Smart riskieren statt dumm riskieren: warum Wachstum Mut braucht | Umair Zaffar
Umair Zaffar October 20, 2026 0 Comments

EU AI Act Article 9: Risk Management System, Implemented in Practice

EU AI Act Article 9 requires a risk management system for high-risk systems that runs across the whole lifecycle, and at SIFAMO we implement it as a risk register per model whose measures the gateway enforces and whose residual risk the business owner signs.

Article 9 is the article most projects want to settle with a Word document: filled in once before go-live, never touched again. That is exactly what the text rules out, because it asks for a continuous process, not a report. It addresses providers of high-risk systems, which includes the bank that builds its credit model itself or substantially modifies one it bought. I am an engineer, not a lawyer. So this post is about how the obligation becomes a living artefact. The basics on deadlines and roles are in my EU AI Act guide.

For context: the EU AI Act has been in force since 1 August 2024. The prohibitions and AI literacy obligations 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 amending Regulation (EU) 2026/1744 (Digital Omnibus), from 2 December 2027, and for Annex I products from 2 August 2028. Article 9 belongs to these high-risk obligations. The Omnibus moved deadlines; it removed no obligation.

What Article 9 requires

Article 9 sits in Chapter III, Section 2. It does not say how much risk a system may carry. It says how the provider has to deal with risk:

  • Art. 9(1): The provider must establish, implement, document and maintain a risk management system.
  • Art. 9(2): The system is “a continuous iterative process planned and run throughout the entire lifecycle of a high-risk AI system” and must be regularly reviewed and updated. Four steps: identify and analyse the known and reasonably foreseeable risks to health, safety and fundamental rights when the system is used as intended (a); estimate and evaluate the risks under reasonably foreseeable misuse (b); evaluate further risks from post-market monitoring under Art. 72 (c); adopt appropriate and targeted risk management measures (d).
  • Art. 9(3): Only risks that can reasonably be mitigated or eliminated through development, design or adequate technical information are in scope.
  • Art. 9(4): Measures must take into account the combined effects of the other requirements in this section: data, documentation, record-keeping, transparency, human oversight, and accuracy, robustness and cybersecurity (Art. 10 to 15).
  • Art. 9(5): The residual risk per hazard and overall must be judged acceptable. There is an order of preference: eliminate or reduce risks through design and development as far as technically feasible; mitigate and control risks that cannot be eliminated; provide information under Art. 13 and, where appropriate, training to deployers. The deployer’s knowledge, experience and context of use have to be considered.
  • Art. 9(6) to (8): The system must be tested to identify the most appropriate measures and to ensure consistent performance for its intended purpose. Testing may take place in real-world conditions under Art. 60. It runs throughout development and in any event before placing on the market, against predefined metrics and probabilistic thresholds.
  • Art. 9(9): The provider considers whether the system is likely to adversely affect persons under 18 or other vulnerable groups.
  • Art. 9(10): Providers already subject to internal risk management requirements under other Union law, such as banks, may integrate the Article 9 requirements into those procedures.

The technical example

A bank runs a pre-check model in retail lending. It does not decide on the loan. It reads the pseudonymised application, the product choice and a structured summary of the submitted documents, and proposes a route: fast lane through the existing rules engine, standard review by a case handler, or specialist review by an experienced team. The credit decision is then made by the rules engine or by a human.

Why is this still high-risk? Annex III point 5(b) covers systems used to evaluate the creditworthiness of natural persons. A routing model does not formally assess creditworthiness, but the route determines who sees the application, how fast and with what expectation. Art. 6(3) has an exception for narrow procedural tasks and preparatory activities. We do not rely on it, because the route shapes the outcome. So we treat the model as a high-risk system, and Article 9 applies in full.

The data flow: the application platform never calls the model. It calls the gateway with a use-case ID. The gateway checks registry record and data zone, replaces names, account numbers and addresses with IDs, calls the model, evaluates the signals and writes the evidence log. The Article 9 obligation bites in three places: the data the model may see (red zone never, gray zone only pseudonymised and within scope), the routes that affect people, and misuse by the bank’s own staff, when a case handler reads the route as a verdict on creditworthiness.

How we implement it

At SIFAMO we do not treat the risk management system as a document beside the system. We treat it as part of the registry record that the governance layer reads at runtime. Every model has a risk register in the registry with its own version. Simplified, it looks like this:

{
  "model_id": "DE-B-2026-0021",
  "purpose": "Credit pre-check: routing of applications (fast lane, standard review, specialist review), no decision",
  "risk_class": "high (Annex III point 5(b)), tier 3 (internal)",
  "risk_register": {
    "version": "credit-routing-risk-v7",
    "hazards": [
      {"id": "H-03", "zone": "gray", "risk": "Proxy discrimination via postcode and first name",
       "measure": "Design: features removed; gate: bias run; gateway: rule G-11",
       "residual": "low", "accepted_by": "Business owner retail credit, 2026-06-15"},
      {"id": "H-07", "zone": "red", "risk": "Raw account data in the prompt",
       "measure": "Gateway: zone check, pseudonymisation with IDs", "residual": "very low"},
      {"id": "H-12", "zone": "all", "risk": "Case handler reads the route as a score (misuse)",
       "measure": "Oversight tier 3 for specialist review, rationale visible, training", "residual": "medium"}
    ],
    "overall_residual": "acceptable",
    "reevaluate_on": ["model change", "prompt change", "index change", "expiry 2026-12-15"]
  }
}

Identify hazards per data zone

We run the hazard analysis under Art. 9(2)(a) and (b) along the three data zones. Red is raw personal data: what happens if it reaches the model? Gray is pseudonymised data with access control: which proxy features remain that can carry discrimination? Green is non-critical data: what misuse is still possible? The zones provide the language the gateway later checks against. We take misuse literally: the most common misuser is not an attacker but the colleague who turns the routing into a score.

Map measures to gateway rules and oversight tiers

Every hazard gets a measure in the order Art. 9(5) prescribes. Design first: postcode, first name and date of birth are not in the input schema at all. Then control: whatever design cannot remove becomes a gateway rule with its own ID, such as the zone check, pseudonymisation or abstention when support is low, or an oversight tier: the specialist-review route is tier 3 and proceeds only with full human review and a visible rationale, standard review gets sampled review with a blind re-check, the fast lane gets monitoring and a comparison against the rules engine. Last, information and training: the deployer manual for case handlers says in one sentence what the route is and what it is not. A measure without a rule ID or a tier in the registry is, for us, not a measure but an intention.

The business owner accepts the residual risk

Art. 9(5) requires a judgement on the residual risk per hazard and overall. Who judges, the text leaves open. We decided: the business owner, meaning the head of the lending business, accepts the residual risk with name and date in the registry, the risk function reviews the judgement, and IT confirms that the measures are wired up. I insist on the business owner because that is where responsibility for the customers sits. An IT department that accepts a discrimination risk has nothing to accept.

Test against predefined thresholds

Paragraphs 6 to 8 are, for us, the pre-deployment gates: an evaluation set owned by the business, regression tests per case, attack scenarios for this use case and the bias run with varied protected attributes. Metrics and thresholds are in the registry before the run, not after it. For Art. 9(9) the evaluation set contains dedicated cases for young adults without credit history and for people with limited German.

Re-evaluate on every change and at the expiry date

This is where the register comes alive. Every change to model, prompt or retrieval index changes the version fields in the registry record, and the gateway routes no version without an approved risk register. Then there is the expiry date: by then at the latest the full re-evaluation runs, otherwise the model loses its route. And Art. 9(2)(c), post-market monitoring, we feed from the evidence log: reviewer actions with their reason category, drift alerts and out-of-scope attempts are held against the risk register every month. A pattern that maps to no hazard becomes a new one. The kill switch held by the business owner is the last measure in the register.

What the auditor sees: the versioned risk register with zone, measure, rule ID or tier, residual risk and signature per hazard, plus the test results against the predefined thresholds and the history of re-evaluations. We use Art. 9(10) by having the bank’s risk function pull these records into its existing risk inventory under MaRisk. More on this approach under AI Governance.

Typical mistakes

  • The risk management system is a document. Written once before go-live, never tied to a version. After the first prompt change it describes a system that no longer exists.
  • Measures without enforcement. “Staff will be made aware” sits as a measure next to a discrimination risk. A measure that does not exist as a gateway rule, an oversight tier or a gate cannot be evidenced.
  • Thresholds are set after the test. Then every model passes. Art. 9(8) requires the order the other way round.

EU AI Act Article 9 checklist

  • Every high-risk model has a versioned risk register linked to the model, prompt and index version in the registry record.
  • Hazards are identified per data zone, including reasonably foreseeable misuse by staff and by applicants.
  • Every measure is mapped to a gateway rule, an oversight tier, a gate or a training item, following the order in Art. 9(5).
  • The residual risk is judged per hazard and overall and accepted by the business owner with name and date, with review by the risk function.
  • Metrics and thresholds are defined and documented before testing; test results are available per case.
  • Minors and vulnerable groups are explicitly addressed in the hazard analysis and in the evaluation set.
  • Every change to model, prompt or data triggers a re-evaluation, the expiry date forces a full repeat, and operational signals flow back on a fixed schedule.

Open questions

  • What does “acceptable” mean? The text gives no yardstick for acceptable residual risk. The harmonised standard on risk management has not been adopted yet.
  • Foreseeability with language models. Art. 9(2) assumes risks can reasonably be foreseen. With language models a remainder stays that no register covers. Interpretability research is the only path I see here.
  • Routing or evaluation? Whether a pre-check model falls under Annex III point 5(b) or under the exception in Art. 6(3) is not settled. We choose the stricter reading.

Frequently asked questions

Does this apply to us as a deployer too?

Article 9 addresses the provider. As a deployer under Art. 26 you have to follow the instructions for use, ensure human oversight and monitor operation, but you do not run your own Article 9 risk management system. As soon as you develop the system yourself, put it into service under your own name or substantially modify it, you become the provider under Art. 25.

Is our existing risk management under MaRisk enough?

Art. 9(10) allows integration, but it does not replace the content. The four steps, the order of measures, the tests against predefined thresholds and the residual risk per hazard have to be demonstrable within the existing procedure, in practice as a risk inventory that knows models as objects of their own, with version and expiry date.

How often do we have to re-evaluate?

On every change to model, prompt or data, and independently of that at a fixed date. We work with the expiry date of the registration, for customer-facing models after a few months. A risk register that is only opened at release time does not satisfy paragraph 2(c).

This post is part of the series on the EU AI Act article by article, all articles at a glance.

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

Leave Comment