In this article7 sections
You already use AI systems that need a risk tier assigned
Classifying your AI tools is the first step to knowing which duties apply. The AI Act has applied in general since 2 August 2026, and the high-risk rules apply from 2 December 2027 for Annex III systems and 2 August 2028 for AI in products covered by Annex I. Under Article 6 of the EU AI Act, the risk tier of a system determines which obligations apply to you as a deployer. Once the high-risk rules apply, breaching deployer duties can be fined up to €15 million or 3% of global annual turnover under Article 99; for SMEs and start-ups the lower of the two applies.
The good news: for most SMEs with 20-200 employees, the classification step itself is not complicated. Eight questions. One decision tree. Thirty minutes.
This post walks you through exactly that.
Why risk classification comes first
The EU AI Act splits AI systems into four tiers: prohibited, high-risk, limited-risk, and minimal-risk. Your obligations as a deployer scale with that tier. One duty applies to every tier: since 2 February 2025, Article 4 requires you to take measures to support the AI literacy of staff who use AI. A chatbot on your website needs little beyond that and a transparency notice. A high-risk CV-screening tool triggers Article 26 obligations from 2 December 2027: human oversight by competent people, keeping logs, monitoring, and informing workers and candidates.
Before you can build a compliance programme, you need a verified inventory of what tier each tool sits in. That is the ai risk classification self assessment.
The 8-question decision tree
Work through these questions in order for each AI system your company uses. Stop at the first "yes" that determines a tier.
Question 1 — Is the system on the prohibited list?
Check Article 5 of the AI Act. Prohibited practices include social scoring (by anyone), emotion recognition at work or in education, real-time remote biometric identification in public spaces for law enforcement, and manipulative techniques or AI that exploits vulnerabilities to distort behaviour. From 2 December 2026 the list also covers AI that creates non-consensual intimate deepfakes or child sexual abuse material. If your system matches any Article 5 description, stop here. You cannot deploy it.
Tier result: Prohibited.
Question 2 — Is it a General Purpose AI (GPAI) model used as a standalone product?
GPAI model rules (Chapter V, including the systemic-risk rules in Article 51) apply to the companies that develop and release the model, such as OpenAI, Google or Mistral, not to you. If you build your own product on a model API, you may become the provider of that AI system, so run that system through the questions below. If you use a GPAI-powered product built by a vendor, continue to Question 3.
Question 3 — Does it fall under Annex III?
Annex III lists eight areas of use. AI systems used for these purposes are high-risk, subject to the narrow exception in Question 4. Read each category carefully:
-
Biometric identification and categorisation
-
Critical infrastructure (energy, water, transport)
-
Education and vocational training (e.g. scoring students, admission decisions)
-
Employment, workforce management, and access to self-employment (this is the big one for HR teams: CV screening, promotion algorithms, performance monitoring)
-
Essential private and public services (creditworthiness, insurance risk)
-
Law enforcement
-
Migration, asylum, border management
-
Administration of justice and democratic processes
If your system makes or substantially influences a decision in any of these categories, answer is yes.
Tier result: High-risk (rules apply from 2 December 2027). Move to Question 4 to check for a narrow exemption.
Question 4 — Does the narrow Annex III exemption apply?
Under Article 6(3), a system listed in Annex III is not high-risk if it does not pose a significant risk of harm to health, safety or fundamental rights, including by not materially influencing the outcome of decisions, and it meets at least one of four conditions:
-
it performs a narrow procedural task;
-
it improves the result of a human activity that has already been completed;
-
it detects decision-making patterns or deviations without replacing or influencing the human assessment, without proper human review;
-
it performs a preparatory task to an assessment relevant to an Annex III use.
One rule overrides all four: an Annex III system that profiles natural persons is always high-risk. The provider that relies on the exception must document its assessment (Art 6(4)) and register the system (Art 49(2)). This exemption is genuinely narrow. A tool that ranks job candidates profiles them, so it is not exempt simply because a human clicks "approve" at the end.
If the exemption does not clearly apply, keep the high-risk classification.
Question 5 — Does it interact directly with people using text, voice, or generated images?
If yes, Article 50 transparency obligations apply. Users must be told they are interacting with an AI. Deepfakes and AI-generated content must be labelled. This tier is sometimes called "limited-risk" or "transparency-risk" in practice.
Tier result: Limited-risk (transparency obligations).
Question 6 — Is any biometric data processed, even temporarily?
Face recognition for timekeeping or voice-print analysis triggers both AI Act obligations and AVG/GDPR requirements. Under GDPR Article 9, biometric data used to identify someone is special-category data, and a Data Protection Impact Assessment (DPIA) will usually be required. Flag this system for dual-track review. Emotion detection in video interviews is a different matter: emotion recognition at work is prohibited under Article 5(1)(f), except for medical or safety reasons. Stop using it.
Question 7 — Does the system generate or process content that could influence public opinion at scale?
For most SMEs the answer is no. But if you run a media company, an edtech platform, or a platform that recommends content to large audiences, re-check Article 50 and any applicable Digital Services Act requirements.
Question 8 — None of the above?
If you have reached Question 8 without triggering another tier, the system is minimal-risk. The only AI Act duty left is Article 4: take measures to support the AI literacy of staff who use it. Keeping a record of what you use and why is still sensible.
Tier result: Minimal-risk.
Filling in your classification register
After running each system through the eight questions, record the result in a simple spreadsheet. Each row should capture:
-
System name and vendor
-
Primary use case (one sentence)
-
Annex III category checked (or "none")
-
Risk tier assigned
-
Key obligations triggered (e.g. Article 26 human oversight, Article 50 transparency notice, DPIA)
-
Date classified and name of person who classified it
This register becomes the backbone of your compliance programme. From 2 December 2027, Article 26 of the AI Act requires deployers of high-risk systems to keep logs, assign human oversight to people with the necessary competence, training and authority, and monitor how the system runs. Without the register, none of that work has a solid starting point.
Keep the register updated. If you add a new AI tool, run it through the eight questions before go-live, not after.
The three mistakes SMEs make most often
1. Assuming the vendor has classified it for you. The provider classifies its system against its intended purpose (Art 6(4)). It cannot see how you actually use it. If you put a tool to an Annex III use it was not designed for, the tier changes, and you may even take on provider duties yourself (Art 25(1)). Check your use, not just the vendor's label.
2. Classifying once and forgetting. AI systems are updated. A tool that scored minimal-risk last year may have added biometric features this year. Build a review trigger into your annual compliance calendar.
3. Misreading the Annex III "employment" category. This is the most consequential error for HR teams. Any AI that ranks, filters, or recommends candidates, allocates tasks based on individual behaviour or personal traits, evaluates performance, or monitors productivity falls squarely in Annex III category 4. If your company uses an ATS with AI-powered scoring, a scheduling tool that allocates shifts based on individual behaviour or traits, or a performance management platform that surfaces risk flags about employees, those are high-risk systems. From 2 December 2027 they need human oversight by trained, competent people, log retention and information for the workers concerned. A fundamental rights impact assessment (Art 27) is only required if you are a public body, provide a public service, or use AI for credit scoring or life and health insurance pricing.
What comes after classification
Once every system in your stack has a tier, the work splits into tracks:
-
Prohibited systems: Immediately cease use or request the vendor to disable the prohibited capability.
-
High-risk systems: Build your Article 26 compliance pack before 2 December 2027. That means use-case documentation, human oversight design, log retention, and a record of the training given to overseers. Add a fundamental rights impact assessment only if Article 27 applies to you.
-
Limited-risk systems: Draft or update your AI transparency notices. Make sure users can identify AI-generated content.
-
Minimal-risk systems: File and move on, and cover them in your AI literacy measures (Article 4). Check again if the tool is updated.
Usually the high-risk list is short, with a handful of limited-risk tools and a longer tail of minimal-risk applications. The classification step usually takes one afternoon. The Article 26 compliance work for each high-risk system takes longer, but having an accurate register keeps it focused.
Start today with one system. Run it through the eight questions. Record the result. That single action is the most useful 30 minutes you will spend on AI Act compliance this year.
Not sure whether your AI tools clear the Annex III threshold? The Khairos Comply free 2-minute check maps your current stack against each risk tier and flags the obligations that apply to your situation.
References
Not sure where your company stands?
Answer 15 questions about how your team uses AI. You get a compliance score and the gaps to close first, as a PDF.