What "High-Risk AI" Actually Means
Most explanations of the EU AI Act list its four risk tiers and leave it there. That isn't useful when your compliance team is staring at a product roadmap trying to figure out which items require legal review. Here is what the framework actually means in practice.
The EU AI Act takes a tiered approach to regulating AI based on the level of risk it poses, ensuring stricter controls for potentially dangerous AI while allowing innovation to flourish in lower-risk areas. The four tiers are: unacceptable risk (prohibited outright), high risk (heavily regulated), limited risk (transparency obligations only), and minimal risk (essentially unrestricted). The critical question for most software teams is not whether they're building prohibited AI — that's usually obvious — but whether they're building high-risk AI without knowing it.
High-risk status is not self-declared. An AI system is considered high-risk under the AI Act if it meets one of these criteria: it serves as a safety component or is a product covered by specific EU laws in Annex I and must pass a third-party conformity assessment; or it is listed in Annex III. Annex III is the list that will catch most commercial software teams off guard, because it defines high-risk not by technical sophistication but by deployment context and decision-making impact.
Understanding the full EU AI Act compliance requirements gives essential background on how the risk tiers interact before walking through the classification logic below.
Step 1: Is Your System an "AI System" at All?
Before asking whether your software is high-risk, you need to confirm it qualifies as an AI system under the Act's definition. This matters because traditional rule-based software — decision trees with hard-coded logic, simple IF/THEN rule engines — may fall outside scope entirely.
The Act defines an AI system as a machine-based system designed to operate with varying levels of autonomy that, based on the inputs it receives, infers from those inputs how to generate outputs such as predictions, content, recommendations, or decisions that can influence real or virtual environments. The key words are infers and varying levels of autonomy. A system that applies fixed programmatic rules set by a human does not infer — it executes. A system that uses machine learning, neural networks, or statistical inference to generate outputs does.
In practice: a fraud detection engine that uses logistic regression on transaction features is an AI system. A rules engine that flags transactions over €10,000 from new accounts is probably not. Most modern AI features — ML-based scoring, LLM-powered recommendations, computer vision classifiers — will qualify.
Step 2: Does It Fall Under Prohibited Practices?
Before getting to high-risk classification, check whether your system might be prohibited under Article 5. Prohibited AI cannot be made compliant through documentation or oversight — it must be removed from the market.
Prohibited categories include social scoring systems that evaluate individuals based on behaviour or characteristics to determine access to services or opportunities, real-time remote biometric identification in public spaces (with narrow law enforcement exceptions), AI systems that exploit psychological vulnerabilities to manipulate behaviour, and emotion recognition in workplace and educational settings.
The EU AI Act introduces outright bans on emotion recognition in workplaces and schools, fundamentally reshaping how these technologies can be deployed. If your product roadmap includes any of these features targeting EU users, the conversation is about product architecture, not compliance controls.
Step 3: Does It Match an Annex III Use Case?
This is the classification decision that will affect the most companies. Annex III lists eight key use cases for AI systems that are considered high-risk if they pose a significant risk of harm to the health, safety or fundamental rights of natural persons. These are: biometric identification and categorisation; critical infrastructure management; education and vocational training; employment and worker management; access to essential services; law enforcement; migration and border control; and administration of justice and democratic processes.
The classification logic is use-case-first, not technology-first. A CV parsing model sounds benign. When it ranks candidates and that ranking influences hiring decisions for EU-based employees, it is Annex III high-risk. The classification question is: what decision does this system inform or make, and in which domain?
AI systems listed under Annex III are always considered high-risk if they profile individuals — meaning automated processing of personal data to assess various aspects of a person's life, such as work performance, economic situation, health, preferences, interests, reliability, behaviour, location or movement. This is the clause that pulls in fraud scoring, credit assessment, insurance pricing models, and a wide range of analytics products whose providers have not thought of themselves as being in regulated territory.
One important escape hatch exists. By derogation from paragraph 2, an AI system referred to in Annex III shall not be considered high-risk where it does not pose a significant risk of harm to the health, safety or fundamental rights of natural persons, including by not materially influencing the outcome of decision making. A provider who wants to claim this derogation must document the assessment before market entry and register the system — a paper trail regulators can audit.
Step 4: Is It a Safety Component in a Regulated Product?
The second pathway to high-risk classification catches AI embedded in physical products. An AI system is high-risk if it is a safety component of a product that already falls under EU sector rules (like the Machinery Directive, Medical Devices Regulation, etc.) and thus must undergo third-party conformity assessment. This affects AI in medical devices, autonomous vehicles, aviation systems, industrial machinery, and railway infrastructure. If you are building AI that goes inside hardware governed by CE marking requirements, you are almost certainly building high-risk AI.
Real Examples: Where Most Software Teams Land
AI hiring or recruitment tool. This is the clearest high-risk classification in commercial software. CV screening, candidate ranking, interview scoring, and even tools that merely surface candidates for human review all fall under Annex III's employment category. If your business uses AI to screen, rank, or match candidates, the EU now regulates those tools as high-risk systems. There is no size or revenue threshold — a ten-person startup using an AI-powered ATS to screen EU applicants is in scope.
Credit scoring or insurance pricing model. Access to essential services — AI for social benefits, credit scoring, insurance, healthcare triage — is in scope under Annex III. A SaaS fintech that offers an AI-powered creditworthiness API to European lenders is a provider of a high-risk AI system, even if the lender is the one making the final credit decision.
General-purpose chatbot or LLM assistant. A customer support chatbot, a writing assistant, or an internal knowledge tool is generally not high-risk unless it is deployed in a high-risk context — such as providing medical triage, legal guidance in enforcement settings, or making eligibility determinations for public services. The system's deployment context matters as much as its technical architecture. AI systems that interact with humans (chatbots, virtual assistants) carry transparency requirements but don't face the full compliance burden of high-risk systems. The transparency requirement means users must be told they are interacting with an AI, but this is far less onerous than the full Annex III obligation stack.
SaaS analytics platform. A business intelligence platform that surfaces trends and correlations for human analysts to act on is typically not high-risk, provided it doesn't profile individuals in Annex III domains. Where analytics platforms get pulled in is when they segment, score, or rank individuals in ways that feed employment, credit, or benefits decisions — even indirectly.
Fraud detection system. This sits in a genuine grey zone. AI systems detecting financial fraud are specifically excluded from the "access to essential services" Annex III category. However, if the same model is used not just for fraud detection but to assess creditworthiness or determine account access, the scope expands. The use case determines the classification, and multi-purpose models may carry multi-tier obligations.
For a deeper look at the technical obligations that flow from these classifications, the EU AI Act for CTOs guide translates the compliance requirements into engineering terms.
The Grey Areas: GPAI, APIs, and Dual-Use Systems
General-purpose AI models — foundation models like large language models — have their own compliance track under the Act. They are not classified as high-risk by default, but when a GPAI model is integrated into a high-risk AI system, the general-purpose AI model integrated into that high-risk system is subject to both high-risk AI system requirements and general-purpose AI model requirements. A company deploying GPT-4 or Claude as the engine behind an AI recruitment tool carries obligations for both the GPAI layer and the high-risk system layer.
APIs present a related classification challenge. If you provide an AI API that clients can deploy in any context, your classification depends on what clients actually do with it. The Act distinguishes between providers (who build and place AI systems on the market) and deployers (who use them in specific contexts). An API provider is likely a provider under the Act; the deployer bears separate obligations for the context they create. Contracts and terms of service that restrict high-risk deployment contexts are not a legal shield — the Act imposes obligations on the chain, not just the end user.
The overlap between AI regulation and data protection law creates an additional layer of exposure. With GDPR penalties reaching up to 4% of global annual revenue or €20 million, and the EU AI Act introducing additional enforcement mechanisms, the financial implications of non-compliance are substantial. AI systems that process personal data to inform high-risk decisions sit simultaneously under GDPR's Article 22 automated decision-making restrictions and the AI Act's high-risk obligations. A single product can trigger both regimes. The intersection of AI and GDPR compliance is one of the more complex areas where organisations currently have the least mature governance.
