The first harmonised European AI standard has arrived. EN 18286:2026 marks an important step in translating the EU AI Act into operational requirements. But does this mean organisations that have already implemented ISO 42001 need to start over? Not at all. The more interesting question is what these standards are designed to achieve — and whether they can keep pace with how AI systems are evolving.
The AI Act tells us what to achieve. Standards help explain how.
On 14 October 2026, EN 18286:2026 became the first harmonised European AI standard to be cited in the Official Journal of the European Union.
The EU AI Act sets out requirements for high-risk AI systems covering risk management, data governance, technical documentation, transparency, human oversight, accuracy, robustness and cybersecurity. But translating these requirements into engineering decisions and operational controls is not straightforward.
What constitutes adequate data quality? How should acceptable levels of bias be determined? What evidence demonstrates that risks have been sufficiently mitigated?
This is where harmonised European standards come in. Developed through CEN-CENELEC following a standardisation request from the European Commission, they provide technical and organisational specifications to support the implementation of the AI Act.
Their significance goes beyond practical guidance. Once cited in the Official Journal, compliance with the relevant provisions of a harmonised standard can provide a presumption of conformity with the legal requirements it covers. This does not mean automatic compliance with the entire AI Act, but it provides a legally recognised route for demonstrating conformity with specific obligations.
EN 18286 is the first AI Act standard to reach this milestone. Its arrival also reveals an important distinction between organisational AI governance and regulatory conformity.
A different perspective on quality
EN 18286 addresses Article 17 of the AI Act, which requires providers of high-risk AI systems to establish and maintain a Quality Management System (QMS).
At first glance, this may sound like another ISO management standard. However, the underlying purpose is different.
ISO 9001 provides a general framework for quality management, including consistent delivery, customer requirements and organisational improvement. ISO/IEC 42001 establishes an AI Management System, supporting responsible AI development and use across an organisation. ISO 13485, designed for medical devices, provides a closer parallel to EN 18286 because of its emphasis on regulated products, design controls, validation and lifecycle evidence.
EN 18286 brings this product-oriented regulatory perspective to AI systems. Here, quality is fundamentally linked to regulatory conformity, including the protection of health, safety and fundamental rights.
The central question is no longer simply whether an organisation has an effective AI management system. It is also whether the provider can demonstrate that a specific AI system, in its intended context of use, complies with the applicable regulatory requirements.
This changes the object of assurance. Organisational policies and processes remain important, but compliance evidence must be traceable to individual AI systems, their requirements, testing, documentation and monitoring.
Business risk and the risk of harm are not the same thing
Perhaps the most interesting difference lies in the concept of risk.
Traditional enterprise risk management, often based on ISO 31000, defines risk as the effect of uncertainty on objectives. Organisations establish their objectives, assess potential risks and determine how much risk they are willing to accept.
The AI Act approaches risk from a different perspective. For high-risk AI systems, the focus is on potential harm, considering its probability and severity. Crucially, this includes risks to health, safety and fundamental rights.
The question is not simply how much risk a company is willing to accept in pursuit of its business objectives. It is what risks its AI systems may impose on others and whether those risks are acceptable under the applicable legal requirements.
Consider an AI-powered recruitment system. It may significantly reduce recruitment costs, accelerate candidate screening and improve process consistency. From a business perspective, it could be an outstanding success. Yet the same system might systematically disadvantage certain groups of candidates.
Business success does not make a legally unacceptable risk acceptable. A company’s risk appetite cannot override requirements concerning equal treatment or other fundamental rights.
This does not mean that ISO 42001 ignores these issues. It explicitly addresses the consequences of AI systems for individuals, groups and society, including through AI system impact assessments.
The difference is more nuanced. ISO 42001 approaches AI-related risks and opportunities within a broader organisational governance framework. The AI Act establishes a regulatory perspective in which the acceptability of risks depends on the system’s intended purpose, potential harms and applicable legal requirements.
These perspectives overlap, but they are not interchangeable.
Fundamental rights change the risk management conversation
European product safety regulations have long required providers to identify hazards, estimate the probability and severity of harm, implement protective measures and evaluate residual risks.
The AI Act introduces an additional dimension: fundamental rights. And this creates challenges that conventional product safety methodologies were not primarily designed to address.
How should we evaluate the severity of discrimination in an AI recruitment system? Can we determine an acceptable threshold for systematically excluding qualified candidates from certain demographic groups? How do we assess the risk of an AI-driven performance management system unfairly affecting promotion decisions?
These questions cannot always be answered through a conventional numerical risk matrix.
The nature of the right involved matters, as do legality, legitimate purpose, necessity and proportionality. Some rights allow restrictions under specific legal conditions. Others cannot simply be balanced against commercial advantages.
The forthcoming EN 18228 risk management standard is particularly interesting in this context. It aims to connect established product safety risk management principles with the specific challenges of fundamental-rights protection.
Although the standard is still under development, the direction is clear: identifying risks is not enough. Providers must demonstrate how those risks are addressed through system design, controls, testing and ongoing monitoring.
A risk register alone does not make an AI system safe or compliant.
From governance policies to system-specific evidence
Another important concept in the emerging AI standards is intended purpose.
An AI system cannot be meaningfully assessed without considering what it is designed to do, who uses it and under what conditions.
Consider two HR applications using similar AI technology. One helps recruiters identify relevant skills in CVs and suggests candidates for further human review. Another automatically ranks applicants and determines who progresses to the next stage.
Even if they use similar models, their intended purposes, decision-making roles and potential consequences for candidates may differ substantially. The appropriate safeguards, human oversight mechanisms and validation requirements must reflect those differences.
Turning regulatory requirements into measurable acceptance criteria therefore requires three interconnected considerations:
- Intended purpose: What does the system do, for whom, and in what operating context?
- Generally acknowledged state of the art: What technical methods, safeguards and performance levels reflect established good practice?
- Risk management: What specific hazards arise in this context, and what measures are necessary to address them?
EN 18286 provides the quality management framework within which such requirements, controls and evidence can be managed.
For each relevant AI system, providers need to connect applicable regulatory requirements to design decisions, verification and validation activities, technical documentation and post-market monitoring.
Regulatory requirement → System requirements → Controls → Verification and validation → Evidence → Ongoing monitoring.
Importantly, EN 18286 does not replace the technical standards addressing individual AI Act requirements. Rather, it provides a framework through which those requirements and their supporting evidence can be managed coherently.
Was implementing ISO 42001 a waste of money?
Absolutely not.
ISO 42001 provides a valuable organisational foundation. It establishes accountability, supports AI risk and impact assessment, defines governance processes, addresses supplier relationships and creates mechanisms for monitoring, auditing and continuous improvement.
Many of these elements can be reused when implementing EN 18286.
What ISO 42001 does not automatically provide is the complete system-specific regulatory evidence needed to demonstrate compliance with every applicable AI Act requirement.
An organisation may have an excellent AI governance policy, a functioning risk register and documented impact assessments while still lacking sufficiently detailed validation evidence or technical documentation for a particular high-risk AI system.
The practical approach is therefore integration, not replacement.
Organisations that have already implemented ISO 42001 should assess their existing processes against EN 18286, identify missing regulatory or system-specific requirements, and extend the existing framework rather than build a parallel management system.
There is also a broader reason why ISO 42001 remains relevant. Not every organisation is an AI provider, and not every AI system is high-risk. Many companies primarily deploy third-party AI systems, including in HR, finance and operational decision-making.
For these organisations, responsible AI governance extends beyond provider-specific regulatory requirements. It includes questions of business strategy, value creation, workforce impact, employee trust, accountability and responsible adoption.
Regulatory compliance is essential, but it is not the entirety of AI governance.
What comes next — and can the standards keep pace with agentic AI?
EN 18286 is only the beginning. Further European AI standards are being developed in areas including risk management, data governance, bias, logging, transparency, human oversight, accuracy, robustness and cybersecurity.
The broader programme currently includes approximately sixteen related standards and standardisation projects. February 2027 has been identified as a target for completing the remaining work, although publication and Official Journal citation are separate milestones.
Even a complete set of horizontal standards will not answer every implementation question. The appropriate testing methods and acceptance criteria for an AI recruitment system, for example, will depend on its intended purpose, operating conditions, relevant risks and the generally acknowledged state of the art.
This is where contextualisation becomes important: translating general requirements into specific, testable criteria for a particular application.
But there is another challenge that I believe deserves much more attention: the changing dynamics of AI systems themselves.
The emerging standards largely build on a product-oriented conformity model, in which an AI system has a defined intended purpose, identifiable boundaries, documented controls and a lifecycle that can be assessed and monitored. This remains a necessary foundation, but increasingly autonomous AI agents introduce a different operational reality.
Imagine an AI recruitment agent that not only screens candidates but independently interacts with external databases, sends communications, schedules interviews, updates applicant records and coordinates with other agents. Its individual functions may have been assessed, yet new risks can emerge from the sequence of autonomous actions, the permissions it holds and its interactions with other systems.
Crucially, these risks can arise without any change to the underlying model or its approved configuration. The compliance challenge is no longer limited to managing system changes. It also involves controlling what an otherwise unchanged system is authorised and able to do at runtime.
The existing and emerging standards already address relevant elements, including lifecycle management, human oversight, logging, monitoring and change control. However, it remains an open question whether these mechanisms sufficiently address dynamic authority, autonomous execution and risks arising from agent-to-agent and agent-to-system interactions.
Human oversight alone may not be effective when autonomous decisions and actions happen faster than a person can realistically review or intervene. Governance may therefore need to place greater emphasis on automated oversight, enforceable permission boundaries, runtime controls and mechanisms that prevent or interrupt unauthorised actions.
This does not make the existing standards irrelevant. Nor does it mean that every agentic AI system falls outside their scope. Rather, it highlights a potential gap between a compliance framework primarily structured around system lifecycle assurance and the realities of continuous autonomous operation.
And it raises an important question: How do we demonstrate not only that an AI system was compliant when assessed, but that its autonomous actions remain within acceptable boundaries during operation?
That question deserves a discussion of its own.
AI governance must evolve beyond compliance documentation
The arrival of EN 18286 makes an important distinction increasingly visible. Organisational AI governance and regulatory conformity are closely connected, but they answer different questions.
AI governance asks whether an organisation develops and uses AI responsibly and effectively, in line with its objectives and the interests of relevant stakeholders. Regulatory conformity asks whether individual AI systems meet applicable legal requirements and whether their providers can demonstrate that compliance.
Both perspectives are necessary. A recruitment algorithm may generate excellent business results while producing discriminatory outcomes. Equally, a fully compliant AI system may fail to deliver meaningful business value.
But agentic AI adds a third dimension: whether the organisation can maintain meaningful control over increasingly autonomous behaviour as it unfolds.
Together, these three dimensions point towards the next stage of AI governance: connecting business objectives, regulatory conformity and operational assurance in a single framework.
That means moving beyond periodic assessments and documented policies towards governance that can respond to changes in context, emerging risks and autonomous actions. Human responsibility remains essential, but effective oversight will increasingly depend on the technical ability to enforce boundaries, detect deviations and intervene when necessary.
The goal is not simply to prove that an AI system was compliant at a particular point in time. It is to ensure that AI continues to create value, respects fundamental rights and operates within acceptable boundaries throughout its lifecycle — including when it acts autonomously.
That, I believe, is where the next chapter of AI governance begins.
One management system, mapped to both standards. Build it once and keep the evidence.