knowledge 27 May 2026 |

Employment Spotlight: EU AI Act - Draft Guidelines on High-Risk AI Classification

On 19 May 2026 the European Commission published draft Guidelines on the classification of high-risk AI systems under Article 6 of the EU AI Act, launching a five-week public consultation running until 23 June 2026. The Guidelines are organised around the two pathways through which an AI system may be classified as high-risk:

• If the AI system is integrated into a regulated product.
• If the AI system comes within one of eight areas set out in Annex III of the AI Act which are viewed as “particularly susceptible” to the risks inherent in AI systems; these include biometrics; critical infrastructure; employment; and access to and enjoyment of essential private and public services. 

If an AI system is classified as high-risk, the AI Act imposes obligations in terms of risk management, data governance and management, technical documentation, record-keeping, transparency of operations, instructions for deployers, ‘human in the loop’, accuracy, robustness and cybersecurity. Obligations are also imposed on providers and deployers of the system.

Although the Guidelines will not be legally binding, they will represent the most authoritative indication to date of how the Commission and national regulators are likely to approach classification.

WHAT DO EMPLOYERS NEED TO KNOW?

The draft Guidelines confirm that the recent political agreement on the AI Digital Omnibus has postponed the application date for the obligations relating to Annex III high-risk AI systems, including employment-related systems, from 2 August 2026 to 2 December 2027. Such high-risk AI systems in use prior to 2 December 2027 are not subject to the AI Act unless they are subject to “significant changes in their design”, or if they are used by public authorities (in which case they must comply by 2 August 2030). Otherwise, the substantive compliance obligations have not changed. Organisations deploying AI in any workforce-related context should treat this extended timeline as an opportunity to prepare and consider whether to submit feedback to the Commission on points that affect their operations.

The overarching principles for employers are straightforward:

• If the AI system materially influences a decision that affects a person’s career prospects, livelihood, or rights, it is likely high-risk. 
• If the AI system performs a purely logistical, informational, or retrospective function without shaping outcomes, it is likely outside scope or eligible for the filter.

KEY TAKEAWAYS FROM THE DRAFT GUIDELINES

Below, we highlight the key takeaways relating to the employment category under Annex III that are particularly relevant for organisations deploying AI in a workforce context.

The Article 6(3) filter offers limited relief

Many organisations have looked to the Article 6(3) ‘filter mechanism’ as a potential route to escape high-risk classification under Article 6(2). Under the filter mechanism, a provider may self-assess that its system is not high-risk even though it falls within an Annex III use case, provided the system meets one of four conditions:

• it performs a narrow procedural task;
• it improves a previously completed human activity;
• it detects decision-making patterns or deviations from prior decision-making patterns and is not meant to replace or influence the previously completed human assessment without proper human review; or
• it performs a preparatory task.

The Guidelines clarify that the above conditions are exhaustive, but alternative, meaning if the AI system falls within at least one of the conditions, it may be deemed not to pose significant risk of harm to the health, safety or fundamental rights of natural persons despite the fact that certain risks might remain.

However, any system that engages in profiling (meaning any automated processing of personal data to evaluate personal aspects relating to a natural person) is categorically excluded from the filter. Given that most AI systems deployed in recruitment, performance management, and task allocation will inevitably involve the processing of personal data to evaluate individuals, this effectively removes a substantial proportion of HR-related AI use cases from the scope of the exemption.

In addition, employers should not assume that introducing a ‘human in the loop’ will alter the outcome; the Commission makes clear that human involvement alone does not change the classification.  While the type and degree of human involvement during the deployment of the system may play a role in the classification of an AI system as high-risk, this is only to demonstrate that the tasks the system is intended to perform are narrow procedural tasks or preparatory in nature, or that the system is only intended to improve a previously completed human activity.

The Commission reads the employment category broadly

Point 4 of Annex III captures AI systems used in ”employment, workers' management and access to self-employment”. In-scope are AI systems used for recruitment and selection (including targeted job advertising, CV filtering and candidate evaluation), as well as systems used to make decisions affecting terms of work-related relationships, promotion, termination, task allocation based on individual behaviour, traits or characteristics and the monitoring and evaluation of worker performance and behaviour in work related relationships.

Critically, the Guidelines confirm that the scope is not limited to traditional employees. Freelancers, independent professionals, platform workers, service providers and self-employed persons all fall within scope.  The Commission has clarified that “work-related contractual relationships” encompasses the broader spectrum of arrangements through which work is performed, extending beyond formal contracts of employment.

Not everything used in employment is high-risk - the boundaries matter

The Guidelines offer a rich set of practical examples that help draw the line between what is in and what is out of scope of the high-risk classification. The following is a high-level summary of examples from the Guidelines which indicate where the boundaries sit.

What’s in (high-risk):

• AI systems intended to be used as an automated job matching and ranking tool that score, rank or shortlist candidates in recruitment by processing structured or unstructured data (e.g. CVs, social and emotional competencies), where the system’s scores and rankings serve as the primary input for decision making.
• Targeted job advertising systems that use profiling to decide which individuals see which vacancies.
• AI-enabled shift schedulers that allocate shifts, rest periods and on-call windows based on behavioural signals such as punctuality, acceptance rates or customer ratings, thereby affecting workers’ pay and progression. 
• AI systems for workload allocation which aggregate behavioural inputs (e.g. turnaround time on prior assignments, responsiveness to emails, voluntary participation in business development activity) into a score which serves as the basis for allocation of work.
• AI systems that dynamically set worker pay based on ratings, completion times and acceptance rates.

What’s out (not high-risk):

• AI systems that are intended to be used for employer brand advertisements to present the company as an attractive place to work and are not tied to a specific vacancy. 
• AI systems that assist candidates (not employers) in tailoring CVs to specific open positions. 
• Employee onboarding chatbots that provide information about policies and training schedules for onboarding after hiring. 
• AI systems to assess employees’ progress in voluntary training modules by analysing quiz results, which only provides feedback and learning recommendations to the employee to support development, and is not used in performance assessment.
• Office-space optimisation and corporate travel-planning tools.

On the borderline (in scope but exempt under the Article 6(3) filter):

• Systems that merely organise CVs into a searchable database without scoring candidates.
• Interview-scheduling tools that coordinate logistics without influencing selection. 
• Retrospective analytics auditing past hiring decisions for bias by analysing anonymised recruitment data, without feeding results into live decisions.
• AI systems intended to be used for sending personalised emails acknowledging receipt of applications with the name and pronouns of the candidate who applied.
• AI writing assistants that refine already-completed promotion evaluations without changing outcomes. 

How can McCann FitzGerald LLP help?

For further information or assistance, please reach out to one of the key contacts below, or your usual contact at McCann FitzGerald LLP.