The EU guidelines on high-risk AI systems: what companies need to know

On 19 May 2026 the European Commission published draft guidelines on the classification of high-risk AI systems — the first set of concrete examples clarifying Regulation (EU) 2024/1689. They are not binding, but they show how authorities are likely to read Article 6, and they give companies operating in Romania clear reference points.

The compliance date for high-risk systems under Annex III is now 2 December 2027, following Regulation (EU) 2026/1744, the Digital Omnibus on AI, in force since 27 July 2026. By then, providers must meet the requirements of Articles 9 to 15 and deployers their obligations under Article 26.

What has changed

The Omnibus moved the strictest deadlines: Annex III systems from 2 August 2026 to 2 December 2027, and AI in products regulated under Annex I to 2 August 2028. It is a preparation window, not a relaxation of the requirements.

The guidelines add what the Regulation itself does not contain: practical examples of systems that fall into each Annex III area, operational criteria for the Article 6(3) exceptions, and the Commission’s reading of key notions such as profiling. After the public consultation, the final version had not been published at the time of writing.

The eight Annex III areas, in brief

For the obligations that follow, see high-risk AI systems.

1. 1.Biometrics — remote identification, emotion recognition, biometric categorisation
2. 2.Critical infrastructure — road traffic, utilities (water, gas, electricity, heating), critical digital infrastructure
3. 3.Education and vocational training — admission, evaluation of results, monitoring behaviour during tests
4. 4.Employment and worker management — recruitment, performance evaluation, task allocation, promotion and termination decisions
5. 5.Essential private and public services — credit scoring, life and health insurance pricing, eligibility for benefits, emergency triage
6. 6.Law enforcement — risk assessments of victims or offenders, evaluation of evidence, profiling in investigations
7. 7.Migration, asylum and border control — risk assessment, examination of visa and asylum applications, detection and recognition of persons
8. 8.Administration of justice and democratic processes — assisting judicial authorities, influencing the outcome of elections

The exceptions in Article 6(3)

Not every Annex III system is high-risk. A system that poses no significant risk to health, safety or fundamental rights is not, where it performs one of these functions:

• a)
  A narrow procedural task

  A limited technical operation with no assessment of its own — for example, automatically formatting documents.

• b)
  Improving the result of a completed human activity

  Refining a result already produced by a person without replacing it — for example, correcting a text someone wrote.

• c)
  Detecting decision-making patterns

  Identifying patterns in existing human decisions without replacing them or influencing them absent proper human review.

• d)
  A preparatory task

  A step preparatory to an assessment that remains with a human — for example, sorting documents before they are analysed.

What companies should do now

1. 1
   Make an inventory of all AI systems

   Every AI tool in use: assistants, recruitment systems, credit scoring, performance evaluation, chatbots, content generators — and AI features inside everyday software such as CRM, ERP and HR tools.

2. 2
   Classify each system

   For each one, check whether it falls into an Annex III area and whether an Article 6(3) exception applies. Document the reasoning; it becomes part of the compliance file.

3. 3
   Test whether human oversight is real

   Article 14 requires oversight by people who understand the system’s limits and are aware of the tendency to rely on its output automatically. A person who approves the system’s recommendations without real review provides oversight on paper only.

4. 4
   Start the technical documentation

   For providers of high-risk systems, the documentation under Article 11 — specifications, performance, limitations, risk mitigation — is built during development, not at the end.

5. 5
   Check the overlap with the GDPR

   An AI system that processes personal data creates cumulative obligations under the AI Act and the GDPR: a DPIA, a lawful basis, transfers to AI providers, processing agreements. Assess the overlap during the inventory.

6. 6
   Plan budget and resources

   Compliance for high-risk systems is not a short task. It needs dedicated people, internal or external, people assigned to oversight, and tools for logging and monitoring.

Conclusion

The Commission’s guidelines turn Article 6 from a general text into an operational framework. For companies operating in Romania with AI systems that affect people — credit scoring, recruitment, performance evaluation, insurance pricing, biometric identification — inventory and classification are the work to do now. The December 2027 date gives time to prepare; it does not remove the need to.

Frequently asked questions