Three ISO/IEC standards form the backbone of AI management requirements for health software: ISO/IEC 42001:2023 for the AI management system itself, ISO/IEC 23894:2023 for AI risk management, and ISO 13485:2016 for the medical device quality management system that AI-enabled products must satisfy. A fourth, ISO/IEC 22989:2022, supplies the vocabulary and concepts these standards depend on. In practice, these are layered rather than interchangeable, and getting the layering right is the difference between a compliant file and an audit finding.
The distinction matters because ISO 42001 governs how an organization manages AI across its entire portfolio. ISO 13485 governs how a specific medical device is designed, validated, and maintained. An AI diagnostic tool sits at the intersection, and compliance requires both. Regulators increasingly expect evidence that both systems talk to each other.
The layered architecture of AI standards
The standards do not compete. They operate at different altitudes, and understanding that altitude determines which requirements bind a given product.
At the top sits ISO/IEC 22989:2022, which standardizes terminology. It defines what counts as an AI system, what machine learning means in a normative context, and how concepts like model training and inference are described. This sounds bureaucratic until you realize that every downstream requirement, from risk classification to post-market surveillance, uses these terms. A quality team that has not adopted consistent definitions will produce documentation that fails review for ambiguity rather than substance.
Below it sits ISO/IEC 42001:2023, the first certifiable AI management system standard. It follows the same Annex SL structure as ISO 27001 and ISO 9001, which means it integrates cleanly with existing management systems. It requires an AI policy, defined roles, an impact assessment process, and a documented risk treatment plan. Certification is possible through accredited bodies, which distinguishes it from guidance-only documents.
ISO/IEC 23894:2023 sits alongside 42001 as the risk management companion. It adapts the ISO 31000 risk framework to AI-specific hazards: data bias, model drift, adversarial inputs, and opacity. It does not certify on its own, but it provides the methodological spine that 42001's risk treatment clause expects.
Then comes ISO 13485:2016, the medical device quality management system standard. It predates the AI wave, but its design controls, verification and validation requirements, and post-market surveillance clauses apply directly to AI-enabled devices. Most regulatory submissions reference it explicitly.
The mechanism that ties these together is documentation traceability. Each standard demands records, and the records must cross-reference. A risk identified under 23894 must appear in the 42001 risk register and in the 13485 design history file. When those three trails diverge, auditors notice.
Why AI-specific requirements diverge from traditional software rules
Traditional medical software is deterministic. Given the same input, it produces the same output, and verification confirms that the code does what the specification says. AI systems break this assumption at the foundation.
A trained model's behavior emerges from data rather than from explicit logic. This means verification shifts from code inspection to performance validation on representative datasets. ISO 13485's clause on design verification assumes you can demonstrate conformance to specified requirements. For a model, "specified requirements" must be expressed as performance metrics with acceptance thresholds, and those thresholds must be justified before training begins, not after results are known.
The second divergence is change management. A conventional device changes when software is updated. A model can change behavior without any code change if it continues learning in deployment, or if the input data distribution shifts. ISO 13485 requires change control, and AI forces a broader definition of what constitutes a change. A silent degradation in accuracy on a demographic subgroup is a change in device performance even if no one touched the software.
ISO/IEC 42001 addresses this by requiring continual monitoring as a management system obligation rather than a one-time validation activity. The system must define how performance is tracked, what thresholds trigger action, and who is accountable. That is a governance requirement, not an engineering one, and it is the gap most traditional quality systems miss.
The third divergence is transparency. ISO/IEC 22989 establishes concepts like explainability and interpretability, and downstream standards expect documentation of model limitations. A clinician using an AI triage tool needs to know the population the model was trained on and the conditions under which it was not validated. That information is a labeling requirement in spirit, even where regulations do not yet mandate a specific format.
How the standards map to regulatory pathways
The ISO layer does not replace regulation. It supplies the evidence structure that regulators evaluate.
In the United States, the FDA's device software guidance and its predetermined change control plan pathway for AI-enabled devices reference quality system expectations that align with ISO 13485. A manufacturer pursuing a predetermined change plan must document the types of modifications anticipated and the validation approach for each. That documentation structure maps naturally onto ISO/IEC 42001's change management and impact assessment clauses.
In the European Union, the Medical Device Regulation requires a quality management system, and harmonized standards create presumption of conformity. ISO 13485 is the reference point for that system. The AI Act adds a separate layer of obligations for AI systems, and ISO/IEC 42001 is positioned as a tool for demonstrating compliance with its management system requirements, though formal harmonization status evolves and should be checked against current official lists rather than assumed.
The practical consequence is that a manufacturer selling into both markets maintains one quality system under ISO 13485, one AI governance layer under 42001, and maps both to each jurisdiction's requirements. Duplicating systems invites inconsistency. Integrating them is the efficient path.
| Standard | Scope | Certifiable | Primary output | |
|---|---|---|---|---|
| ISO/IEC 22989:2022 | AI concepts and terminology | No | Shared vocabulary | |
| ISO/IEC 42001:2023 | AI management system | Yes | Governance framework and risk register | |
| ISO/IEC 23894:2023 | AI risk management | No | Risk assessment methodology | |
| ISO 13485:2016 | Medical device quality system | Yes | Design history file and validation records | |
| IEC 62304 | Medical device software lifecycle | No (referenced) | Software safety classification and lifecycle records | |
| ISO 14971 | Medical device risk management | No (referenced) | Hazard analysis and risk file |
The two IEC and ISO entries at the bottom matter because they complete the chain. IEC 62304 assigns software safety classes from A to C, and an AI model's clinical function determines its class. ISO 14971 governs hazard-based risk management for the device, while 23894 handles the AI-specific risks that 14971's framework does not natively address, such as training data bias.
The edge cases that break naive compliance
The hardest cases are not the obvious ones. A standalone diagnostic classifier with a fixed model is relatively tractable. The problems cluster around four scenarios.
Adaptive models that update in deployment create a moving validation target. If the model learns from clinical feedback, the validated version is not the deployed version after some interval. Manufacturers must either lock the model between releases or define a controlled update process with pre-specified validation, and the latter is far harder to document credibly.
Multi-component systems where the AI is one module among several raise boundary questions. If a triage model feeds a rule engine that feeds a clinician-facing display, which component owns the risk when the composite output is wrong? ISO/IEC 42001's system-level view helps, but the design history file must still assign responsibility clearly.
Foundation models adapted for clinical use introduce third-party dependence. The base model's training data is often undisclosed, which limits the bias assessment that ISO/IEC 23894 expects. A manufacturer can document what it knows and flag what it does not, but claiming a complete risk assessment over an opaque upstream model is a red flag in any audit.
Warning: claiming full validation of a third-party foundation model without documentation of its training data and known limitations. Warning: treating a deployed model as unchanged when input data distributions have shifted. Warning: certifying to ISO/IEC 42001 while the device-level quality system under ISO 13485 remains unintegrated. Warning: setting performance thresholds after seeing test results rather than before. Warning: omitting subgroup performance data because aggregate accuracy looks acceptable.
A counterintuitive detail worth internalizing: aggregate accuracy is often the least informative metric in a regulatory review. A model with strong overall performance can still fail on a subgroup that represents a meaningful share of the intended use population. ISO/IEC 23894's emphasis on harm to individuals and groups, rather than to averages, is what makes it more than a relabeled generic risk standard. Plainly: a model can look good on paper and still harm a specific group of patients.
From requirements to an operating program
Translating standards into practice requires sequencing, not simultaneous adoption. The following steps reflect how mature organizations build the stack.
First, adopt terminology from ISO/IEC 22989 across quality, engineering, and regulatory affairs. Inconsistent definitions generate documentation conflicts that surface late.
Second, establish ISO 13485 as the base quality system if it is not already in place. An AI management system built on a weak device quality foundation will not hold.
Third, implement ISO/IEC 23894's risk methodology within the ISO 14971 hazard framework. Map AI-specific risks such as drift and bias to existing hazard categories rather than creating a parallel risk file.
Fourth, build the ISO/IEC 42001 management system with explicit interfaces to the device quality system. Define who owns model monitoring, who reviews performance thresholds, and how findings flow into corrective action.
Fifth, define and document pre-specified acceptance thresholds for each performance metric and subgroup before validation begins. This is the single most scrutinized element in AI device reviews.
Sixth, establish a post-market surveillance process that tracks model performance in the field, not just complaints. Drift detection requires data collection that complaint handling does not capture.
A useful threshold framing: safety-related model failures warrant the same corrective action rigor as hardware failures. If a model's sensitivity on a critical subgroup falls below the pre-specified floor, that is a corrective action trigger, not a monitoring note.
Timeline of standards development
| Year | Milestone | Relevance to health AI |
|---|---|---|
| 2016 | ISO 13485:2016 published | Current device quality system baseline |
| 2019 | ISO/TR 24291 published | Early AI concepts report, informational |
| 2022 | ISO/IEC 22989:2022 published | Normative AI vocabulary |
| 2023 | ISO/IEC 23894:2023 published | AI risk management guidance |
| 2023 | ISO/IEC 42001:2023 published | First certifiable AI management system |
| Ongoing | AI Act harmonized standards work | EU conformity pathway development |
The timeline shows a compression: three foundational AI standards landed within roughly eighteen months. Organizations that built quality systems before 2022 often need retrofitting rather than greenfield adoption, which is harder.
What the standards do not cover
The gaps are as important as the requirements. No ISO standard currently certifies model accuracy for a specific clinical use. Certification under 42001 confirms that a management system exists and operates; it does not confirm that a given model performs well. That distinction is frequently misunderstood by procurement teams who treat a certificate as a performance guarantee.
The standards also do not resolve liability allocation between model developers, device manufacturers, and deploying health systems. ISO/IEC 42001 requires defined roles and responsibilities within an organization, but cross-organizational accountability remains a contractual and legal matter.
Nor do they prescribe clinical evidence levels. A management system can be fully conformant while the underlying clinical validation is thin. Reviewers who check only for certificates will miss this, and the standards themselves do not prevent that error.
Finally, the standards are living documents. ISO/IEC 42001 will be revised, and the AI Act harmonization process will shift which standards carry presumption of conformity. Treating any standard as fixed is itself a compliance risk. The practical posture is to build the management system to survive revision, with modular documentation that separates governance from technical specifics.
This article is general educational information about standards frameworks, not a regulatory opinion or compliance certification. Device classification, validation requirements, and jurisdiction-specific obligations vary and should be confirmed with a qualified regulatory professional and current official guidance before any submission.