AI Act High-Risk Self-Check: Is Your AI Really High-Risk?

In almost every first conversation, the same worry lands on the table: "Lars, we now use AI in quality control and run a chatbot in customer service. Does that make us high-risk, and do we have to spin up the entire compliance machinery?" My answer is reassuring almost every time: as a rule, no. High-risk is not the default case but a clearly defined exception.

The AI Act thinks in four risk classes

The AI Regulation (Regulation (EU) 2024/1689) follows a risk-based approach. It does not regulate "AI" across the board, but instead asks: what could this application do in the worst case? Four levels follow from that.

• Prohibited practices (Art. 5): a small number of clearly named applications such as social scoring, manipulative behaviour, or certain biometric practices. These are banned.
• High-risk AI: narrowly defined fields of use carrying significant obligations. This is where the real focus of the regulation lies.
• Limited risk (Art. 50): transparency obligations. Anyone running a chatbot or generating AI-produced content and deepfakes must, as a rule, label it.
• Minimal risk: everything else. No specific obligations under the AI Act.

The decisive point for mid-sized companies: the vast majority of everyday AI applications land in the lower two levels. High-risk is something you essentially have to "earn", and only via one of two clearly described routes.

High-risk only arises via two routes

An AI is not high-risk because it is "important" or "powerful". It is high-risk only if it hits one of the two routes laid down in law. Anyone who knows both can roughly classify their own application in a few minutes.

Route A: Annex III - the eight areas of use

Annex III lists eight areas in which AI is deemed high-risk because it directly affects people's fundamental rights, safety, or life chances. Check honestly whether your application falls into one of them:

• Biometrics (e.g. biometric identification)
• Critical infrastructure (e.g. control of electricity, water, or transport networks)
• Education and vocational training (e.g. grading of exam performance, access to education)
• Employment and workforce management (e.g. AI for candidate selection or employee evaluation)
• Access to essential private and public services (e.g. creditworthiness assessment, risk assessment in life or health insurance)
• Law enforcement
• Migration, asylum, and border control
• Administration of justice and democratic processes

Don't see yourself here? Then Route A is, as a rule, already settled for you. This is exactly where the all-clear comes for many industrial and service businesses: a machine that detects defective parts via camera, or a model that forecasts material demand, belongs to none of these eight areas.

Route B: Annex I - AI as a safety component in products

The second route concerns AI that sits as a safety component in an already regulated product subject to a conformity assessment. Think of machinery, medical devices, lifts, or toys. When the AI takes on safety-relevant functions there, it is captured through the respective product regulation.

For many manufacturers this is familiar territory: they already go through a conformity assessment anyway. The AI Act hooks into this existing logic rather than creating an entirely new world. For pure deployers who merely use such products, Route B is usually not relevant.

High-risk is not a gut feeling. Either your AI hits one of the eight Annex III areas, or it sits as a safety component in a regulated product under Annex I. Neither applies? Then, as a rule, it is not high-risk.

The most important exception: Annex III does not automatically mean high-risk

Now comes the rule that brings the most relief in practice and that many overlook. Even when an AI nominally falls into an Annex III area, it can be classified as not high-risk. The regulation provides for this exception in Art. 6(3).

It applies when the AI performs only a preparatory or narrowly limited supporting task and does not pose a significant risk to health, safety, or fundamental rights. A system that, for instance, merely pre-sorts incoming documents or provides a purely supportive secondary function without making the actual decision can fall out of the high-risk class despite its Annex III link.

Important: you must be able to justify and document this classification cleanly. It is not a free pass but a deliberate, traceable assessment. This is precisely the transition from the question "Are we high-risk?" to the question "Can we prove our answer?"

Typical mid-market AI - and where it really lands

Let's look at the applications I encounter most often. The classification almost always turns out reassuring.

• Predictive maintenance: as a rule, minimal risk.
• Camera-based quality control and image processing in manufacturing: as a rule, minimal risk.
• Demand and requirements forecasting: as a rule, minimal risk.
• Internal chatbot or customer-service chatbot: usually limited risk, i.e. a transparency or labelling obligation, not high-risk.
• Text generation and AI assistants: usually limited risk with labelling of AI-generated content.

It only becomes high-risk once the same technology moves into one of the sensitive areas. The same text analysis is harmless when it sorts service tickets, and potentially high-risk when it decides on job applications. It is not the technology that matters, but the purpose and context.

What applies when? The deadlines at a glance

The AI Act does not enter into force all at once but in stages. That gives you time to sort things out rather than fall into a panic.

• 02.02.2025: the prohibitions (Art. 5) and the obligation on AI literacy (Art. 4) apply.
• 02.08.2025: obligations for general-purpose AI models (GPAI) plus the governance and authority structure.
• 02.08.2026: high-risk obligations under Annex III.
• 02.08.2027: high-risk obligations for products under Annex I.

Note: the obligation on AI literacy under Art. 4 applies to everyone who deploys AI, regardless of risk class. Your staff should understand what they are working with. That is not a high-risk question but common sense with a legal basis.

The self-check in four questions

If you need a quick initial assessment for a specific application, work through these four questions in order. They do not replace a full legal evaluation of the individual case, but they sort most cases cleanly in advance.

• 1. Is it a prohibited practice under Art. 5 (e.g. social scoring)? If yes, the application is banned. For typical mid-market AI, almost never the case.
• 2. Does the application fall into one of the eight Annex III areas, or is it a safety component under Annex I? If no, it is, as a rule, not high-risk.
• 3. If Annex III: does the exception under Art. 6(3) apply because the AI performs only a preparatory or narrowly limited task? Then it can still fall out of high-risk, properly documented.
• 4. Otherwise check: do you need transparency under Art. 50 (chatbot, labelling of AI content)? If yes, limited risk. If no, minimal risk.

The right question is not "Are we using AI?" but "For what exactly, and with what effect on people?" From that, the risk class follows almost by itself.

Why an AI Management System makes the difference here

Regardless of which class your AI falls into, you need a solid answer when a customer, an insurer, or an authority asks. That is exactly what ISO/IEC 42001 is built for. The standard gives you a structure to systematically capture, assess, and govern your AI applications.

The self-check above is the snapshot. An AI management system to ISO/IEC 42001 turns it into a durable, traceable process: you document why an application falls into which class, keep it up to date, and can present it on request. In practice this is often worth more than the class itself, because it shows that AI is being deployed here deliberately and under control.

A note on classification: this overview does not replace a legal review of the individual case. It helps you ask the right questions and approach the topic with a clear head rather than a knot in your stomach. In the vast majority of mid-market cases, the honest answer in the end is: not high-risk, but good that you took a closer look.