High-risk AI is the central category of the EU AI Act and the one that carries the most demanding obligations short of an outright ban. A high-risk AI system is one the regulation considers capable of causing significant harm to the health, safety, or fundamental rights of people. Such systems are not prohibited — they are permitted, but only if they meet a comprehensive set of requirements designed to make them safe, transparent, and accountable. For most organisations, the question of whether their AI is high-risk is the single most consequential point in an AI Act assessment.
How the AI Act identifies high-risk systems
The Act takes two routes to the high-risk label. The first covers AI used as a safety component of a product, or as a product itself, that is already regulated under existing EU product-safety legislation — think medical devices, machinery, or vehicles. The second, and the one most organisations grapple with, is a list of specific use cases set out in Annex III: AI used in areas such as critical infrastructure, education and vocational training, employment and worker management, access to essential private and public services, law enforcement, migration and border control, and the administration of justice and democratic processes. If a system serves one of these listed purposes, it is presumptively high-risk.
The Annex III domains
Annex III is worth understanding in concrete terms because it is where many ordinary business applications unexpectedly land. AI that screens job applications or evaluates employees falls under employment. AI that scores creditworthiness or allocates public benefits falls under access to essential services. AI used in admissions or grading falls under education. Many companies that would never describe themselves as operating “high-risk AI” discover that a hiring tool or a credit model places them squarely in the category. This is why a careful mapping of each system against the Annex III list is indispensable.
The requirements for high-risk systems
High-risk systems must satisfy a connected set of requirements throughout their lifecycle. A risk-management system must run continuously to identify and mitigate risks. Data governance must ensure training, validation, and testing data are relevant, representative, and as free of errors as possible. Detailed technical documentation must demonstrate compliance, and the system must keep automatic logs of its operation. It must be transparent enough that deployers can interpret and use its output appropriately, must allow for effective human oversight, and must achieve an appropriate level of accuracy, robustness, and cybersecurity. Together these requirements aim to make high-risk AI trustworthy by design rather than by afterthought.
Conformity assessment and CE marking
Before a high-risk system can be placed on the EU market, it must undergo a conformity assessment — a procedure that verifies it meets the Act’s requirements. For many systems this is a self-assessment by the provider against the requirements; for some it involves a notified body. A system that passes is documented in an EU declaration of conformity and may bear the CE marking, signalling that it complies. The provider must also register the system in an EU database. This pre-market gate is a defining feature of the high-risk regime and means compliance work has to be substantially complete before launch, not after.
Obligations of providers and deployers
The heaviest obligations fall on the provider — the party that develops the high-risk system and places it on the market. Providers must build and maintain the quality and risk-management systems, prepare the documentation, run the conformity assessment, and monitor the system after it is on the market. Deployers — organisations that use a high-risk system professionally — carry lighter but real duties: they must use the system in line with its instructions, ensure meaningful human oversight, monitor its operation, and keep the logs it generates. Where a deployer significantly modifies a system or markets it under its own name, it can itself become a provider, inheriting the heavier obligations.
Human oversight
A distinctive requirement of the high-risk regime is meaningful human oversight. The Act insists that high-risk systems be designed so that natural persons can effectively oversee them — understanding the system’s capabilities and limits, remaining alert to the risk of over-relying on its output, being able to interpret its results correctly, and being able to intervene or halt the system. The goal is to keep a human meaningfully in control of consequential decisions rather than letting an automated output stand unchallenged. Designing for genuine oversight, rather than a token “human in the loop”, is one of the more demanding aspects of compliance.
Timeline and preparation
The high-risk obligations are among the later provisions of the AI Act to apply, with the headline date for many high-risk systems falling in 2026 and certain product-safety-linked cases extending into 2027. That staggered timing is not a reason to wait: building a risk-management system, assembling technical documentation, establishing data-governance practices, and arranging a conformity assessment all take time. Organisations that begin their gap analysis early give themselves room to remediate; those that start late risk having to choose between a delayed launch and a non-compliant one.
What organisations should do
The practical path begins with classification: inventory every AI system and map each against the high-risk criteria and the Annex III list. For any system that qualifies, the next step is a structured gap analysis against the full set of requirements, followed by a remediation plan and, ultimately, the conformity assessment. Even where a system turns out not to be high-risk, documenting why is valuable evidence of diligence. Innopulse supports this process with its AI Risk Check tool, which walks organisations through the classification and shows, use case by use case, whether and to what degree the high-risk obligations apply.
Conclusion
High-risk AI is the core of the EU AI Act: a category of permitted but heavily regulated systems whose potential to affect health, safety, and fundamental rights justifies extensive requirements around risk management, data, documentation, transparency, human oversight, and robustness, enforced through a pre-market conformity assessment. Because so many ordinary applications — hiring tools, credit models, education systems — fall into the Annex III domains, careful classification is essential, and because the obligations are substantial, early preparation ahead of the 2026 timeline is the only prudent approach.