Risk Classification of AI Systems

Imagine you are preparing a deployment for a customer. The product works, the model is accurate on the test data, and the customer is eager to go live. Then a legal or procurement reviewer asks a question you cannot answer: does this system fall under the EU AI Act, and if so, what must be in place before it runs? For anyone deploying AI into real organizations, that question now has a concrete answer. The answer depends on the system's risk level and on where its outputs land.

This article follows one habit of thought: before asking whether an AI system is good, accurate, or useful, ask how risky this particular use is, and what that risk level requires. We will trace that question step by step, from the Act's basic identity, through the risk sorting, to the nine areas of obligation and the cross-border question that catches many teams off guard.

Treat the date range as a planning signal rather than a single deadline. Some duties are already relevant once the Act is in force, and others are scheduled to apply later. Your job is to know which duties attach to which systems, and when each one begins to bind.

The Act does not treat every AI system the same way. It classifies AI systems by risk level, and the classification determines which obligations apply. The practical question for any system is therefore not 'Is this AI?' but 'How risky is this particular use, and what does that risk level require?'

The sorting is tied to the use of the system, not to the model alone. The same underlying model can sit in different risk positions depending on what it is used for and what its outputs affect. This is why a compliance review starts with the deployment context rather than with the model architecture.

The sorting is easiest to hold in your head as a set of tiers. The Act's risk levels run from prohibited uses at the top, through high risk and limited risk, down to minimal risk. The diagram below shows the tiers and the general character of the duties attached to each. The source pack does not list the exact tests for each tier, so use it as a map and verify the boundaries in the legal text.

Let us trace the reviewer's question through a concrete deployment. The steps below are a procedure for organizing the analysis, not a legal ruling. Each step names what you must establish before you can move on.

Once a use is identified as high risk, the Act imposes specific obligations. The source lists nine areas. Each one describes a property that the system and its surrounding processes must have, rather than a single document to file. Read the list as a set of properties to demonstrate, not as a stack of forms.

The nine areas are not independent boxes. Risk management runs across the whole life of the system and feeds the others. Data governance shapes what accuracy and robustness can realistically be. Transparency and human oversight describe the relationship between the system and the people who use it or are affected by it. Accuracy, robustness, and cybersecurity describe how the system performs and how well it is protected. Technical documentation and record-keeping form the evidence trail that lets an organization show how it met the other seven.

A common assumption is that a regulation only binds companies located in the regulating region. The EU AI Act has extraterritorial reach, meaning it can apply to parties outside the EU. The source gives a specific case: a US vendor deploying for a US customer can still be subject to the Act if the customer's outputs affect EU residents.

The flowchart is a simplification. The source pack identifies the outputs reaching EU residents as the case it describes, and it does not enumerate every other condition. A 'no' answer to the outputs question is therefore a reason to confirm the position against the official text, not a guarantee that the Act is out of scope. In a customer conversation with a multinational organization, add a direct question: who do the system's outputs reach? That answer may bring the Act into scope even when both parties are American.

The Act entered into force on August 1, 2024. Its obligations then phase in over a series of stages running through 2026 and 2027. The diagram shows this sequence. The source pack does not give the individual stage dates, so consult the official schedule for the exact date each obligation becomes enforceable.