High-risk AI mapped: classification and obligations in one overview

The high-risk regime of Chapter III is the centre of gravity of the AI Act: by far the most obligations, the conformity assessment and the CE marking hang from it. It is also the part whose date of application is expected to shift, under the Digital Omnibus agreement of 7 May 2026 (see the timeline), to 2 December 2027 (Annex III) and 2 August 2028 (Annex I). What has to be done does not change — only when.

Two routes into the high-risk classification

Route 1 — Annex I: the AI system is a safety component of a product already covered by European product legislation (machinery, medical devices, toys, lifts), or is such a product itself, and requires third-party conformity assessment under that legislation.

Route 2 — Annex III: the system falls within one of eight listed areas, including biometrics, critical infrastructure, education, recruitment and employment decisions, access to essential services (including creditworthiness and insurance), law enforcement, migration and the administration of justice.

The filter provision (Article 6(3)): an Annex III system is nevertheless not high-risk if it poses no significant risk to health, safety or fundamental rights — for instance because it performs a narrow procedural task or merely prepares a human assessment. Whoever invokes the filter must document that assessment; profiling of natural persons always counts as high-risk.

Provider obligations

The provider carries the heaviest package (Articles 8 through 17): a continuous risk management system; requirements on training, validation and test data including bias controls; technical documentation; automatic logging; transparent instructions for use to the deployer; design for effective human oversight; accuracy, robustness and cybersecurity; and a quality management system. Then follow the conformity assessment, the EU declaration of conformity, CE marking and registration in the EU database (Article 49).

Deployer obligations

Whoever puts a high-risk system to use has its own list (Article 26): use in accordance with the instructions, assignment of competent human oversight, relevant and representative input data, monitoring of operation, retention of logs and informing workers when deployed in the workplace. Public bodies and providers of essential private services must in addition carry out a fundamental rights impact assessment (FRIA, Article 27) before first use. For financial institutions, the bridging provisions connect these duties to existing governance — see the separate analysis on the interplay with DORA.

What the delay does and does not mean

Until the Digital Omnibus appears in the Official Journal, 2 August 2026 formally remains the applicable date; thereafter 2 December 2027 becomes the planning horizon for Annex III. Not delayed are the provisions that already apply — the prohibitions, the literacy obligation and the GPAI regime. The content of the high-risk package does not change either. Organisations that build the components now (register, risk management, data quality, documentation) do no work that will be lost; they merely spread the load.