High-risk AI systems: does the AI Act apply to you?

A high-risk AI system is an artificial intelligence system that the AI Act subjects to its strictest requirements because it can affect people's health, safety or fundamental rights. It qualifies through one of two routes: being a safety component of an already regulated product (Annex I), or being used in a sensitive area such as recruitment, credit or education (Annex III). Since the July 2026 Digital Omnibus, the corresponding requirements will apply on 2 December 2027 for Annex III and 2 August 2028 for Annex I.

The EU AI Act sorts uses of artificial intelligence into risk levels. High risk sits between prohibited practices and uses that are merely subject to transparency: these systems are allowed, provided they meet a set of technical and organisational requirements before being placed on the market and throughout their use.

Classification does not depend on the technology used but on the system's intended purpose, as defined by its provider. The same text analysis engine can be minimal-risk when sorting emails and high-risk when assessing job candidates. Article 6 of the regulation sets the classification rules, with two separate routes.

The stakes are high: a high-risk system must meet the requirements of Articles 9 to 15, undergo conformity assessment and bear the CE marking, and its use creates obligations for the organisation deploying it. Misclassifying a system can lead to fines of up to €15 million or 3% of worldwide annual turnover.

Article 6(1) covers AI built into products already covered by EU harmonisation legislation. Two cumulative conditions must be met:

1. the AI system is a safety component of a product covered by Annex I, or is itself such a product;
2. that product must undergo a third-party conformity assessment (by a notified body) before being placed on the market.

Section A of Annex I includes toys, recreational craft, lifts, equipment for explosive atmospheres, radio equipment, pressure equipment, cableway installations, personal protective equipment, gas appliances, medical devices and in vitro diagnostic medical devices. Section B covers civil aviation, vehicles, marine and rail equipment: for these products, AI Act requirements apply through sector-specific legislation.

What the omnibus changed for Annex I

Regulation (EU) 2026/1744 narrowed this route. According to the consolidated version of Article 6, AI used solely for user assistance, performance optimisation, convenience or quality control is not a safety component, unless its failure would endanger health or safety. A product that requires third-party assessment for reasons unrelated to health and safety (radio interference, for example) does not meet the second condition either. The Machinery Regulation was also moved from Section A to Section B. Finally, this route now applies from 2 August 2028.

Article 6(2) refers to Annex III, which lists stand-alone uses, not linked to a product, considered sensitive for people's rights. The annex is organised in eight areas; within each, only the uses precisely described are covered.

Being listed in Annex III is not always enough. Article 6(3) provides that a system is not high-risk if it does not pose a significant risk to health, safety or fundamental rights, in particular because it does not materially influence the outcome of decision-making. This is the case if it meets at least one of four conditions:

• it performs a narrow procedural task, for example turning unstructured documents into structured data or sorting incoming mail by category;
• it improves the result of a previously completed human activity, such as rephrasing a text already written by an officer;
• it detects decision-making patterns or deviations from prior decisions, without replacing or influencing the human assessment without proper human review;
• it performs a preparatory task for an assessment relevant to an Annex III use, such as translating or indexing documents.

The profiling lock

The exception never applies when the system performs profiling of natural persons within the meaning of the GDPR, that is, automated evaluation of personal aspects (work performance, economic situation, reliability, behaviour…). In its analysis of the Commission's draft guidelines, the law firm Osborne Clarke notes that recruitment, credit assessment, migration and policing tools usually involve some form of profiling.

The duty to document and register

Relying on an exception is not merely an internal choice. Under Article 6(4), the provider must document its assessment before placing the system on the market, provide it to authorities on request and register the system in the EU database (Article 49(2)). The omnibus lightened the content of this registration, but not the duty to document. During an inspection, Article 80 allows the market surveillance authority to re-examine the classification and to penalise a provider that knowingly misclassified its system to avoid the requirements.

As of this article (16 September 2026), the guidelines required by Article 6(5) are not final: the Commission published a draft on 19 May 2026, open for consultation until 23 July, and plans to adopt them by the end of 2026.

Combining the two routes and the exceptions, classification boils down to a decision tree. It assumes two prior checks: the software is indeed an AI system within the meaning of Article 3, and its use is not a practice prohibited by Article 5.

Once a system is classified as high-risk, its provider must meet seven requirements, taking into account its intended purpose and the state of the art (Article 8). The table summarises them.

Recruitment and HR management

Software that automatically screens CVs, ranks candidates after a video interview or recommends promotions falls under Annex III (point 4). As it evaluates people, it performs profiling: the Article 6(3) exception is ruled out. On the other hand, a tool that merely anonymises CVs before a recruiter reads them may be a narrow procedural task, provided this is documented.

Credit and insurance

Assessing a borrower's creditworthiness or assigning a credit score is high-risk (point 5(b)), as is risk assessment and pricing in life and health insurance (point 5(c)). Systems that detect financial fraud are expressly excluded. Banks and insurers deploying these systems must also carry out a fundamental rights impact assessment.

Healthcare

Healthcare is concerned by both routes. Diagnostic support software that is a class IIa or higher medical device, and therefore certified by a notified body, is high-risk via Annex I, applying from 2 August 2028 with an assessment integrated into the Medical Devices Regulation procedure. An emergency department patient triage system or an emergency call dispatch system falls under Annex III (point 5(d)). Conversely, an operating theatre scheduling tool or a report transcription tool is not, in principle, high-risk.

Education

Systems that decide admission to an institution, grade papers, steer students to an education level or monitor cheating during an exam are high-risk (point 3). A tool that helps a teacher rephrase their own comments could fall under the exception for improving a previously completed human activity.

Essential public services

A public administration or social security body that uses AI to assess entitlement to a benefit, grant it, reduce it or reclaim it falls under Annex III (point 5(a)). Public bodies deploying these systems must carry out a fundamental rights impact assessment and register their use in the EU database. A chatbot that informs users about procedures is not high-risk in itself, but it is subject to transparency obligations.

High-risk requirements will apply on 2 December 2027 for Annex III systems and 2 August 2028 for Annex I systems. Systems placed on the market before these dates will only be caught if their design changes significantly, a rule now assessed per type and model of system (specific rules apply to systems used by public authorities). This time is for preparation, especially as harmonised standards and final guidelines are still pending.

1. List the AI systems you use or develop, with their exact intended purpose.
2. Run each system through the decision tree and document the outcome, especially any exception relied on.
3. Question the providers of systems that are likely high-risk: compliance roadmap, available documentation, instructions for use, human oversight measures. This is the job of your third-party risk management process.
4. Prepare how human oversight will be organised: who supervises, with what training, with what power to stop or correct.
5. Link this work to your existing analyses: risk analysis, DPIA, business continuity plan.

For an overview of the regulation (risk levels, full timeline, penalties, authorities), see our main article EU AI Act: what the AI regulation changes for you. To run and track risk analyses of your AI systems, Phinasoft's risk analysis module accepts your own method.