High-risk AI systems: classification, deadlines and duties

Article 6 defines two independent tracks. The Annex III track (Article 6(2)) lists standalone use cases considered high-risk because of their impact on people: biometrics (remote biometric identification, biometric categorisation by sensitive attributes, emotion recognition where not prohibited outright); critical infrastructure safety; education and vocational training (admission, evaluation, exam proctoring); employment (recruitment, screening and filtering of applications, decisions on promotion, termination, task allocation and monitoring); access to essential private and public services (including creditworthiness scoring and life/health insurance pricing); law enforcement; migration and border control; and administration of justice and democratic processes.

The Annex I track (Article 6(1)) covers AI that is a product, or a safety component of a product, regulated by listed EU harmonisation law and subject to third-party conformity assessment — think medical devices, aviation, lifts, toys. Note the Omnibus narrowed the safety-component definition (Article 3(14)): the component's intended purpose must be to prevent or mitigate health-and-safety risks; assistance, optimisation and convenience functions are excluded.

The Digital Omnibus (Regulation (EU) 2026/1744) moved the high-risk application dates. Annex III systems must comply from 2 December 2027. Annex I Section A products follow from 2 August 2028. Certain pre-existing systems used by public authorities have a longstop of 2 August 2030, and legacy high-risk systems already on the market are generally caught only upon significant changes in their design after the applicability date (Article 111(2)).

Do not read the postponement as a reprieve to do nothing: a provider building an Annex III system needs 12-18 months to stand up risk management, data governance, documentation and conformity assessment. Companies that treat December 2027 as a delivery date, not a start date, will be the ones that make it.

Article 6(3) lets a provider conclude that a system falling under Annex III is not high-risk because it does not pose a significant risk of harm to health, safety or fundamental rights — where the system only performs a narrow procedural task, improves the result of a previously completed human activity, detects decision-making patterns without replacing human review, or performs a preparatory task. Three points companies get wrong:

• The derogation is open to all Annex III areas, including biometrics. There is no general carve-out excluding Annex III point 1 from Article 6(3); what matters is whether one of the conditions in points (a)-(d) is genuinely met.
• Profiling is an absolute bar: an Annex III system that performs profiling of natural persons is always high-risk — no derogation possible.
• Self-assessment is not silent: the provider must document the assessment and register the system under Article 49(2). The Omnibus kept this registration duty (the proposal to drop it was rejected), so 'we decided internally it's not high-risk' without documentation and registration is non-compliance in itself.

Our Risk Classification template walks through Annex III mapping, the 6(3) conditions and the documentation record.

From the applicable date, providers must meet the full Chapter III, Section 2 programme: a continuous risk management system (Article 9); data governance over training, validation and testing data (Article 10); technical documentation per Annex IV (Article 11 — SMEs may use the simplified documentation form); automatic logging (Article 12); transparency and instructions for use for deployers (Article 13); human oversight by design (Article 14); accuracy, robustness and cybersecurity (Article 15). On top come a quality management system (Article 17), conformity assessment and CE marking, EU database registration (Article 49), post-market monitoring (Article 72 — the Commission must issue guidance including a template by 2 September 2027) and serious-incident reporting (Article 73: immediately and at the latest 15 days after awareness; 2 days for widespread infringements or serious and irreversible critical-infrastructure disruption; 10 days in the event of a death).

Deployers get a shorter but real list (Article 26): operate the system in accordance with the provider's instructions for use; assign human oversight to people with the competence, training and authority to intervene; ensure input data is relevant and sufficiently representative insofar as they control it; monitor operation and suspend use if the system presents risk; keep automatically generated logs; and inform workers and their representatives before using high-risk AI on them. Public bodies and private deployers of certain Annex III systems (including credit scoring and insurance pricing) must additionally run a fundamental rights impact assessment (Article 27) before first use — the AI Office is required to develop a template questionnaire, but none has been published as of mid-2026, so use an internal template and reconcile later. Our FRIA and Human Oversight templates cover both duties.

The Omnibus moved the Machinery Regulation (EU) 2023/1230 from Annex I Section A to Section B. The consequence is structural: AI-enabled machinery no longer follows the AI Act's direct high-risk conformity path — the Chapter III, Section 2 requirements do not apply directly to Section B products the way they do to Section A products such as medical devices. Instead, AI-related safety requirements will be channelled into the machinery framework itself through Commission delegated and implementing acts. If you build AI into machinery, your compliance track now runs primarily through machinery law, with the Annex I application date of 2 August 2028 as the horizon — watch the secondary legislation, not just the AI Act text.