EU AI Act: How to Tell If Your AI Is High-Risk

Most companies building or deploying AI systems right now are asking the wrong question. They're asking "is our AI dangerous?" when the EU AI Act is asking something more precise—and, frankly, more interesting: "does your AI system operate in a context where getting it wrong could harm real people in legally defined ways?"

That distinction matters enormously. It's the difference between a compliance exercise you can delegate to legal and a fundamental rethink of how you document, deploy, and govern your technology.

The Architecture of Risk

The EU AI Act doesn't try to classify AI by capability or sophistication. It classifies by context and consequence. The European Commission's digital strategy framework makes this explicit: AI use cases that can pose serious risks to health, safety, or fundamental rights are classified as high-risk. Not AI systems that could pose such risks in theory—use cases where the deployment context makes those risks foreseeable.

This is a meaningful design choice. A facial recognition model isn't inherently high-risk under the Act. The same model used to make decisions about who gets a mortgage, or to verify identity at a border crossing, is a different matter entirely.

The mechanics of that classification run through Article 6, which the AI Act Service Desk describes as operating through two distinct criteria: first, whether the AI system serves as a safety component in a product covered by specific EU laws listed in Annex I and must pass third-party conformity assessment; second—and this is where many organizations will find themselves unexpectedly implicated—whether the system falls under one of the use cases listed in Annex III.

Annex III is where the practical weight of the regulation lands.

What's Actually on the List

The Annex III categories cover terrain that will feel familiar to anyone watching how AI has already been deployed across industries: biometric identification and categorization, critical infrastructure management, education and vocational training, employment decisions, essential private and public services, law enforcement, migration and border control, administration of justice, and democratic processes.

The breadth is deliberate. Europe watched what happened when algorithmic systems made consequential decisions about people—the Dutch childcare benefits scandal, the UK's exam grade algorithm debacle during the pandemic, the documented biases in US recidivism prediction tools—and built a category specifically designed to catch the systems most likely to replicate that pattern.

But there's a significant carve-out that companies should understand. According to WilmerHale's analysis of the Act, AI systems identified in Annex III will not be considered high-risk if they don't pose a significant risk of harm to individuals' health, safety, or fundamental rights—including by not materially influencing the outcome of decision-making. That last phrase is load-bearing. An AI tool that helps a human make a decision is treated differently from one whose output is the decision, or effectively predetermines it.

The high-level summary at artificialintelligenceact.eu adds another layer: systems listed under Annex III are always considered high-risk if they profile individuals. Profile, not just analyze. The distinction points at systems that build persistent behavioral or characteristic models of specific people—the kind of architecture that underlies much of the adtech and HR tech ecosystem.

The Part Everyone Gets Wrong

Here's where a lot of organizations are going to miscalculate: they're thinking about high-risk classification as a question about their model. It's actually a question about their system—meaning the full stack of how the model is integrated, surfaced, and used.

KDnuggets puts this cleanly: how a system is documented, marketed, deployed, and used can be just as important as its technical capabilities. The same underlying model can sit in a low-risk or high-risk classification depending on what it's connected to and how its outputs are used downstream.

Consider an AI resume screening tool. The model might be a fairly standard classification system, not meaningfully more sophisticated than tools used for spam filtering. But if it's deployed in a hiring workflow where its ranking effectively determines which candidates get interviews, it's operating in the employment domain covered by Annex III, and its outputs are materially influencing decisions about people's economic lives. That's a different compliance posture than the same model deployed to sort internal documents.

This is a feature of the regulation, not a bug. The EU's architects were specifically trying to catch cases where organizations claim a human is "in the loop" while the human is really just rubber-stamping an algorithmic recommendation. The materiality test is designed to pierce that.

What High-Risk Actually Requires

If you land in the high-risk category, the compliance burden is substantial. The framework requires robust data governance, detailed technical documentation, automatic logging and record-keeping, transparency to users, human oversight mechanisms, and demonstrated accuracy and cybersecurity standards. There's also a conformity assessment process—for some categories, that assessment must be conducted by an independent third party.

The Commission has committed to publishing practical guidance to help organizations work through these determinations—a resource that will matter considerably given how much interpretive work the current framework leaves to deployers.

The reason the compliance requirements are this extensive is that high-risk classification is meant to track situations where getting it wrong has consequences that can't easily be undone. A biased content recommendation is annoying. A biased tool that helps determine who gets access to credit, employment, or social services can define the contours of a person's life.

The Harder Question

The regulation's architecture reflects a coherent theory: that risk in AI is primarily about context, power, and consequence, not raw capability. That's probably the right theory. But it also creates genuine ambiguity at the edges that organizations are going to have to navigate without perfect information.

The open question isn't really whether companies will comply with the clear-cut cases. It's how the edge cases get resolved—who gets to make the materiality determination, whether self-assessment is credible in high-stakes domains, and whether the enforcement infrastructure will have the resources to actually audit the systems that get classified as low-risk but probably shouldn't be.

The Act creates the categories. What happens inside them will depend heavily on how seriously the institutions implementing it treat the spirit of the thing, not just the letter.

---

Marcus Chen-Ramirez is a senior technology correspondent for Buzzrag covering AI, software development, and the intersection of technology and society.