and how findings feed back into the risk file and the change log.

Plan storage, access and backup early, because records that cannot be retrieved during the retention period are of little use in a review. The retention period itself is set by Article 18 of Regulation (EU) 2024/1689, so quote the official text rather than a summary.

Frequently Asked Questions

Is Annex IV documentation required for every AI system?

No. Annex IV only applies to high-risk AI systems, not to every AI tool a company uses. Classification depends on the intended purpose and use case, so check the official categories before assuming a system is covered, and record the reasoning you applied.

Can a deployer rely on the provider's documentation instead?

Only partly. Provider documentation is a useful starting point, but a company that substantially modifies a system or places it on the market under its own name may take on provider duties. Fine-tuning data, local configuration and your own evaluation results still need records you control.

How long must the technical documentation be kept?

The official text sets the retention period, which runs after the system is placed on the market or put into service. Keep the versions that match each release retrievable for that whole period, and store them somewhere access can be controlled.

Does Rui Codex prepare Annex IV documentation for clients?

Rui Codex provides technical implementation support for AI regulatory measures, such as wiring logging, versioning and evidence collection into your software. It does not provide legal, tax or compliance advice, audits or certification, so regulatory review stays with your own advisers.

Sources

Rui Codex is led by a founder with 200+ projects delivered and 12+ years of hands-on experience. To discuss how these records fit your systems, request a project consultation or email info@ruicodex.com.

Tags: AI Act Annex IV high-risk AI technical documentation AI compliance AI automation EU regulation Belgium

Need Help Implementing This?

Our team can help you put these insights into practice. From AI automation to custom software development, we build solutions that deliver real results.

Send a project request