AI Act Article 9 requires providers of high-risk AI systems to establish, implement, document and maintain a risk management system. The process must run across the whole lifecycle: it identifies and evaluates foreseeable risks, adopts measures to reduce them, tests the system and stays current after release.
Key takeaways
- Article 9 treats risk management as a continuous process that runs for the whole lifecycle of the system, not as a one-off document.
- Providers must identify foreseeable risks, including reasonably foreseeable misuse, and estimate them for the intended purpose.
- Residual risk must be judged acceptable after reasonable measures, and the reasoning must be documented and passed on to deployers.
- Testing and post-market monitoring feed new evidence back into the risk analysis.
What does AI Act Article 9 require?
Article 9 requires providers of high-risk AI systems to set up, run, document and maintain a risk management system. The obligation sits with the provider, which is the organisation that develops the system or has it developed and places it on the market under its own name.
The text of the regulation is in Regulation (EU) 2024/1689, the AI Act, on EUR-Lex. This article is general information and not legal or tax advice. Whether a particular system is high-risk depends on Article 6 and the Annex III categories, and that classification should be checked with qualified counsel before any risk work begins.
The risks Article 9 cares about are risks to health, safety and fundamental rights. Business losses matter, but they are not the regulatory focus. A risk file that only lists downtime or cost overruns misses the point of the article.
Article 9 also asks providers to consider whether the system could adversely affect people under 18 and, where relevant, other vulnerable groups. That consideration should appear in the analysis itself, not in a footnote.
How do you identify and estimate risks step by step?
Start by defining what the system is meant to do, then look for ways it can cause harm in that use and in foreseeable misuse. Each identified risk is estimated for likelihood and severity before any control is chosen.
Step 1: Define the intended purpose and context
Describe the users, the people affected, the input data, the environment and the decisions the system influences. Most weak risk files fail here, because a vague purpose makes every later estimate vague as well.
Step 2: Identify known and foreseeable risks
Work through the lifecycle stages: data collection, training, validation, deployment, operation and retirement. For each stage, ask how the output could be wrong, how it could be misread, and how someone could use it outside its intended purpose.
Step 3: Estimate and evaluate each risk
Rate each risk for likelihood and severity under normal use and under foreseeable misuse. Keep the scale simple and explain it, because an unexplained rating cannot be reviewed.
- Name the affected group and the type of harm for every risk.
- Record the data and test evidence behind each rating.
- Flag risks that depend on other systems or on human behaviour.
- Use post-market data, where it exists, to update the ratings.
When is residual risk acceptable?
Residual risk is acceptable only when, after reasonable measures have been applied, the remaining risk is judged acceptable for the intended purpose and the provider can explain why. Article 9 does not set a numeric threshold, so the judgement itself must be documented.
The usual order of measures is to eliminate or reduce risk through design first, then apply protective measures such as limits, fallbacks and human oversight, and finally give clear information to deployers about what remains. Each step has a cost in accuracy, speed or usability, and that trade-off should be recorded.
The order fails in predictable ways. A confidence threshold can reduce harmful outputs while silently increasing missed cases for one group. A human review step can be ignored when staff are under time pressure. An instruction for deployers protects no one if it is never read or followed.
Technical implementation of these measures is where many teams need outside help. Rui Codex offers technical measures for AI Act compliance as implementation support, not as legal advice.
How do testing and monitoring keep the risk file current?
Testing checks that the chosen measures work for the intended purpose, and monitoring after release feeds new evidence back into the analysis. The risk file changes whenever either one shows something the earlier analysis did not expect.
Testing should reflect the conditions in which the system will actually run, including the data quality problems that appear in production. Results should be broken down by relevant subgroups, because an average can hide a failure that affects one group badly.
Monitoring needs defined triggers. Examples include a rise in overrides, a drift in input data, a spike in complaints, or a change in how deployers use the output. Each trigger should lead to a documented review, not an informal conversation.
Hypothetical examples of a risk file in practice
Here are two hypothetical scenarios:
Scenario A. A hypothetical screening tool ranks job applications for a recruitment team. The analysis finds that some input fields act as proxies for protected characteristics. The provider removes those fields, tests outcomes across subgroups, and adds a documented human review step before any rejection. The residual risk note tells the deployer which cases still need careful review.
Scenario B. A hypothetical maintenance-planning assistant recommends when a production line should stop for inspection. The analysis finds that a wrong recommendation could cause an unsafe shutdown or a missed fault. The provider adds a fallback to the existing manual schedule when confidence is low, and reviews operator overrides every month to decide whether the threshold needs adjusting.
In both cases the value of the file lies in the links between the risk, the measure, the test and the monitoring trigger. Without those links, the file describes intentions rather than controls.
Where does this approach fail?
The approach fails most often when the risk file is written once before launch and never revisited. It also fails when the intended purpose is defined so broadly that no realistic risk can be ruled in or out.
- When the intended purpose is vague, the risk estimates have no reliable basis.
- When testing uses clean data only, production failures remain hidden.
- When monitoring has no owner, triggers are logged but never acted on.
- When residual risk is accepted without written reasoning, the decision cannot be defended later.
Rui Codex is led by a founder with 200+ projects delivered, and it works with operations and IT leads who need these processes built into real systems. Published project details are at the Rui Codex case studies page.
Frequently Asked Questions
Does AI Act Article 9 apply to every AI system?
No, Article 9 applies only to high-risk AI systems, so the first task is to confirm whether your system falls into that category under Article 6 and the Annex III list. Systems outside the high-risk category do not need this formal risk management system, although good risk practice still helps. Classification is a legal question, so it should be checked with qualified counsel before you build the risk file.
Is the risk management system a document you write once?
No. Article 9 describes a continuous, iterative process that runs through the whole lifecycle of the system, with regular systematic review and updating. The written file is the record of that process, so it must change when the design, the data, the intended purpose or the monitoring evidence changes.
Does Article 9 require a third-party certificate?
Article 9 itself does not ask for a certificate. Whether a third-party assessment applies depends on the type of system and the conformity assessment route in the AI Act, so the provider should check the text and seek qualified advice. Technical evidence from testing and monitoring is needed in every case.
Who should own the risk management work inside a company?
The provider's management should assign clear ownership, because the obligation sits with the provider rather than with a single engineer. In practice the work draws on engineering, operations and legal input, and a named owner keeps the risk file current. Rui Codex offers technical implementation support for this work, not legal or compliance advice.
Conclusion
Article 9 turns risk management into an ongoing engineering discipline. A useful risk file defines the intended purpose clearly, links each risk to a measure and a test, accepts residual risk only with written reasoning, and uses monitoring to trigger review. Teams that build these links early find the regulation far easier to live with. If you are preparing a high-risk system and need technical support for the implementation side, request a project consultation at Rui Codex or email info@ruicodex.com to discuss your project.
Sources
- Regulation (EU) 2024/1689 (Artificial Intelligence Act) - EUR-Lex, Publications Office of the European Union