prEN 18228 explained: the AI risk management standard behind Article 9
prEN 18228 is the draft European standard meant to turn Article 9 into a repeatable process. It is not law and it is not finished — but it is the shape auditors will expect, and the reason the high-risk deadline moved to December 2027.
Article 9 of the AI Act requires a risk management system for high-risk AI. It says what the process must achieve, not how to run it. That gap is what harmonised standards are for, and prEN 18228 is the CEN-CENELEC draft covering AI risk management.
prEN 18228 is a draft. Clause numbering and wording can change before publication, and citing a draft clause in a compliance file is risky. Use it to structure your process, and cite it as a draft in the document. This article describes the expected shape, not the final text.
Why it exists
Before the AI Act, AI risk was handled either as a generic enterprise risk (ISO 31000) or as a model-quality exercise. Neither fits. Enterprise risk does not see model drift; model quality does not see fundamental rights. prEN 18228 is meant to bridge them: a risk process that covers health, safety and fundamental rights across the lifecycle, and that plugs into the quality management system required by Article 17.
How it relates to the standards you already know
| Standard | Role | Relationship |
|---|---|---|
| ISO 31000 | Generic risk management principles | The vocabulary and the process skeleton |
| ISO/IEC 23894 | AI risk management guidance | The AI-specific interpretation of 31000 |
| ISO/IEC 42001 | AI management system | The organisational system the risk process sits inside |
| prEN 18228 | European AI risk management standard | The harmonised route to presumption of conformity with Article 9 |
| prEN 18229-1 | Logging and record keeping | Feeds the risk process with operational evidence |
The process a compliant system runs
- 1Scope and criteria. Define the system, its intended purpose, the affected parties and what counts as unacceptable risk. This is where most projects fail: the scope is written to exclude the hard cases.
- 2Identification. Enumerate risks per lifecycle stage: design, training, validation, deployment, monitoring, decommissioning. Include risks from misuse and from the context of use, not only from the model.
- 3Analysis. For each risk: cause, event, consequence, affected group, existing controls, likelihood and severity. Qualitative scales are acceptable if they are defined and applied consistently.
- 4Evaluation. Compare against the acceptance criteria. Record who accepted residual risk and on what basis.
- 5Treatment. Choose controls: design change, data change, guardrail, human review, restriction of use, disclosure. Record the expected effect and verify it.
- 6Monitoring and review. Feed incidents, drift, complaints and post-market data back into the register. The process is iterative, not annual.
- 7Documentation and communication. The register, the decisions and the residual risk statement go into the technical file (Annex IV section 4).
Article 9 adds two constraints people miss: the process must run across the entire lifecycle, and it must address risks to fundamental rights, not only to health and safety. A risk register with no row for discrimination is incomplete for any system that scores people.
What the evidence looks like
| Artefact | Minimum content | Owner |
|---|---|---|
| Risk policy | Scope, criteria, roles, review cadence | Management |
| Risk register | ID, hazard, affected group, stage, controls, residual rating, owner | Product / risk |
| Acceptance record | Who accepted which residual risk and why | Accountable person |
| Treatment plan | Control, expected effect, verification method, date | Engineering |
| Review minutes | What changed, what was re-assessed | Risk owner |
The register is the artefact auditors ask for first, and the one most often written retrospectively. A register assembled the week before an audit has no history, no rejected risks and no evidence of change — three signals that the process is not real.
Why the delay happened
Regulation (EU) 2026/1744 moved the Annex III deadline to 2 December 2027 and the Annex I deadline to 2 August 2028, citing the failure of CEN and CENELEC to deliver the harmonised standards. prEN 18228 is one of those standards. Until it is published, conformity assessment rests on the Act's essential requirements plus whatever alternative technical solutions you can document.
Practically, that means two things. You cannot claim presumption of conformity from a standard that does not exist yet, so describe your method explicitly. And you should design the process to survive the final standard: use the ISO vocabulary, keep the register structured, and avoid tying your documentation to draft clause numbers.
TraceSeal maps seven risk-management controls to the evidence each one needs, and keeps their status alongside the rest of the file.
See the control setFrequently asked
Is prEN 18228 mandatory?
No. Standards are voluntary; compliance with the AI Act is mandatory. Applying a harmonised standard gives a presumption of conformity, which is why the delay in publishing them moved the deadline. Until publication, you document the alternative you used.
Can we just use ISO/IEC 23894?
You can use it as your method, and many teams will. The risk is evidentiary: a European harmonised standard is the shortest path to presumption of conformity. Until prEN 18228 is final, cite ISO/IEC 23894 and explain how it maps to Article 9.
How big should the risk register be?
Large enough to cover every affected group and lifecycle stage, small enough that each row is maintained. Twenty to forty live risks is typical for a focused system; a hundred rows that nobody reviews is worse than twenty that are.
Does fundamental rights risk need a lawyer?
It needs legal input at least once, to define the relevant rights and the unacceptable outcomes. After that, product and risk owners can maintain the register, with legal review when the purpose or the population changes.
Sources
Keep reading
Which AI systems Annex III makes high-risk, how the Article 6(3) filter works, the full obligation checklist under Articles 9–15, 17, 43 and 72, and the timeline after the AI Digital Omnibus.
What goes into each of the nine Annex IV sections, what an auditor checks in each one, and the parts teams almost always leave empty.