EU AI Act for Companies: Which Risk Class Applies to Your System — and What Needs to Happen by When?
Expert review: · Content updated:
The AI Act doesn't classify AI systems by model, but by purpose: the same foundation model can be minimal-risk in one use case and high-risk in the next. Four classes determine the obligations: prohibited, high-risk, limited-risk, minimal-risk. The costliest misconception for companies: AI-powered recruiting is almost always high-risk — even if it feels like just an add-on feature in existing applicant software.
This guide was reviewed by an expert on 23 July 2026 and its content was updated on 31 October 2026.
Which four risk classes does the AI Act distinguish?
The classification depends on purpose, not model: the same foundation model can be minimal-risk in one use case and high-risk in the next. The class is thus a property of the specific application — not the technology behind it. Most enterprise use cases fall into the two middle classes.
Art. 5 — since February 2025
Social scoring, manipulative systems, emotion recognition in the workplace and in education institutions (Art. 5(1)(f), except for medical or safety reasons), real-time remote biometric identification in publicly accessible spaces for law enforcement. Narrow exceptions exist only for this real-time remote identification by law enforcement (Art. 5(1)(h), 5(2)–(7)).
Annex III
Applicant screening, performance evaluation, credit scoring, risk assessments by law enforcement authorities (Annex III point 6). Predicting the risk of a person committing a criminal offence based solely on profiling or on their personality traits remains prohibited (Art. 5(1)(d)). Conformity assessment, documentation, human oversight, registration.
Art. 50 — labeling requirement
Chatbots, AI-generated content, emotion recognition outside the workplace and education (prohibited there under Art. 5(1)(f)). Users must know they are interacting with an AI.
The vast majority of systems
AI-powered search, spam filters, recommendation systems, simple automation. No obligations specific to the risk class — but the AI literacy obligation of Art. 4 applies to all providers and deployers regardless of risk class, and GDPR still applies.
A four-tier pyramid, from top to base: prohibited (Art. 5), high-risk (Art. 6 and Annex III), limited-risk with transparency obligations (Art. 50), and minimal-risk with no class-specific obligations. Obligations grow with the risk; the wide base stands for the vast majority of systems. The classification depends on purpose, not model.
Which obligations apply to each risk class?
Obligations grow with the risk, and they do not hit everyone involved in the same way. The AI Act mainly distinguishes between providers, who develop an AI system or place it on the market under their own name, and deployers, who use it under their own authority (Art. 3(3) and (4)). The table summarizes scope and core obligations per class; the wording of the Regulation remains authoritative.
| Risk class | What falls under it | Core obligations | Example |
|---|---|---|---|
| Prohibited (Art. 5) | The eight practices in Art. 5(1), including manipulative techniques, social scoring, untargeted scraping of facial images from the internet or CCTV footage, and emotion recognition in the workplace and in education institutions. | Placing on the market, putting into service and use are prohibited. Infringements can be fined up to EUR 35 million or up to 7% of worldwide annual turnover, whichever is higher (Art. 99(3)); for SMEs, including start-ups, the lower of the two amounts applies (Art. 99(6)). | Software that infers the emotions of employees from video footage (Art. 5(1)(f)). |
| High-risk (Art. 6) | AI as a safety component or product under Annex I that is subject to third-party conformity assessment (Art. 6(1)), and systems in the eight areas of Annex III (Art. 6(2)). | Providers: risk management (Art. 9), data governance (Art. 10), technical documentation (Art. 11), record-keeping (Art. 12), transparency and instructions for use (Art. 13), human oversight (Art. 14), accuracy, robustness and cybersecurity (Art. 15), quality management (Art. 17), conformity assessment (Art. 43), registration for Annex III systems (Art. 49). Deployers: obligations under Art. 26, in some cases a fundamental rights impact assessment (Art. 27). | AI system that analyses and filters job applications (Annex III point 4(a)). |
| Limited risk: transparency obligations (Art. 50) | Systems that interact directly with people, systems that generate synthetic audio, image, video or text content, emotion recognition and biometric categorisation where not prohibited, and deepfakes. | Disclosure that an AI is involved (para. 1), machine-readable marking of generated content (para. 2), informing affected persons about emotion recognition or biometric categorisation (para. 3), disclosure of deepfakes (para. 4). These duties apply in addition to any high-risk obligations (Art. 50(6)). | Customer service chatbot. |
| Minimal risk | All AI systems that do not fall into any of the other three classes. | No class-specific obligations. Providers and deployers must still ensure a sufficient level of AI literacy among their staff (Art. 4); voluntary codes of conduct are possible (Art. 95). | Spam filter or AI-powered search. |
Cutting across the four classes are the obligations for providers of general-purpose AI models (Art. 53 to 55). They attach to the model, not to the purpose. Integrating such a model unchanged into your own application does not, as a rule, make you the model provider; you are the downstream provider of the AI system built on it (Art. 3(68)) and must classify that system yourself.
Which eight areas does Annex III list?
Annex III is the yardstick for most companies: if a system falls under one of the eight areas, it is in principle high-risk under Art. 6(2). What counts is not the area alone but the specific intended purpose that Annex III describes for each area. The Commission can amend or add use cases by delegated act (Art. 7). One example from the Annex per area:
- Biometrics: Remote biometric identification, where permitted under Union or national law. Pure verification that a person is who they claim to be is excluded (point 1(a)).
- Critical infrastructure: AI as a safety component in the management and operation of critical digital infrastructure, road traffic, or the supply of water, gas, heating or electricity (point 2).
- Education and vocational training: Systems that evaluate learning outcomes, including when those outcomes steer the learning process (point 3(b)).
- Employment, workers management and access to self-employment: Analysing and filtering job applications and evaluating candidates (point 4(a)).
- Access to essential private and public services and benefits: Evaluating the creditworthiness of natural persons or establishing their credit score. Systems used to detect financial fraud are excluded (point 5(b)).
- Law enforcement: Evaluating the reliability of evidence in the course of investigating or prosecuting criminal offences, where permitted (point 6(c)).
- Migration, asylum and border control: Assisting in the examination of applications for asylum, visas and residence permits and related complaints, where permitted (point 7(c)).
- Administration of justice and democratic processes: Assisting a judicial authority in researching and interpreting facts and the law and in applying the law to a concrete set of facts (point 8(a)).
Not every system in one of these areas is automatically high-risk. Under Art. 6(3) the classification does not apply if the system poses no significant risk to health, safety or fundamental rights, in particular because it only performs a narrow procedural task, improves the result of a previously completed human activity, detects decision-making patterns without replacing the human assessment absent proper review, or only performs a preparatory task. Two limits apply: a system that performs profiling of natural persons is always high-risk. And anyone relying on the exception must document the assessment before placing the system on the market and still register it (Art. 6(4), Art. 49(2)).
Provider or deployer: who carries which obligations?
Most companies buy high-risk AI rather than building it. That makes them deployers, and their list of obligations is shorter than the provider’s, but not empty. It becomes critical when their own changes shift the role.
Art. 16: meet and demonstrate the requirements
Ensures the system meets the requirements of Art. 8 to 15, runs a quality management system, completes the conformity assessment, draws up the EU declaration of conformity, affixes the CE marking and registers Annex III systems.
Art. 26: use as intended and oversee
Uses the system in line with the instructions for use, assigns human oversight to competent persons, keeps automatically generated logs for at least six months and informs workers’ representatives and affected workers before use in the workplace.
Art. 25: when the deployer becomes the provider
Anyone who puts their own name or trademark on a high-risk system, substantially modifies it, or changes the intended purpose of a system so that it becomes high-risk takes on the provider obligations under Art. 16.
By when does my system need to meet the requirements?
The deadlines are staggered, not a single cutoff: bans and transparency duties have long applied, general-purpose AI obligations since August 2025. The big shift affects high-risk Annex III systems — their obligations were pushed back 16 months to December 2027 by the Digital Omnibus, while labeling and transparency duties still kick in earlier, unchanged.
Prohibited practices take effect: Art. 5 (social scoring, manipulative systems, etc.) and AI literacy obligations (Art. 4) apply from this date.
Obligations for general-purpose AI: foundation model provider duties, governance structures, and the fine regime take effect.
Transparency & registration duties: general information and registration obligations for AI systems become applicable.
Labeling requirement for AI content: Art. 50(2) (labeling of AI-generated content) becomes applicable.
High-risk systems (Annex III) — postponed: new deadline via the Digital Omnibus (agreement May 7, 2026) — 16 months later than originally planned. Substantive obligations (Art. 6–27) unchanged.
High-risk as a safety component (Annex I): AI as a safety component in already-regulated products (e.g. machinery, medical devices) — its own, later deadline.
What exactly did the Digital Omnibus change?
The Digital Omnibus only shifts deadlines, not the substantive requirements. Reason: harmonized standards for companies to follow were not available before the end of 2026 — the EU didn't want to impose an obligation without a reliable benchmark, but deliberately chose not to water down the content itself.
Important: no framework replaces another. GDPR, the voluntary GPAI Code of Practice (with "safe harbor" effect), and the ISO/IEC 42001 certification standard all apply in parallel and cumulatively with the AI Act — focusing on just one deadline means missing the other three. Waiting through 2026 also wastes lead time: risk management and documentation can be built up independent of the exact deadline.
The misconception I encounter most often in projects: HR AI is almost always high-risk. Annex III point 4 explicitly names employment and workforce management — including AI-powered candidate screening, promotion decisions, and performance evaluation. Activating AI recruiting features in existing HR software very likely makes you a high-risk deployer. That classification belongs checked before go-live, not after.
Glossary
- High-Risk AI
- An AI system classified as high-risk under Art. 6 — as a safety component or product under Annex I, or through its use in one of the areas listed in Annex III.
- Conformity Assessment
- The process of demonstrating whether a high-risk AI system meets the requirements of Chapter III, Section 2 (Art. 3(20), Art. 43).
- Annex III
- The list of use areas in which AI systems are considered high-risk under Art. 6(2) — for example employment and workforce management, education, or creditworthiness assessment.
- GPAI
- General-purpose AI: an AI model with significant generality that can competently perform a wide range of distinct tasks and be integrated into downstream systems (Art. 3(63)).
Sources
- Regulation (EU) 2024/1689 (Artificial Intelligence Act), Official Journal of the EU — source for the articles and deadlines cited — https://eur-lex.europa.eu/eli/reg/2024/1689/oj
Know where you stand in 2–3 minutes
The EU AI Act check maps your system to one of the four risk classes without sign-up and shows the concrete obligations and deadlines that apply.
EU AI Act check: find your risk class for free →FAQ
Does the new deadline (December 2027) also apply to systems already in operation?
Yes — the postponement applies to all Annex III systems regardless of whether they're newly introduced or already in use. Companies that are already prepared don't need to wait for the deadline and can become compliant earlier.
Is AI-powered recruiting automatically high-risk?
In the vast majority of cases, yes. Annex III explicitly names employment and workforce management as a high-risk area — including AI add-on features in existing HR software like applicant screening. The classification belongs checked before rollout, not after.
Does the AI Act affect me outside the EU too?
Yes, if your system's output is used in the EU (market-location principle) — regardless of where the provider or deployer is based.
Is the 2027 deadline enough reason to do nothing yet?
Technically yes, practically no: conformity assessment, risk management, and documentation need several months of lead time for more complex systems — the extended deadline is a buffer, not a free pass.
Daniel Ostner
From Chapter 4 of the Enterprise AI Guide book
Related guides
Building AI Governance
Solid AI governance runs on three levels: strategic, tactical, operational. A minimal model with 5 roles and 6 core decisions is enough to get started.
Governance Decision Tree
Five questions determine whether an AI use case gets approved. Following the decision tree cleanly means approval in 5 business days instead of 5 months.