Point 4(b) of Annex III classifies as high-risk AI systems intended to be used to make decisions affecting terms of work-related relationships, the promotion or termination of work-related contractual relationships, to allocate tasks based on individual behaviour or personal traits or characteristics, or to monitor and evaluate the performance and behaviour of persons in such relationships.
These AI systems concern situations taking place after the recruitment and selection process, and encompass the exercise of certain managerial prerogatives and the organisation of work throughout the duration of the employment or contractual (for self-employed) relationship up until its termination.
AI systems in workplace management, although capable of enhancing efficiency and consistency, can also exert decisive influence over workers’ livelihoods, future career prospects and rights, and can lead to risk of discrimination . Using AI to manage or make decisions about the workforce is not unlawful per se under the AI Act, but its risks, in particular concerning transparency, responsibility and structural power imbalance justifies classifying certain AI systems as high-risk system under point 4(b) of Annex III and subject ing them to the requirements and obligations provided in Chapter III AI Act47.
i. AI systems intended to be used to make decisions affecting terms of work-related relationships
The term ‘decisions’ used in point 4(b) of Annex III should be understood in a functional, rather than a formalistic sense (focusing on substance over form). In the context of employment relations, a decision is any act or omission48 attributable to the employer that produces material effects on the worker’s contractual position or relationship or the modalities of its work. It is not limited to unilateral modifications of contract or relationship in the strict sense, but extends to operational determinations that materially affect the enjoyment by the worker of essential rights and obligations under the contract or relationship. The terms ‘intended to be used to make decisions ’ should be understood as including both cases where the AI system takes the decision in an automated manner and cases where a human operator formally takes a decision, but significantly relies on the output of the AI system.
As regards the words ‘terms of work-related relationships’used in point 4(b) of Annex III , these are the fundamental elements to the work contract or relationship that, if they were not present or were different, might have resulted in one or the two parties not consenting to the contract or relationship. This encompasses those conditions that define the balance of reciprocal rights and duties of employer and employee, including working conditions, such as pay entitlements, working time arrangements, leave rights, and resting periods. The EU acquis (including the Directive on Transparent and Predictable Working Conditions)49 and Member State labour law, and should be considered when ascertaining w hich elements fall within the scope of ‘terms of work-related relationships’. Decisions on such terms are not limited to unilateral modifications of the contract or relationship in the strict sense, but extend to determinations that materially affect the enjoyment of the rights and obligations of the worker arising from the ab ovementioned fundamental elements (e.g., denial of a leave request or a significant change to regular working hours).
The term ‘work-related relationships’ includes employment relationships but also self-employment, including those contractual relationships of self-employed persons providing services through platforms (see Recital 57) and other solo self-employed persons who are in a situation comparable to workers.
The training entitlement provided by the employer is an ‘essential aspect of the employment relationship’ under Article 4(2) of Directive Transparent and Predictable Working Conditions50 . However, optional or ancillary training with no entitlement and no legal or collectively agreed obligation does not automatically fall within the use case of point 4(b) of Annex III. For the purpose of high-risk classification , training acquires decisive legal significance where completion or assessment is made a condition for advancement or promotion , or it is otherwise linked to remuneration, grade or continued employability and keeping the job . In such cases, its relevance becomes a determinant of career progression or continued employability. If an AI system evaluates performance in training modules , and such an evaluation feed s into performance appraisals, continued employability, or eligibility for higher-grade positions, it should be considered to fall within the use case of point 4(b) of Annex III.
However, it is equally important to recognise that not every day-to-day managerial or operational organisational adjustment amounts to ‘decisions affecting terms of the workrelated relationship’ - within the meaning of point 4(b) of Annex III. To fall withi n that use case, the decision needs to reach a threshold of significance. A functional interpretation of point 4(b) of Annex III in light of Recital 57 AI Act cannot be stretched to include all day-to-day or operational decisions. While influencing the factual conditions under which work is carried out, decisions that do not materially alter the rights or obligations arising from the underlying contract or work relationship do not constitutes ‘decisions affecting terms of the work-related relationship’ within the meaning of point 4(b) of Annex III. This includes individual operational decisions taken by managers on a day-to-day basis within the limits of those ‘terms of the workrelated relationship’, such as allocation of-office space and break or lunch time within the context of an assigned shift (as long as no changes occur to the total break and lunch time enjoyed by the worker).
ii. Promotion and termination of work-related contractual relationships
The terms ‘promotion and ‘termination of work-related contractual relationships ” used in point 4(b) of Annex III should be understood as any act or omission attributable to the employer that produces effects on the worker’s position by, respectively, elevating grade, responsibilities, remuneration or access to career development, or any element that would contextual ly lead to an effective form of promotion; and any act or omission attributable to the employer which ends the employment relationship or other work-related contractual relationship or deprives the worker of the essential rights and benefits attached to it51.
As regards ‘promotion’, EU employment law has recognised that decisions concerning career advancement constitute legally significant determinations52 . Similar considerations could apply when interpreting that term as used in the context of point 4(b) of Annex III. As regards ‘termination’, it could be understood in a similar manner to that used in Union law, namely as the final, irrevocable endpoint of the work-related contractual relationship decided by the employer or contracting party. Other forms of removal from work, if not definitive, may qualify for material modifications of the terms of the work-related relationship, and thus may fall within that notion and within the notion of ‘termination’ as used in point 4(b) of Annex IIITermination includes . dismissal, redundancy, and non-renewal of fixed-term contracts when the decision to not renew is AI driven, or any act producing equivalent effects. As previously stated, omissions also constitute elements on which termination can be claimed.
Although non-standard work arrangements and particularly platform work might blur the distinction between ‘suspension’, ‘deactivation’ and ‘termination’, where an AI system definitively suspends a platform account, the effect is to deprive the individual of access to their professional activity, irrespective of contractual form. Lengthy or frequent suspensions of access to the platform may produce similar effects. More generally, the functional equivalence between dismissal of an employee and termination or suspension of a contract with a nominally self-employed person53 confirms that determinationson contract termination, by act or omission, may amount to restrictions on access to self-employment covered by point 4(b) of in Annex III.
While termination is the decision with the most far-reaching consequences in the employment setting, the filter mechanism under Article 6(3) AI Act may still apply to AI systems falling within the use case of point 4(b) of Annex III . For example, where an AI system merely keeps a record and flags to the appropriate manager the imminent lapsing of a fixed-term contract before its end point, free from discretionary consideration , evaluative judgment, and without suggesting a decision, the outcome follows directly from the contract itself and not from a decision of the AI system. In such instances, the system does not alter the rights of the worker in substance and therefore should not be considered as high-risk.
iii. AI systems intended to be used to allocate tasks based on individual behaviour or personal traits or characteristics
The terms ‘intended to be used to allocate tasks based on individual behaviour or personal traits or characteristics’ in point 4(b) of Annex III should be read in conjunction with Recital 57 AI Act, which clarifies that such systems may have a significant impact on future career prospects, livelihood and rights of the worker. A n allocation of tasks is not just an operational arrangement , but may be a determinant of professional development. For instance, the distribution of more complex, visible or lucrative assignments may open pathways to advancement, while the allocation of repetitive or less valued tasks may in practice preclude promotion and potentially affecting current or future income.
If an allocation of tasks is driven by taking into account individual behaviour or personal traits, the risk of structurally disadvantaging certain groups is increased, in addition to dignity , data protection, and privacy concerns. This is also reflected in Recital 57 AI Act, which clarifies that such systems ‘may perpetuate historical patterns of discrimination, for example against women, certain age groups, persons with disabilities, or persons of certain racial or ethnic origins or sexual orientation’.
The scope of point 4(b) of Annex III should be understood to cover any individual task allocation ‘based on individual behaviour or personal traits or characteristics’, which may include data based on personality testing, behavioural scoring, or inferred productivity traits.
Behavioural indicators, such as punctuality, responsiveness to customer requests, reliability metrics, or performance ratings, fall within the scope of ‘personal traits or characteristics ’ . AI systems that, for instance, withhold access to delivery slots from workers with lower acceptance rates or slower response times or that rank self-employed external lawyers from a pool based on response time and performance ratings provided by internal company counsel condition work opportunities on such behavioural profiling.
The dividing line between individual behaviour or personal traits and characteristics, and those that are not, is whether the attribute in question pertains to the worker’s personal traits or social identity, rather than to neutral , objective and external factors that preclude the execution of the required tasks, such as unavailability or inadequate location54.
Task allocation systems that rely only on objective, neutral, and external criteria tied directly to the requirements of a job fall outside the scope of the use case listed in point 4(b ) of Annex III. Objective, neutral, and external criteria cannot be proxy for nationality or ethnicity. These include attributes that do not pertain to the worker’sindividual behaviour, personal traits or social identity, and that are also not covered by the EU non-discrimination acquis. For example, an AI system allocating tasks based on availability as indicated by workers in scheduling tools or distributing assignments according to possession of a required professional accreditation, such as a forklift permit or professional bar accreditation, would not constitute reliance on personal traits within the meaning of point 4(b) of Annex III. Similarly, AI systems assigning work based on geographical proximity to a delivery location should generally not be classified as high-risk.
iv. AI systems intended to be used to monitor and evaluate the performance and behaviour of persons in such relationships
AI systems can considerably extend an employer’s means to supervise, monitor and evaluate the performance and behaviour of workers or other persons in work-related relationships. The deployment of such systems is not per se unlawful: managerial oversight is inherent in the employment relationship and employers may legitimately seek to enhance productivity, protect property, or ensure workplace safety. However, such deployment can raise concerns of privacy and personal data protection, but also concerns related to health, safety, discrimination, and dignity in the workplace.
The terms ‘to monitor and evaluate the performance and behaviour of persons in such relationships’ used in point 4(b) of Annex III must be understood as an autonomous high-risk category. The conditions of ‘monitor’ and ‘evaluate’ are alternative and they cover AI systems intended to monitor ‘or” evaluate workers’ performance or behaviour systematically , including during probationary periods55, regardless of whether this affects a formal term contractually established, such as pay or working time.
The notion of ‘performance’ includes quantitative productivity measures, such as the number of tasks completed or output speed. It also includes qualitative productivity measures such as the quality of work56 . For example , AI systems that automatically rate workers on the number of deliveries per hour fall under that notion. So should qualitative indicators, such as feedback scores, performance assessments, appraisals or rankings generated by AI systems that feed directly into employee files.
The notion of ‘behaviour’ may encompass punctuality, adherence or potential adherence to workplace rules (e.g., employee risk profiling) , responsiveness to clients, patterns of interaction with colleagues , and engagement with trade unions57 . An AI system that assesses (including potentially downgrading) workers for low acceptance rates of shifts or for declining tasks should be considered to evaluate behaviour.
AI systems deployed by employers exclusively to meet external legal or regulatory obligations (for instance transaction logging to comply with financial market abuse rules) should not be considered to monitor or evaluate an individual’s performance or behaviour for employment purposes, as long as they are also not used for this purpose. Likewise, out of scope of the notions monitoring or evaluating an individual’s performance or behaviour for employment purposes are AI system s deployed purely and exclusively for medical and safety reasons or to protect property and company assets. These AI systems will be not covered by the use case listed in point 4(b) of Annex III.
AI systems that monitor the behaviour or performance of workers in order to support them in the performance of their tasks without being designed to pressure workers to achieve additional productivity, such as by preventing information or products being sent to incorrect recipients though automated flagging, should not be classified as high-risk, provided the AI system only flags potential issues to the worker, not to the employer or any other person within the employer organisation. Similarly, AI systems that convey data on performance, such as systems that aggregate data from successful sales transactions and compare them across company average and only notify them to the worker, should not be classified as high-risk.
a) Practical examples of AI systems falling within the high-risk use case of point 4(b)
- A retail/logistics company deploys an AI - enabled scheduler to assign shifts, rest periods and on-call windows Inputs provided to the AI system include , among others, behavioural/performance signals (punctuality, non-show history, acceptance/decline rates for offered shifts, customer/manager ratings). The optimiser ranks workers for each slot and auto-allocates shifts, allocating the highest performing workers to the most important shifts . Workers ranked lower receive less important and possibly fewer shifts, lower variable pay, and less predictable patterns. The optimiser also downgrades priority if a worker declines ‘priority offers”. The AI system will evaluate the performance and behaviour of workers and assign tasks based on this evaluation and, as such, falls within the scope of point 4(b) of Annex III. Use of this AI system will affect workers’ remuneration, their access to priority tasks and possibly progression. As such, the system cannot benefit from the exceptions in Article 6(3) AI Act.
- An online tutoring platform deploys an AI system to manage self-employed teacher accounts Inputs include student satisfaction ratings, lesson completion rates, and punctuality logs from the videoconferencing system. The system aggregates these into a ‘tutor performance score.”If the score falls below 4/5 for three consecutive weeks, the system automatically suspends the tutor’s account for a month. If low ratings persist for another review period, the account is permanently deactivated. The practical effect is that tutors can no longer accept lessons, which ends their work relationship with the platform. The system’s effect is the functional termination of the work relationship. The term ‘termination’ used in point 4(b) of Annex IIIis to be understood in substantive terms, covering any decision that deprives workers of continued engagement.
- AI system for workload allocation among associates in a law firm The system ingests data on billing hours, turnaround times on prior assignments, responsiveness to emails, and voluntary participation in firm activities (e.g. training, client development). These behavioural inputs are aggregated into a score, which serves as a basis for the allocation of client matters. Since the system allocates tasks based on the individual behaviour of associates , materially affecting both present work conditions and long-term career trajectories, it falls within the use case of point 4(b) of Annex III.
- AI system for pricing and pay determination in platform work A ride-hailing platform deploys an AI system to dynamically set driver compensation. Inputs include real-time demand, driver acceptance rates, passenger ratings, and average completion times. The system calculates per-ride fares and adjusts individual drivers’ pay multipliers. For example, drivers with consistently lower passenger ratings or slower completion times receive a reduced pay coefficient, while highly rated drivers earn bonuses. These adjustments are made automatically and applied in real time, directly affecting income. Human oversi ght exists formally, but in practice pay levels are determined exclusively by the algorithm and appeals rarely alter outcomes. Since remuneration decisions fall within the terms of a work-related relationship, the system falls within the use case of point 4(b) of Annex III58.
- AI system intended to be used for assignment of civil servants to posts A public administration uses an AI system to allocate successful candidates (e.g. teachers or doctors) to specific posts after they have passed a competition exam. The system takes into account test scores, geographical preferences, availability of vacancies, and seniority. It generates final assignment lists that determine which candidate goes to which post. Appeals are possible, but in practice th e AI system’s allocation output is decisive, as administrative services rely on the system to finalise placements. Since the system is used for the allocation of roles in the public sector, directly and meaningfully determining working conditions and career paths of civil servants, it falls within the use case of point 4(b) of Annex III.
b) Practical examples of AI systems falling outside the high-risk use case of point 4(b)
- AI system for tracking delivery operations A courier and shipping company uses an AI system to track shipping data and detect potential mistakes in labelling or issues with distribution. The system assesses the content of labels, monitors parcels in transit and detects potential deviations between the expected and actual route taken by the parcel in transit to its destination. If the AI system detects deviations, it will inform only the employee in charge. The employee in charge will then adopt the necessary procedures for correction (if any a re needed ). As the system’s purpose is ensuring smooth workflow operations and contractual compliance, supporting workers in their tasks rather than evaluating workers’ performance, the monitoring activity is incidental and falls outside the use case of point 4(b) of Annex III59.
- AI system for training performance evaluation A company uses an AI system to assess employees’ progress in voluntary training modules by analyzing quiz results and course completion data. The system only provides feedback and learning recommendations to the employee to support individual development. Its function is limited to measuring learning outcomes within a training context, not to appraising job performance, determining promotions, or shaping managerial decisions.
- AI system for office space optimization A company uses an AI-enabled booking tool to optimise desk allocation in a hybrid workplace by matching employee reservations with available desks or meeting rooms. While the system organises physical workplace resources, it does not allocate tasks, determine promotions, evaluate workers, or affect employment rights.
- AI system for corporate travel planning A company uses an AI application to optimise business travel arrangements, suggesting cost-effective flights, hotels, and itineraries for employees required to attend meetings or conferences. The AI application’s suggestions do not prevent the employee from making alternative choices or even fully manually booking as long as it does not breach objective rules of company policy (e.g., maximum prices). The system supports logistics and budgeting, but does not determine the distribution of work, performance evaluations, career progression or termination decisions.
- AI system suggesting delivery areas to platform couriers A food delivery platform deploys an AI system that recommends geographic areas where demand is expected to be higher, based on factors such as weather, past orders, and real-time traffic data. Couriers receive notifications such as ‘high demand expected in Area X between 1-2 pm’ but they are free to ignore the suggestions. There are no penalties for declining recommendations, no specific registry of couriers that accepted or rejected the recommendation, and the system does not reduce their access to future tasks. The recommendation tool provides only optional, non-binding supporting information and does not represent task allocation on the basis of individual behaviour.
c) Practical examples for AI systems falling within the use case but exempted under the filter mechanism of Article 6(3)
- AI system compiling performance records for reporting A manufacturing company uses an AI system to automatically collect from the company’s timekeeping application and organize employee monthly attendance records into structured reports and dashboards that the AI system sends to managers at a fixed date . The system does not monitor employees’ attendance, generate performance scores, make recommendations, or flag employees, it merely consolidates existing informationthat was entered by employees in the timekeeping application. Managers continue to perform qualitative evaluations independently. Although the system operates in employment monitoring and falls under the use case of point 4(b) of Annex III, it does not engage in monitoring nor does it make evaluative judgments and falls under the exception for AI systems intended to perform a preparatory task listed in Article 6(3)(d) AI Act.
- AI system refining human-drafted promotion evaluations A consultancy firm uses an AI-enabled writing assistant to refine managers’ promotion reports after the evaluations are fully completed. Managers decide on recommendations, draft justifications, and assign ratings based on company criteria before the AI system improves language clarity, ensures consistency with corporate style, and flags potentially biased wording. At the end of the process, the manager is required by internal policy to double-check the revised evaluations. The system does not change outcomes, scores, or create new content; its role is supportive. Although the system relates to promotion decisions, it does not influence decision-making and falls within the exception for AI systems intended to improve the result of a previously completed human activity listed in Article 6(3)(b) AI Act.
- The AI Act rules on classification of high-risk AI systems and the relevant requirements and obligations are directly and horizontally applicable to operators. However, without prejudice to the full and uniform application of the rules under the AI Act,even in cases where their AI systems fall within the filter, operators should be aware that they may be subject to obligations arising from other EU legal instruments, such as the GDPR. For example, in the context of platform work, the Platform Work Directive is of particular importance.
- The functional reading of ‘decision” in employment law encompasses not only formal acts but also omissions producingbinding effects on the worker’s position. A ‘failure to promote” is a legally relevant determination: see Case C-409/95,Marschall,EU:C:1997:533, CaseC-407/98, Abrahamsson and Anderson v Fogelqvist, EU:C:2000:367.
- Including the essential aspects of the employment relationship referred in Article 4(2). Directive (EU) 2019/1152 of the European Parliament and of the Council of 20 June 2019 on transparent and predictable working conditions in the European Union, OJ L 186,11.7.2019, p. 105.
- Informationon the training entitlement, if any, shall be provided by the employer individually to the worker in the form of a document within one month of the first working day.
- Constructive dismissal is not an autonomous concept of Union law, but it is widely recognised in Member States’ labour law. For example, in Spain, Article 50 of the Estatuto de los Trabajadores allows an employee to terminate the contract with entitlementto compensation where the employer’s conduct entails a ‘serious breach” of obligations, such as unilateral downgrading of conditions or harassment.
- Directive 2006/54/EC, identifies ‘promotion” as a protected domain alongside, inter alia, access to employment, self-employment and training .. Also SeeCase C-409/95, Marschall, EU:C:1997:533, Case C-407/98, Abrahamsson and Anderson v Fogelqvist, EU:C:2000:367 , both concerning the application of sex equality rules to promotion decisions. While centred on gender discrimination, these cases illustrate the Court’s broader approach in treating promotion as a legally significant employment determination. InCase C-715/20,K.L,EU:C:2024:139the Court held that fixed-term workers must be afforded reasons for termination equivalent to those required for permanent employees, thereby recognising the possibility of a functional equivalence of non-renewal and dismissal.
- Union equality law expressly protects not only ‘access to employment” but also ‘access to self-employment” and to occupation more broadly: see, as referred before,Directive 2000/78/EC and Directive 2000/43/EC.
- Characteristics such as those protected by the fundamental right of non-discrimination under Article 21 of the Charter (sex, age, disability, etc.) are clearly covered, among others.
- Recital 57AI Act.
- AI systems designedto measure physiological indicators such muscle movement, fatigue, brain waves or heart rate may be considered as high-risk when the abovementioned physiological indicators are used to monitor or evaluate performance. The same will not be the case if the use of the AI system is related only to purely medical or safety reasons.
- Without prejudice to the lawfulness of or the monitoring or evaluation of the behaviour under applicable Union and national laws.
- An AI system may be used to dynamically set prices for end users without necessarily falling within the scope of Annex III, point4(b) of the AI Act. However, for this to occur the AI system may not structure or influence employment conditions (such as pay).
- Provided that it is not intended to be used also for performance assessment.