EU AI Act Risk Classification: A Practical Guide to the Four Tiers

The EU AI Act assigns every AI system placed on the EU market or used in the EU to one of four risk tiers. Each tier carries distinct legal obligations — from outright prohibition to a simple transparency label. For organisations with 50 to 500 employees running a portfolio of SaaS tools, getting this classification right determines whether a tool must be retired, subjected to conformity assessment, or merely disclosed to end-users.

This guide walks through each tier, explains how high-risk obligations land on common mid-market tools, and provides a step-by-step classification workflow your team can apply today.

The four risk tiers at a glance

The AI Act uses a pyramid structure:

1. Prohibited practices — systems that are banned outright.
2. High-risk AI systems — systems subject to the heaviest compliance obligations (conformity assessment, risk management, documentation, human oversight).
3. Limited-risk (transparency) AI systems — systems that must meet transparency duties but not the full high-risk compliance regime.
4. Minimal-risk AI systems — systems with no specific obligations under the AI Act (e.g. spam filters, AI-enabled video games).

Most commercial SaaS AI tools fall into tiers 2 or 3. Understanding where your tools sit is the starting point for any compliance programme.

Prohibited practices: Art. 5 obligations that hide in plain sight

The prohibitions under Article 5 have been in force since 2 February 2025. Several map directly onto tools that mid-market companies already deploy:

• Subliminal manipulation — any AI technique that uses subliminal or superliminal techniques to distort behaviour, causing significant harm. A marketing-optimisation tool that exploits cognitive biases to manipulate purchasing decisions could qualify.
• Social scoring — AI systems that evaluate or classify natural persons or groups based on social behaviour or known, inferred, or predicted personal characteristics. Some employee-monitoring SaaS platforms that score workers on “trustworthiness” or “engagement” hover close to this boundary.
• Emotion recognition at work or in education — AI systems that infer emotions in the workplace or in educational institutions are prohibited, with narrow exceptions for medical or safety uses. That means any HR tool claiming to detect “engagement” or “frustration” through facial analysis during remote work is unlawful.
• Untargeted scraping for facial-recognition databases — systems that scrape facial images from the internet or CCTV to build or expand facial-recognition databases are banned. This matters if your company uses any third-party image-tagging or people-search SaaS product.
• Biometric categorisation by sensitive attributes — inferring race, political opinions, trade-union membership, religious beliefs, sex life, or sexual orientation from biometric data is prohibited. An AI tool that categorises job applicants by inferred “cultural fit” from biometric inputs crosses this line.

Before classifying your AI portfolio, scan for any tool whose core functionality matches one of these prohibitions. If it does, the tool must be discontinued or fundamentally redesigned — no amount of documentation will make a prohibited practice lawful.

Annex III high-risk AI systems: where mid-market tools get caught

Annex III enumerates eight broad categories of high-risk AI systems. For EU companies with 50 to 500 employees, the categories most likely to apply are:

Employment and worker management

Any AI system used for CV screening, candidate ranking, promotion or termination decisions, task allocation, or performance monitoring is high-risk. A 200-person SaaS company using an AI-powered recruiting platform that ranks candidates by predicted job fit is deploying a high-risk system — even if the tool is provided by a third-party vendor.

Education and vocational training

AI systems used for admissions decisions, exam proctoring, or evaluating learning outcomes are high-risk. An e-learning company whose platform uses AI to assess student performance or recommend curricula falls within this category.

Access to essential services — creditworthiness and insurance

AI systems used for credit scoring, creditworthiness assessment, life and health insurance risk assessment, and insurance pricing are high-risk. Any fintech or insuretech tool that uses AI to determine a natural person’s access to financial services or insurance products triggers the full high-risk compliance regime.

Other Annex III categories

The remaining high-risk categories include: biometric identification and categorisation, management and operation of critical infrastructure, law enforcement, migration and border control, and administration of justice and democratic processes. These are more relevant to public-sector deployers or specialised providers but remain important for vendors selling into those markets.

Provider vs deployer: who bears which obligation

The AI Act draws a sharp distinction between providers and deployers:

• A provider develops or places an AI system on the market under its own name or trademark. Most mid-market companies are not providers of AI systems — they are buyers.
• A deployer uses an AI system under its authority. This is the role that most 50–500-employee EU companies occupy.

As a deployer of a high-risk AI system, your duties under Article 26 include:

1. Using the system in accordance with the provider’s instructions.
2. Ensuring human oversight by competent individuals.
3. Monitoring the system’s operation and reporting malfunctions to the provider.
4. Informing affected individuals that a high-risk AI system is being used.
5. Keeping input data logs relevant to the high-risk system’s output.
6. Informing workers and their representatives about the deployment.

Article 27 additionally requires certain deployers — public bodies, private entities providing public services, and deployers of credit-scoring or insurance-pricing systems — to conduct a Fundamental Rights Impact Assessment (FRIA) before deploying a high-risk AI system.

Art. 50 transparency duties for limited-risk systems

The AI Act’s Article 50 transparency obligations enter into force on 2 August 2026. They apply to limited-risk AI systems that interact with natural persons or generate content. Key obligations include:

• Chatbots and conversational AI must clearly disclose that the user is interacting with an AI system.
• Synthetic audio, image, video, or text content that could be mistaken for human-generated must be marked in a machine-readable format.
• Deepfakes — AI-generated or manipulated content that looks like a real person — must be disclosed to the recipient.

These obligations are less onerous than the high-risk regime but legally binding. A customer-support chatbot that does not disclose its AI nature, or an AI-generated marketing image lacking machine-readable metadata, would violate Art. 50 from August 2026.

A six-step classification workflow

Use this workflow to classify every AI system in your portfolio:

Step 1 — Inventory your AI systems

List every tool, platform, or module that uses AI, including embedded features in your existing SaaS stack (CRM lead scoring, recruitment screening tools, etc.).

Step 2 — Screen against Art. 5 prohibitions

Check whether any system’s core functionality matches a prohibited practice. If it does, stop using it or remove the prohibited feature.

Step 3 — Check Annex III categories

Determine whether the system falls into one of the eight Annex III high-risk categories. Pay particular attention to employment, education, and access to essential services.

Step 4 — Apply Art. 6 exemption criteria

A system is only high-risk if it is intended as a safety component or if its output is a decision-making factor with material legal or significant effects. If neither condition is met, obligations may be reduced.

Step 5 — Assess transparency duties

For systems that are neither prohibited nor high-risk, check whether Art. 50 transparency obligations apply (chatbot interfaces, synthetic content generation, deepfakes).

Step 6 — Document your classification

Record the classification rationale for each system, including the tier assigned, the articles relied upon, and the date of assessment. This documentation is the evidence base for supervisory authorities.

The 2 December 2027 deadline for high-risk systems

The Digital Omnibus regulation (Council final approval: 29 June 2026) delayed the high-risk obligations for Annex III systems to 2 December 2027. This does not mean companies can ignore high-risk obligations until then — it means the clock is ticking and organisations have a defined runway.

Given that conformity assessment, risk management systems, and documentation requirements take months to implement, companies deploying Annex III high-risk AI should begin their compliance work no later than mid-2027. For systems involving credit scoring, insurance pricing, or employment decisions, the preparation period is particularly important because Art. 27 FRIA and Art. 26 deployer duties must be satisfied before deployment.

For an in-depth risk classification of your AI portfolio, consider Renomira, a purpose-built platform for inventorying AI tools and mapping them to the AI Act’s requirements.

Frequently asked questions

Is every AI system used in HR software automatically high-risk?

Not every AI tool in HR software is high-risk, but the most common use cases are. CV screening, candidate ranking, promotion and termination decisions, and performance monitoring all fall within Annex III. An AI feature that merely summarises interview notes without producing a recommendation may be limited-risk, but the threshold is low.

Do start-ups with fewer than 50 employees face the same obligations?

Yes. The AI Act applies to all deployers regardless of size. Fines scale with global turnover, but the compliance obligations — inventory, classification, and transparency — apply equally. There is no SME exemption from the core requirements.

What is the difference between a provider and a deployer?

A provider develops an AI system or places it on the market under its own name. A deployer uses a third-party AI system under its authority. Most 50–500-employee EU companies are deployers. The provider handles conformity assessment and technical documentation, but the deployer must still ensure lawful use, human oversight, and — where required — a Fundamental Rights Impact Assessment.

Does the AI Act apply to AI tools hosted outside the EU?

Yes. The AI Act has extraterritorial scope. If an AI system’s output is used in the EU, the Act applies regardless of where the provider is established. A US-based SaaS vendor selling an AI recruiting tool to an EU company is subject to provider obligations, and the EU company to deployer obligations.

When do the transparency obligations take effect?

Art. 50 transparency obligations apply from 2 August 2026. Chatbots must disclose their AI nature, synthetic content must be machine-readably marked, and deepfakes must be explicitly disclosed. This is separate from the high-risk obligations, which are delayed to 2 December 2027 for Annex III systems.