How to Classify Your First AI System Under the EU AI Act in 15 Minutes

What you’ll achieve in the next 15 minutes

By the end of this quick exercise, you’ll have:

• A working risk-tier classification for your AI system under the EU AI Act
• A short, auditable record of why you chose that tier
• A clear next step list for compliance actions appropriate to the tier

This guide is designed for professionals doing this for the first time in a compliance workspace—practical, fast, and repeatable.

---

Before you start: collect 5 inputs (2 minutes)

You don’t need a perfect dossier to classify a system the first time, but you do need the basics. Gather:

1. System purpose (one sentence): what decision or output it produces.
2. Where it’s used: internal tool, customer-facing product, public service, etc.
3. Who is affected: employees, customers, patients, students, general public.
4. What it outputs: recommendations, classifications, scores, content generation, biometrics, predictions.
5. Whether outputs can materially impact rights/opportunities: access to services, employment, education, credit, healthcare, law enforcement outcomes, or safety.

Write these in your workspace as a short “System Snapshot.” You’ll refer to it repeatedly.

---

Step 1: Confirm you’re classifying an “AI system” (2 minutes)

Classification starts by ensuring your tool is in scope.

Ask:

• Does it infer outputs (predictions, content, recommendations, classifications, decisions) from inputs?
• Does it use statistical/ML methods, logic-based approaches, or other computational techniques to produce those outputs?

If the answer is “yes” in practical terms, proceed. If it’s purely deterministic business logic (e.g., fixed rules with no inference), it may not be an AI system for EU AI Act purposes—but still document the rationale.

Workspace tip: Create a checkbox: In scope as AI system: Yes/No + one-line justification.

---

Step 2: Screen for “Unacceptable Risk” first (3 minutes)

Start at the top. If your system matches prohibited practices, it’s not a matter of “high vs limited risk”—it’s fundamentally non-compliant.

In your workspace, run a quick red-flag check:

• Does it manipulate people in a way likely to cause significant harm (e.g., exploit vulnerabilities, materially distort behavior)?
• Does it exploit vulnerabilities of specific groups (age, disability, socio-economic situation) in a harmful way?
• Does it involve social scoring by public authorities or on their behalf (broadly scoring trustworthiness leading to unjustified detrimental treatment)?
• Does it use biometric categorisation based on sensitive traits (depending on the use case and specifics)?
• Does it enable certain types of biometric identification in publicly accessible spaces for law enforcement, beyond narrow exceptions?

If any answer is “yes” or “uncertain,” stop and escalate to legal/compliance review. For first-time classifiers, uncertainty here is a signal to slow down.

Workspace output:

• Unacceptable risk screen: Pass/Fail/Needs review
• Notes on which red flag triggered and why

---

Step 3: Check if it’s “High-Risk” (the main decision) (5 minutes)

Most real-world classification effort goes here. A practical approach is to test two routes:

Route A: Is it a product-level regulated use?

If your AI is a safety component of a product or used in a regulated product context, it may qualify as high-risk. Even if you’re unsure, note this route and seek confirmation.

Workspace prompt: “Is this AI a safety component or part of a regulated product/service context?”

• Answer: Yes/No/Unsure
• Explanation: one sentence

Route B: Is it used in a high-impact area?

Ask whether your AI is used to make—or meaningfully influence—decisions in areas that affect people’s rights or life chances. Examples of typical high-impact domains include:

• Employment and worker management (hiring, firing, task allocation, performance evaluation)
• Education and training (admissions, assessment, proctoring decisions)
• Access to essential private/public services (creditworthiness, insurance eligibility, welfare decisions)
• Healthcare (triage support, diagnostic support, treatment prioritization)
• Migration, asylum, border control
• Law enforcement and public security
• Administration of justice and democratic processes (depending on the specific function)

You’re not trying to be exhaustive—you’re determining whether the system’s function could materially affect outcomes in one of these areas.

Two questions that settle most borderline cases:

1. Is the output used to make or strongly influence a decision about a person?
2. Could an error plausibly cause significant harm or denial of opportunity?

If both are “yes,” treat the system as high-risk unless you have a clear reason not to.

Workspace output:

• High-risk assessment: Yes/No/Unsure
• Domain: pick from a list
• “Decision impact” note: 2–3 bullet points on how the output is used

---

Step 4: If not high-risk, decide between “Limited Risk” and “Minimal Risk” (3 minutes)

If the system is not prohibited and not high-risk, you’ll usually land in one of these two tiers.

Limited Risk (transparency obligations likely)

These are systems where the key compliance focus is transparency—making sure people know they’re interacting with AI or that content is AI-generated/manipulated where relevant.

Common indicators:

• A chatbot or conversational agent interacting with users
• An AI system generating content that could be mistaken for human-made
• Systems that influence user behavior but don’t decide access to rights or essential services

Workspace decision:

• “Does a user need to be informed they’re interacting with AI or seeing AI-generated/manipulated content?”
• If yes → likely limited risk (and document the transparency measure you’ll implement).

Minimal Risk

Most internal productivity tools, benign recommendation engines, or non-sensitive analytics will fall here—where harms are low and the law’s main requirements are not triggered beyond general good practice.

Workspace output:

• Risk tier: Limited or Minimal
• Rationale: one paragraph

---

Step 5: Create a “Classification Record” you can defend (2 minutes)

In a compliance workspace, the classification is only as useful as the record behind it. Capture:

• System Snapshot (from the start)
• Final tier: Unacceptable / High / Limited / Minimal
• Reasoning trail: the answers to Steps 2–4
• Assumptions and unknowns (be explicit)
• Owner and date (who made the call, when)
• Next action list (what must happen now)

Keep it short but complete. Aim for something a colleague can review in under five minutes.

---

What to do next (based on your tier)

Use your tier to trigger the right workflow immediately.

If High-Risk

Prioritize a structured compliance plan. Your next actions typically include:

• Define intended purpose, users, and use environment
• Establish risk management activities and acceptance criteria
• Map data governance (data quality, representativeness, bias risks, provenance)
• Document model development and system design choices
• Ensure human oversight measures are real and operational
• Prepare for logging, monitoring, and incident handling
• Align instructions for use and user documentation with how the system is deployed

Workspace tip: Create a checklist template so every high-risk project starts with the same control set.

If Limited Risk

Focus on transparency and user-facing controls:

• Add clear AI interaction disclosures (when users converse with AI)
• Label AI-generated or AI-manipulated content where required and relevant
• Provide simple user guidance on appropriate use and limitations

If Minimal Risk

Adopt good governance without overbuilding:

• Record the classification and basic rationale
• Do lightweight monitoring for drift or misuse
• Establish a point of contact for incidents or user complaints

If Unacceptable Risk or “Needs Review”

Stop deployment planning and escalate:

• Document the exact feature triggering concern
• Propose alternatives (feature removal, scope limitation, different use case)
• Seek formal legal/compliance decision

---

Common classification mistakes (and how to avoid them)

• Mistake: Classifying by model type instead of use case.
  Fix: Classify based on what the system does and where it’s used, not whether it’s “genAI,” “ML,” or “rules-based.”

• Mistake: Ignoring downstream use.
  Fix: Document how outputs are acted upon. A “recommendation” can become a decision if it’s routinely followed.

• Mistake: Underestimating impact on individuals.
  Fix: Use the two settling questions from Step 3 (decision influence + plausible harm).

• Mistake: No record of assumptions.
  Fix: Add an “Unknowns” section. Unknowns are acceptable; hidden unknowns are not.

---

A 15-minute checklist you can reuse

• [ ] Write System Snapshot (purpose, users, affected people, outputs, impacts)
• [ ] Confirm in-scope AI system
• [ ] Run Unacceptable Risk screen
• [ ] Run High-Risk routes (regulated product context + high-impact domain test)
• [ ] If not high-risk, decide Limited vs Minimal
• [ ] Save a one-page Classification Record with owner, date, assumptions
• [ ] Trigger the next-step workflow based on tier

Done well, your first classification won’t just give you a label—it will create a defensible trail and a clear compliance pathway you can repeat for every new AI system.