Point 4(a) of Annex III classifies as high-risk AI systems intended to be for the recruitment or selection of natural persons, in particular to place targeted job advertisements, to analyse and filter job applications, and to evaluate candidates.
The classification of AI systems intended to be used in recruitment and selection as high-risk is justified by the fact that those processes constitute the principal entry point into the labour market. Decisions taken at this stage can condition an individual’scareer prospects and access to livelihoods. At the same time, the structural power asymmetry between internal and external recruiters and candidates, combined with the high degree of information held by employers on the candidate and contracting parties for self-employed, the job requirements, and the labour market creates a context in which ensuring the protection of job applicant’s rights is essential.
For an AI system to be classified as high-risk pursuant to point 4(a) of Annex III , the activity for which it is intended to be used should relate to or impact the substance of the recruitment process: covering preparatory steps (special advertising, prospecting, pre-application screening) and the selection process (shortlisting, grading, ranking, or testing of candidates). The notions of ‘recruitment” and ‘selection” require a functional interpretation, since they are partly overlapping but not identical.
i. The notion of ‘recruitment’
Recruitment constitutes the processof preparing the constitution of a new work-related relationship, commencing already with the identification and attraction of potential applicants , and continuing until the conclusion of a contract, if applicable. This encompasses early-stage prospection activities, including the placement of targeted job advertisements (see point iii below), the identification of potential candidates in databases, or other informal approaches undertaken prior to the launch of a formal procedure42. AI systems may also be intended to be used in recruitment, for instance, for the screening of CVs, candidate sourcing, or applicant ranking.
AI systems may also analyse hiring data to predict which candidates are most likely to succeed based on skills and past performance or to help generate job descriptions. These may or may not fall within the scope of point 4 of Annex III depending on whether the system truly pertains to the domain of recruitment in a manner that impacts in substance career prospects or workers’ rights . Where that is the case, the system may be exempted from high-risk classification if it fulfils any of the conditions in Article 6(3) AI Act. To assess this, it is key to examine the AI system’s potential impact on prospective or actual candidates’ employment and career opportunities, or the risk of discrimination43. AI system generating job descriptions An AI system exclusively generates job descriptions based on a list of tasks to be carried out and a set of necessary qualifications and skills previously defined by a human recruiter . AI tools of this type can fall within the high-risk use case of point 4 of Annex III, ifit is within their intended purpose to play a role in the recruitment process in a manner that affects applicants’ prospects. However, the specific AI system referred above should be considered as performing a narrow procedural task (Article 6(3)(a)) as this specific use case poses only limited risks and does not meaningfully influence the application process and, hence, should be exempt from the high-risk classification. In contrast, an AI system of the same type, but where the AI system itself generates the necessary qualifications and skills based on a high-level description of the job position orthat has an additional function allowing it to substantially affect the application process by evaluating the candidates’ CVs against the job description that it created and to make recommendations of suitability on this basis cannot be considered as falling within the filter and should be classified as high-risk.
The notion of recruitment includes onboarding processes where the outcome of an AI system performing such a process may affect access to work, in particular where onboarding constitutes a relevant aspect of the recruitment process and not the first step after its conclusion ( e.g. when the employer performs checks on preliminarily hired workers) . Recruitment is concluded when the worker is understood to be hired and dismissal procedures must follow through specific channels foreseen to that effect.
ii. The notion of ‘selection’
Selection has two distinct meanings. On the one hand, it reflects the fact the high-risk area of Annex III, point 4 also encompasses ‘access to selfemployment’. While such processes might not be - ‘recruitment’ in the strict sense, natural persons who are selfemployed may also undergo a-selection process before they are contracted. Therefore, AI systems used in the selection of self-employed persons by their contracting parties should fall within the scope of Annex III, point 4(a). On the other hand, ‘selection’ also refers to the specific stage in recruitment where candidates are assessed, filtered, and chosen for advancement or hiring. It covers activities such as the analysis of applications, automated shortlisting, scoring, ranking, or assigning predictive scores of job suitability.
iii. Place targeted job advertisements
Although point 4(a) of Annex III, lists examples of selection, such as ‘analysing and filtering job applications’ and ‘evaluating candidates,’ this list should not be seen as being exhaustive44 . The explicit inclusion of targeted job advertisements in point 4(a) of Annex III follows from the fact that the placement of advertisements that are algorithmically tailored to reach certain groups of individuals determine who becomes aware of, and who is encouraged to apply for , specific vacancies. Where an AI system placing targeted job advertisements relies on profiling, it should always be considered high-risk, given the exclusion inArticle 6(3), last subparagraph, AI Act of AI systems based on profiling from benefiting from the filter mechanisms contained in that provision. Even where such an AI system does not involve profiling, the system should be considered high-risk, if it meaningfully conditions access to employment opportunities or can lead to a risk of discrimination.
A job advertisement must be understood to represent an active indication that an employer is seeking candidates for a concrete vacancy. Employer branding or generic advertisements related to the firm, which do not in practice relate to a vacancy, fall outside the use case listed in point 4(a) of Annex III.
Moreover, where the personalisation of the persons to whom an advertisement is shown results directly from non-discriminatory, objective and reasonable criteria45 that are inherent requirements of the job description, it may be excluded from the scope of ‘targeted job advertisements”. Likewise, the placing of job advertisements by AI system that are based exclusively on contextual elements (contextual targeting) , such as content of the location where the advertisement is placed (e.g., website) and that are not directly related to specific natural persons or groups of natural persons , should also be considered to fall outside the use case listed in point 4(a) of Annex III . In such a case, the target audience of the location will be the same as the relevant audience that the advertiser is trying to reach to fill a vacancy . However, the job advertisement can still be shown to anyone who accesses the location (i.e., there is no exclusion).
iv. Analyse and filter job applications
The terms ‘to analyse and filter job applications’ used in point 4(a) of Annex III should be understood broadly and not as limiting the use case to AI systems that always perform both functions simultaneously. A functional interpretation is required: where an AI system merely analyses applications in a descriptive sense (such as converting file formats or classifying degrees into standard categories without applying evaluative weight), it may be considered to perform a preparatory task within the meaning of Article 6(3)(d) AI Act. In contrast, AI systems that analyse job applications with an effect that may result in a restriction of the applicant in accessing certain job opportunities, for example by producing suitability scores, assessing compatibility of a candidate’s profile or experience with the selection criteria or job description , or producing competence profiles that condition shortlisting, fall within scope of the high-risk use case listed in point 4(a) of Annex III . In this case, ‘analysis” and ‘filtering” form part of a continuum: analysis that influences access to later stages operates de facto as filtering, and therefore falls within the scope of Annex III, point 4(a).
v. Evaluate candidates
The terms ‘to evaluate candidates’ used in point 4(a) of Annex III must also be understood broadly. Evaluation refers to any assessment of a candidate’s (including self-employed) suitability, merit or potential, whether through scoring, ranking, predictive modelling, or qualitative judgements derived from data inputs. The provision does not require that the evaluation carried out by the AI system directly determines the hiring outcome . It suffices that it appreciably influences the decision-making process, for example by laying out shortlists , prioritising certain applicants or scoring assessments within the context of the recruitment or selection process that may result in advantages for certain candidates . AI tools that merely present factual information without involving evaluative weight (such as software displaying unaltered employment history) may fall outside scope if they do not generate assessments of suitability.
a) Practical examples of AI systems falling within the high-risk use case of point 4(a)
- AI systems intended to be used as an automated job matching and ranking tool designed to assist internal and external recruiters in the recruitment and selection processes by analyzing and filtering candidates should be classified as high-risk. Such a system processes structured and unstructured data (e.g., CVs, skills, education, past placements, cognitive, social and emotional competencies) and compares it to job descriptions or historical hiring data to generate quantitative scores, rankings (e.g., ‘top 5 candidates’), and/or qualitative fit categories (e.g., ‘high fit,’ ‘low fit’). These outputs are used to shortlist candidates for internal and external recruiters, who retain discretion to review, override, or supplement algorithmic recommendations. However, the system’s rankings and scores serve as a primary input for decision-making. For example, it could lead to low ranking or exclusion of candidates of a certain gender or of candidates with disabilities. It therefore falls to be classified as high-risk pursuant to point 4(a) of Annex III.
- AI system intended to be usedfor candidate or contractor sourcing across online platforms A recruitment agency deploys an AI-enabled tool to search social media, professional sites or job boards, and its own CV databases. Recruiters input pre-defined criteria such as years of experience, technical skills, or educational background. The system p rocesses unstructured and structured data, identifies matching profiles, and generates a shortlist of potential candidates or contractors who are natural persons. This system falls within the use case of point 4(a) ofAnnex III, since it directly and meaningfully affects the recruitment and selection of natural persons by filtering and identifying candidates. None of the exceptions in Article 6(3) apply considering the tasks carried out by the AI system and its material influence on the outcome of decision-making.
- AI system ranking self-employed service providers An online platform designed to help consumers in finding a self-employed service provider to assist them with various tasks uses an AI-enabled job-matching tool that ranks service providers and determines which are presented to potential consumers. This system falls within the use case of point 4(a) of Annex III since it directly and meaningfully affects selection of natural persons by filtering and ranking potential service providers. Because its output shapes access to future assignments and may have a decisive impact on the careers of self-employed persons, along with their livelihoods and rights, none of the exceptions in Article 6(3) apply.
- AI-based vision monitoring system intended to be used for pilot recruitment An AI-enabled system is intended to be used to assess the visual capacities of pilot candidates. Using cameras and sensors the system monitors relevant metrics (e.g., visual acuity and reaction speed) then evaluates whether candidates meet the thresholds required for passenger flights. These results directly determine candidates’ eligibility for certain roles. This system falls within the use case of point 4(a) of Annex III, since it is used in recruitment and selection to evaluate candidates’ suitability (i.e.,in this case whether they are eligible to fly long-haul or short-haul flights). None of the derogations in Article 6(3) apply considering the nature of the tasks being carried out by the AI system and its material influence on the outcome of the decision-making.
- AI system intended to be used for scoring applicant answers in a recruitment process An AI system is intended to be used to evaluate written or oral responses46 given by job applicants during an online assessment. The system assigns each applicant a numerical score based on linguistic and substantive criteria, then generates a ranking used to determine who is invited to the interview stage. This system falls withi n the use case of point 4(a) of Annex III, since it is used in recruitment and selection. The AI performs core evaluative functions and meaningfully impacts the substance of the recruitment and selection. None of the exceptions in Article 6(3) AI Act apply, in particular, considering the fact that the AI system will materially influence the decision-making.
- AI system intended to be used to support apprenticeship recruitment at a commercial firm A firm uses an AI system to support the recruitment of apprentices. The system is applied to manage and process applications for apprenticeship positions, potentially by screening CVs, matching applicants to roles, or generating shortlists based on pre-defined criteria such as qualifications or age. Although apprenticeships include training elements, the AI is used within the context of hiring for a formal employment relationshipand therefore falls within the use case of point 4(a) of Annex III. None of the exceptions in Article 6(3) AI Act apply considering the nature of the tasks being carried out by the AI system.
- AI system intended to be used to place targeted job advertisements to specific users on social media The system collects and analyses a wide range of data navigation patterns and user characteristics. Advertisers provide targeting criteria, such as age, educational and professional background, or sector, and the system identifies and prioritises potential audiences. The system goes beyond merely matching requirements such as whether the candidate possesses certain professional accreditations or their geographical location. Since the system determines access to employment opportunities through targeted advertising, it falls within the use case of point 4(a) of Annex. Since the system’sfunctioning can materially affect which candidates are able to learn about job vacancies, thereby influencing their ability to pursue employment , it cannot benefit from the exceptions in Article 6(3) AI Act.
- AI system intended to be used by employment agencies to assign candidates to vacancies The system processes candidate data such as information in CVs, including skills, education history, and prior work experience. On the employer side, the system incorporates job requirements, occupational classifications, and labour market data. The system then generates recommendations by ranking candidates for specific vacancies or by suggesting suitable job postings to jobseekers. These outputs are used by cas eworkers in the agency, who may rely heavily on the system to manage large volumes of applications. Since the system evaluates candidates in the context of recruitment and access to employment opportun ities, it falls within the use case of point 4(a) of Annex III. Since the system’srecommendations have a direct bearing on which candidates are referred to employers and which vacancies are presented to jobseekers, it cannot benefit from the exceptions in Article 6(3) AI Act.
- AI system intended to be used to perform background checks on job applicants during recruitment The system aggregates and analy s es multiple types of data. Standard inputs include official records, such as education and professional certifications, employment history, social network history, and credit or financial data , where legally permissible. The system also incorporates open-source and online information . Employers or recruiters receive outputs in the form of composite risk scores, categories such as ‘low,” ‘medium,” or ‘high risk,” or specific alerts highlighting potential issues (e.g., unexplained employment gaps, flagged financial liabilities, or controversial online activity). These outputs can be used as filters in high-volume hiring, where candidates flagged as ‘high risk” may be deprioritized or excluded before a caseworker reviews their file. While human review is formally part of the process, in practice the system’s output heavily influences which candidates advance in the recruitment process, hence materially influencing the outcome of the recruitment process. Since the system materially influences career prospects and is used in the scope of recruitment , it falls within the use case of point 4(a) of Annex III . Since the system performs profiling, it cannot benefit from the exceptions in Article 6(3) AI Act.
b) Practical examples of AI system falling outside the high-risk use case of point 4(a)
- AI system intended to be used exclusively to identify non-inclusive or discriminatory wording in ad descriptions The system’sidentification logic is based on patterns and terms defined and reviewed by human experts, and the system flags potentially problematic phrases. At first glance, the system operates within the broad employment domain, but the job ad screening tool is not intended to be used for the recruitment or selection of natural persons and therefore falls outside the use case of point 4(a) of Annex III.
- AI system intended to be used for employer brand advertisements to present the company as an attractive place to work The advertisements highlight aspects , such as workplace culture, employee benefits, or career development opportunities, but are not tied to any specific job opening and are non-discriminatory. The system may optimise the placement of ads to reach broader or more suitable audiences, using metrics , such as website traffic, engagement rates, or general demographics. However, the system is limited to employer branding advertisements and does not concern specific job vacancies, and therefore falls outside the use case of point 4(a) of Annex III.
- AI system intended to be used for employer reputation monitoring A company uses an AI system to scan online platforms and social media for comments or reviews mentioning the employer, with the purpose of tracking the organisation’s reputation as a workplace and identifying general trends in how the company is perceived by potential applicants. No analysis or tracking of specific comments or specific users(which are anonymized at collection) are undertaken.Although the AI system is linked to recruitment more broadly, since employer reputation may affect how attractive the company appears to jobseekers, it is not intended to be used for the direct recruitment or selection of natural persons and therefore falls outside the use case of point 4(a) of Annex III.
- AI system intended to be used for employee onboarding support The system provides personalized information about company policies, training schedules, and answers common questions during their onboarding process. Since the system supports internal HR operations after hiring and does not influence the recruitment or selection of natural persons or affect in any manner their access to work, the system does not fall within the use case ofpoint 4(a) of Annex III. The system may fall within the use case of point 4(b) ofAnnex III, if it affects the terms of the work-related relationship or feeds into performance evaluations or monitors workers.
- AI system intended to be used for assisting candidates in tailoring their CV to specific open positions The AI system analyses the candidate’s CV along with the description of the open position provided by the candidate. Based on these elements, the AI tool recommends changes to the CV of the candidate with the aim of increasing the likelihood of selection for an interview. These recommendations are exclusively shared with the candidate. While the AI system may indirectly impact the success of the candidate in obtainin g the desired position, its use is initiated and managed by the candidate, outside of the potential employer’s control and occurs outside of the recruitment and selection process. It therefore falls outside the use case of point 4(a) of Annex III.
- AI system intended to be used for assisting candidates in finding the best available position The system is used by candidates to analyse both job vacancies and information provided by the candidate covering their skills , experience, and professional expectations . Based on these elements, the system ranks job vacancies and recommends the most suitable positions for the candidate. These recommendations are exclusively shared with the candidate. While the AI system may indirectly impact the success of the candidate in obtaining the desired position, its use is initiated and managed by the candidate, outside of the potential employe r’s control and occurs outside of the recruitment and selection process. It therefore falls outside the use case of point 4(a) of Annex III.
c) Practical examples of AI systems falling within the use case but exempted under the filter mechanism of Article 6(3)
- AI system intended to be used for verifying professional accreditations of applicants from official registries (e.g. bar membership number provided by each candidate with official registries maintained by national or regional bar associations). The outputs of the system are factual: either ‘confirmed” or ‘not confirmed,” with any issues flagged for manual follow-up by HR or legal staff. While the system is applied during recruitment, it does not filter or assess candidates beyond confirming a credential. Because the system does not materially shape candidate selection beyond this binary factual confirmation and performs a clearly defined and limited function, it can be considered to perform a narrow procedural task.It therefore falls under the exception listed in Article 6(3)(a) AI Act.
- AI system intended to be used for recognising and organising information in CVs received for an open position and to organi s e it within an internal database that can be searched by recruiters. Such a system will be used for a clearly defined and limited function and does not have a material impact on the recruitment or selection of the applicant and. It therefore falls under the exception for AI systems intended to perform narrow procedural taskslisted in Article 6(3)(d) AI Act.
- AI system intended to be used for scheduling interviews The system is used to automate the coordination of interview appointments with job applicants. The system operates by integrating with the company’s calendar tools and the availability preferences provided by each candidate. It proposes interview time slots based on mutual availability and logistical constraints, such as time zone differences or maximum daily meetings per recruiter. It also sends reminders to recruiters and candidates a few days before the interview reminding them of the date, time and location of th e interview. The system includes a function allowing candidates to indicate specific accessibility needs, such as requests for sign language interpretation, extended interview duration, or alternative formats for communication. These preferences are then automatically i ncorporated into the scheduling process to ensure the appropriate arrangements are made. While the scheduling system operates in the recruitment context, it does not perform any function that contributes to the selection of candidates. Since the system’s purpose is purely logistical and it does not initiate nor influence employment-related assessments, it falls within the exception for narrow procedural tasks listed in in Article 6(3)(a) AI Act.
- AI system intended to be used to check human patterns in hiring The AI system is used to audit past hiring decisions by analysing anonymized recruitment data, including CV scores, interview notes, and hiring outcomes. The system employs statistical modelling to detect potential biases or inconsistencies . It does not participate in ongoing recruitment or influence current candidate evaluations, nor does it assess or make decisions based on identified or identifiable recruiters’ or applicants’ personal characteristics. Its role is purely retrospective based on anonymized information and for already completed human assessments. Although it falls within the use case of point 4(a) of Annex III , the system benefits from the exception listed in Article 6(3)(c) AI Act for AI systems intended to detect decision-making patterns or deviations from prior decision-making patterns without replacing or influencing previously completed human assessments.
- AI system intended to be used for sending personalized acknowledgement emails acknowledging receipt of applications with the name and pronouns of the candidate who applied. Such a system does not influence in any way the likelihood of the candidate being selected for the position. It falls within the exception for AI systems intended to be used to perform narrow procedural tasks listed in Article 6(3)(a) AI Act.
- This understanding is consistent with the case law of the Court of Justice of the European Unionwhichconfirms that EU equality law applies already at the stage of access to employment. In Case C-54/07,Centrum voor gelijkheid van kansen en voor racismebestrijding v Firma Feryn NV ,EU:C:2008:397and subsequently in Case C-507/18,Associazione Avvocatura per i diritti LGBTI ,EU:C:2019:922the Court held that discriminatory statements by an employer concerning prospective employees constituted discrimination, even in the absence of a formal recruitment procedure.
- Forinstance, job descriptions or advertisements may have a strong influence on who will be successful in the recruitment procedure, especially if they set out the qualifications required for the job and/or the selection criteria. Drafting them based on biased data, such as past recruitment patterns, can increase the risk of discrimination for disadvantages groups.
- The applicationof point 4(a) is limited by the intended purpose of the AI system, as such AI systems intended to select or evaluate tender offers or responses to requests for proposals will not be classified as high-risk just because they are theoretically capable of se lecting or evaluating an offer from a self-employed sole proprietor. However, to fall out of scope of the high-risk classification, the use of the AI system may not result in a de facto selection (or exclusion) of the self-employed person (instead of the proposal) nor an evaluation of the self-employed person (instead of the proposal).
- This would be the caseif the requirements stem from agenuine and determining occupational requirement whose absence may make it impossible for the candidate to perform the functions stated in the job description, and not when it is about a preferred characteristic, or a specific profile that while desirable does not have a direct and objectivelink to the required qualifications/skills.In assessing the mentioned non-discriminatory, objective and reasonable criteria, it is important to consider the EU non-discrimination acqui s, including Article 4 of Directives2000/78/EC and 2000/43/ECand Article 14(2) of Directive 2006/54 /ECof the European Parliament and of the Council of 5 July 2006 on the implementation of the principle of equal opportunities and equal treatment of men and women in matters of employment and occupation (recast) , OJ L 204, 26.7.2006, pp. 23–36.
- Oral responses could be evaluated, for example, through an AI system capable of processing video-and audio and interactingwith the applicant through an avatar.