OIT - AI in human resource management The limits of empiricism

X AI in human resource management The limits of empiricism Authors / Janine Berg, Hannah Johnston

November / 2025 ILO Working Paper 154© International Labour Organization 2025 Attribution 4.0 International (CC BY 4.0) This work is licensed under the Creative Commons Attribution 4.0 International. See: https:// creativecommons.org/licenses/by/4.0/. The user is allowed to reuse, share (copy and redistribute), adapt (remix, transform and build upon the original work) as detailed in the licence. The user must clearly credit the ILO as the source of the material and indicate if changes were made to the original content. Use of the emblem, name and logo of the ILO is not permitted in connection with translations, adaptations or other derivative works. Attribution – The user must indicate if changes were made and must cite the work as follows: Berg, J., Johnston, H. AI in human resource management: The limits of empiricism. ILO Working Paper

154. Geneva: International Labour Office, 2025.© ILO.

Translations – In case of a translation of this work, the following disclaimer must be added along with the attribution: This is a translation of a copyrighted work of the International Labour Organization (ILO). This translation has not been prepared, reviewed or endorsed by the ILO and should not be considered an official ILO translation. The ILO disclaims all responsibility for its content and accuracy. Responsibility rests solely with the author(s) of the translation. Adaptations – In case of an adaptation of this work, the following disclaimer must be added along with the attribution: This is an adaptation of a copyrighted work of the International Labour Organization (ILO). This adaptation has not been prepared, reviewed or endorsed by the ILO and should not be considered an official ILO adaptation. The ILO disclaims all responsibility for its content and accuracy. Responsibility rests solely with the author(s) of the adaptation. Third-party materials – This Creative Commons licence does not apply to non-ILO copyright materials included in this publication. If the material is attributed to a third party, the user of such material is solely responsible for clearing the rights with the rights holder and for any claims of infringement. Any dispute arising under this licence that cannot be settled amicably shall be referred to arbitration in accordance with the Arbitration Rules of the United Nations Commission on International Trade Law (UNCITRAL). The parties shall be bound by any arbitration award rendered as a result of such arbitration as the final adjudication of such a dispute. For details on rights and licensing, contact: rights@ilo.org. For details on ILO publications and digital products, visit: www.ilo.org/publns.

ISBN 9789220428610 (print), ISBN 9789220428627 (web PDF), ISBN 9789220428634 (epub), ISBN 9789220428641 (html). ISSN 2708-3438 (print), ISSN 2708-3446 (digital) https://doi.org/10.54394/NMSH7611

The designations employed in ILO publications, which are in conformity with United Nations practice, and the presentation of material therein do not imply the expression of any opinion whatsoever on the part of the ILO concerning the legal status of any country, area or territoryor of its authorities, or concerning the delimitation of its frontiers or boundaries. See: www.ilo. org/disclaimer. The opinions and views expressed in this publication are those of the author(s) and do not necessarily reflect the opinions, views or policies of the ILO. Reference to names of firms and commercial products and processes does not imply their endorsement by the ILO, and any failure to mention a particular firm, commercial product or process is not a sign of disapproval. Information on ILO publications and digital products can be found at: www.ilo.org/researchand-publications ILO Working Papers summarize the results of ILO research in progress, and seek to stimulate discussion of a range of issues related to the world of work. Comments on this ILO Working Paper are welcome and can be sent to berg@ilo.org.

Authorization for publication: Caroline Fredrickson, Director, Research Department ILO Working Papers can be found at: www.ilo.org/research-and-publications/working-papers Suggested citation: Berg, J., Johnston, H. 2025. AI in human resource management: The limits of empiricism, ILO Working Paper 154 (Geneva, ILO). https://doi.org/10.54394/NMSH761101 ILO Working Paper 154

Abstract The rapid integration of artificial intelligence (AI) into Human Resource Management (HRM) is transforming how organizations recruit, manage, and evaluate their workforces. While proponents champion AI as a means to enhance efficiency, reduce bias, and align HR practices with strategic business goals, this paper argues that such optimism is misplaced. Drawing on a critical review of AI's application across four core HRM functions—recruitment, compensation, scheduling, and performance management—this paper identifies significant risks and limitations arising from the fundamental structure of AI systems. Central to the analysis is a three-parameter framework for assessing AI tools: their objective, the data they rely upon, and how they are programmed. The paper shows that across HR functions, AI systems frequently operationalize reductive or poorly aligned objectives, rely on low-quality or biased data, and are programmed in non-transparent ways that undermine their usefulness. These structural shortcomings not only undermine the effectiveness of AI systems but also introduce legal, ethical, and practical risks for firms and their workers.

Keywords: artificial intelligence, human resource management, data analytics, algorithmic management About the authors

or biased data, and are programmed in non-transparent ways that undermine their usefulness. These structural shortcomings not only undermine the effectiveness of AI systems but also introduce legal, ethical, and practical risks for firms and their workers.

Keywords: artificial intelligence, human resource management, data analytics, algorithmic management About the authors Janine Berg is Senior Economist and Head of the Effective Labour Institutions Unit in the Research Department of the ILO. Since joining the ILO in 2002, she has conducted research on the economic and social effects of labour laws as well as provided technical assistance on policies for generating jobs and improving working conditions. She is the author of several books and numerous articles on employment, labour market institutions and the digital transformation of work.

Hannah Johnston is an Assistant Professor in the School of Human Resources Management at York University in Toronto, Canada, specializing on the digitalization of work. Prior to joining York, Hannah was a postdoctoral fellow at Northeastern University in Boston and also worked professionally at the International Labour Organization. Hannah has a longstanding interest in the platform economy and is a collaborator with Oxford University’s Fairwork Project. Her recent publications can be found in journals including Industrial and Labor Relations Review, Work and Occupations, and the International Labour Review.02 ILO Working Paper 154 Abstract 01 About the authors 01 X Introduction 04 X 1 How did we get here? The rise of “people analytics” and managing workers through data 05 X 2 The workings of AI systems: Objective, data and programming 08 Recruitment 09 Sourcing 10 Screening, Interviewing and Selection 11 Compensation 13 Scheduling work 16 Performance management 19 X 3 AI in HRM: Unbridled optimism 23 X 4 What is an HR manager to do with AI? 26 X Conclusion 28 References 29 Acknowledgements 36 Table of contents03 ILO Working Paper 154 List of Figures Figure 1. McKinsey infographic on the benefits of people analytics 06

X 4 What is an HR manager to do with AI? 26 X Conclusion 28 References 29 Acknowledgements 36 Table of contents03 ILO Working Paper 154 List of Figures Figure 1. McKinsey infographic on the benefits of people analytics 06 Figure 2. The workings of AI systems 0904 ILO Working Paper 154 X Introduction Motivated by a desire to more efficiently and effectively manage people in organizations, HR managers are using the programming and analytical capacities of AI to fulfill key HR functions – including selection and recruitment of personnel, compensation determination and structure, performance review and evaluation, and the organization of working time. Largely absent from the rush to integrate AI, however, has been a comprehensive assessment of whether, and under what circumstances, AI is useful for the management of people within organizations. Yet despite the lack of assessment, the ‘AI for HR’ industry and the adoption of these tools and systems by individual firms and organizations is burgeoning. This paper presents a framework for understanding and evaluating the potential benefits and possible risks or harms presented by AI systems in workforce management. Following a section of the paper documenting the historical context of the Human Resources field that has given rise, first to people analytics and then to AI, the paper presents a framework based on three inter-related parameters that can help assess the quality, legality, and suitability of AI systems used in the field. These are: (1) the system objective, (2) the data it is built on and relies on, and (3) how the AI system is programmed. Drawing on existing literature about how AI is being used for workforce management, the paper applies the three-parameter framework to map the contours of AI use relative to four key Human Resource Management functions where adoption of AI technologies has been prominent: recruitment, compensation, scheduling, and performance management. Our discussion section, “Unbridled optimism” views the disciplinary and occupational history of HR alongside the findings on HR’s emergent use of AI. We argue that the search for occupational legitimacy by HR professionals has fostered a preoccupation with numeracy and positivism

Our discussion section, “Unbridled optimism” views the disciplinary and occupational history of HR alongside the findings on HR’s emergent use of AI. We argue that the search for occupational legitimacy by HR professionals has fostered a preoccupation with numeracy and positivism that has provided fertile ground for the ‘evidence-based solutions’ that AI systems purport to offer. This tendency towards positivism, we contend, is likely to result in widespread adoption of AI under circumstances that create risks for workers, liabilities for firms, and costs for society. Given the seemingly inevitable transformation of work due to AI, we argue for the need for HR professionals to improve their understanding of the workings of AI systems so that they can better judge their potential and limitations, and that this is best achieved when they participate in the design of systems that are implemented in their workplaces.05 ILO Working Paper 154 X 1 How did we get here? The rise of “people analytics” and managing workers through data

Human Resource Management emerged in the 1950s as a distinct field of study and practice concerned with the management of people in organizations. Carved out from the broader discipline of Industrial Relations, which has historically examined labour relations in the context of unions and collective bargaining – or collective employment relations – Human Resource Management is most associated with an employer’s relations with individual employees (Kaufman 2001). As practitioners, HR managers are typically engaged in facilitating a range of personnel functions – including recruitment and selection, compensation, scheduling and performance management and promotion – to achieve the organization’s goals. Since the 1980s, the dominant paradigm guiding HR managers has been ‘strategic HRM’(Paauwe and Boon 2018). Building on HR’s origins and the foundational concept of scientific management, strategic HRM aims to link firm performance to the specific methods and practices deployed within the firm to manage their workers. When first introduced, this approach represented a significant shift in firm management. Prior to this shift, organizational ‘strategy’ referred to a firm’s perspecstrategic HRM aims to link firm performance to the specific methods and practices deployed within the firm to manage their workers. When first introduced, this approach represented a significant shift in firm management. Prior to this shift, organizational ‘strategy’ referred to a firm’s perspective or ‘world view’, its intended plan and patterned behaviour, as well as a firm’s use of ploys to outwit competitors (Mintzberg 1987). However, in the 1980s firm strategy began to focus more on causality, with researchers and practitioners attempting to identify inputs (e.g. management practices) and outputs (e.g. market performance), quantify them, and derive causal connections between the two. Operationalizing this strategic approach, as many have argued, has largely been “geared towards specific numerical targets” (Wood and Kispál-Vitai 2017). Firms seek not only to derive general conclusions or trends through quantification, but also to compare a range of strategies for the purpose of distinguishing a singular ‘best practice’. This epistemological shift has fueled the development of new metrics and datapoints that can be used to analyze the relationship between workforce management and firm performance. In turn, HR managers can use these new sources of data about the workforce to inform decision-making. This type of Evidence-Based Management technique (EBM) has been largely promoted in strategic HRM to help overcome pitfalls, such as relying on personal experience or managerial whims, or the tendency to mimic the strategies or approaches of top performers (Pfeffer and Sutton 2006; Rousseau 2006); Reay, Berta, and Kohn 2009). The result has been a vast arsenal of digitally enabled workplace and work-related tools that capture, collect and analyze worker behavior and performance data. Since the 1980s, this has driven the growing field of people analytics – defined as “the use of measurement and analysis techniques to understand and optimize the people side of business”(Enderes and Shannon 2019) – and ultimately has paved the path towards the adoption of AI for HRM.

alytics – defined as “the use of measurement and analysis techniques to understand and optimize the people side of business”(Enderes and Shannon 2019) – and ultimately has paved the path towards the adoption of AI for HRM. Within the field of people analytics and the development of AI, more data and better data are commonly viewed as precursors for robust and powerful systems. People analytics data may pertain to workers’ demographics, descriptive information about the location or nature of the job, performance, training or professional history, or tenure. While every organization has data on its workforce, advanced analysis for workforce optimization and planning can only be achieved when data are high quality, robust and plentiful, and when organizations have adequately trained staff to process and make sense of them. Figure 1 is an infographic from a McKinsey publication on the “virtue” of people analytics, stressing that the power of such systems to provide workforce insights increases with the volume and quality of workforce data (Ledet et al. 2020).06 ILO Working Paper 154 X Figure 1. McKinsey infographic on the benefits of people analytics Source: (Ledet et al. 2020). When first introduced, HR professionals used people analytics to assess patterns about their workforce, informing a wide range of HR functions including recruitment and hiring, promotion, compensation, and health and safety (Giermindl et al. 2022). With more and higher-quality “big data” and increases in computing power, people analytics is being propelled from correlation analysis into the world of pattern-based prediction, and Human Resource professionals, once responsible for executing a wide range of functions, are, in some instances, relinquishing these responsibilities to algorithms and AI. This decades-long strategic shift provides critical context for understanding why the field of HRM has embraced the use of AI; the particular harms that may emerge from the use of AI; and why the field of HR is largely blind to them. AI is distinguished by vast quantities of data and rapid quantitative analysis. These new sources of data have held particular appeal to a profession that has oriented itself towards EBM. Additionally,

may emerge from the use of AI; and why the field of HR is largely blind to them. AI is distinguished by vast quantities of data and rapid quantitative analysis. These new sources of data have held particular appeal to a profession that has oriented itself towards EBM. Additionally, AI is popularly portrayed as innovative and cutting edge, and the use of AI technologies and tools is thus seen as a way to elevate the HR profession. This motivation is also underpinned by important context as historically many have tended to regard HR professionals as administrative functionaries, who – lacking power – are engaged in mere bureaucratic service delivery without adding any real value to the organization (Wright 2008; Legge 1978). Although the strategic focus of HRM in the 1980s (along with other activities such as the formation of professional associations and educational and training courses) provided one avenue for the occupation to recast itself on equal footing with other managerial professions (Legge 1978; Cayrat and Boxall 2023), critics maintain that the field has still not provided evidence of its worth (Wright 2008; Alvesson 2008; Kryscynski et al. 2018; Cayrat and Boxall 2023; Hammonds 2005). Thus, this persistent quest for occupational legitimacy presents another important motivation for HR’s embrace of people analytics and AI: by embracing evidence-based management (EBM),07 ILO Working Paper 154 and the data accumulating and data-intensive tools and technologies that facilitate it, HR practitioners might provide evidence of their value to the firm. This approach, however, has not been without criticism. Scholars argue that while empirical research linking HR practices to organizational performance can at times demonstrate a clear association, the relationship between these variables is under-theorized (Fleetwood and Hesketh 2008). The empirical approach of HR scholars has proven adequate for generating predictive dimensions of theory rooted in prior observations of ‘what’ is happening and ‘how’; however, robust theories should also be capable of explaining the mechanisms and reasoning behind

2008). The empirical approach of HR scholars has proven adequate for generating predictive dimensions of theory rooted in prior observations of ‘what’ is happening and ‘how’; however, robust theories should also be capable of explaining the mechanisms and reasoning behind the mechanisms and causal relationships (Guest 2025; Fleetwood and Hesketh 2008). It is with respect to this latter explanatory dimension that HRM theory has fallen short. Without an adequate or clearly articulated theory, research will “also lack an adequate rationale for the choice of phenomena that will eventually become the variables” used to drive future predictive theorization (ibid); 127). Under these conditions, rather than explaining why organizations succeed or fail and what other systemic conditions or discrete practices may contribute to this outcome (Guest 2025), HR managers and researchers alike are prone to replicate past practice simply because such practices have been previously examined. HR theories, in turn, remain poorly articulated and lack conceptual clarity (Boon et al. 2019). This risk of ‘measurement without theory’ is amplified in the context of AI, which by definition is not about building an explicit theory, but rather about building patterns through big data (Elragal and Klischewski 2017). Proponents of AI integration, including many global consulting firms, promise exactly this: added value from the use of their newly developed technologies and tools – and specifically, that the predictive capacity of AI will help to reveal the recipe for an effective allocation of human resources. Companies are rapidly obliging. ISG, an organizational change management company, found in 2023 that one out of every three organizations was prioritizing AI and analytics in their HR and technology strategy. This finding was based on a survey of enterprise leaders at firms employing between 5,000 and 50,000 workers. As the report explained, “leading HR technology providers, such as such as Oracle, [IBM], SAP SuccessFactors and Workday, are focused on embedding AI, machine learning and analytics in the core platform” (ISG 2023). Indeed, most of the software is provided through third-parties and is typically an add-on to other services algy providers, such as such as Oracle, [IBM], SAP SuccessFactors and Workday, are focused on embedding AI, machine learning and analytics in the core platform” (ISG 2023). Indeed, most of the software is provided through third-parties and is typically an add-on to other services already provided. This facilitates the work of HR professionals who often lack adequate training in what kinds of analytical questions to ask or how to interpret quantitative findings (Giermindl et al. 2022; Kryscynski et al. 2018). Yet despite high rates of technological integration, including AI, the same ISG survey found that less than half of organizations surveyed realized business value from their investments. What has emerged is thus a paradox: although troves of data and developments in the field of AI promise to provide HR professionals with the necessary data and analysis to help them effectively and efficiently allocate human resources, these technologies have not yet delivered much value to firms and organizations. Why not? What does the actual evidence on the implementation of AI in HRM bear? Does the design of these AI systems allow them to deliver as promised?08 ILO Working Paper 154 X 2 The workings of AI systems: Objective, data and programming

To assess how effective AI has been for realizing the overarching goals of HR, it is necessary to delineate what these goals are, and to unpack the workings of AI system to better understand potential pitfalls. AI systems are composed of three inter-related parameters: (1) the system objective, (2) the data it is built on and relies on, and (3) how it is programmed (See figure 2). The quality of each of these parameters differentiates systems that work well from those that do not. Beginning with the system objective, while defining such an aim may appear straightforward, how these aims are operationalized is more complicated. When AI systems are developed with neutral aims – such as determining the shortest route between locations – it is easy to rely on identifiable and relevant variables and the findings are easy to interpret. But most human reBeginning with the system objective, while defining such an aim may appear straightforward, how these aims are operationalized is more complicated. When AI systems are developed with neutral aims – such as determining the shortest route between locations – it is easy to rely on identifiable and relevant variables and the findings are easy to interpret. But most human resource functions involve the “fleshy, messy, indeterminate stuff of everyday life” – stuff that are difficult to capture in a discrete variable (Katz 2001). This can, as described by Sandy Gould (Gould 2024), create practical challenges, which she explains as the need to, “accept resource constraints on what is measured, as well as the ontological limitations (some things are never going to be open to ‘direct’ measurement by any conceivable means)” (106). Data, meanwhile, are a necessary input to AI systems. Limitations in data have been the focus of much critique, with two particularly salient concerns. First, there is the question of data quality. AI systems rely on training data to ‘learn’ the connections and patterns that provide the foundation upon which decisions are made (Whang et al. 2023). When these data are poor quality, AI systems yield poor quality outputs. As the saying goes: ‘garbage in, garbage out’. A second issue concerns the suitability of the data. In bespoke AI systems that are developed internally by or for a singular firm, training data may be internal to the firm and consist of past operations (Kresge 2020). When systems are developed for ‘off the shelf’ use, they rely on more generalized datasets that can either be purchased via a growing data market (Zuboff 2019) or otherwise compiled from available data that are deemed to be relevant sources for the system (Muldoon et al. 2023). The appropriateness of more generalized datasets ought to be evaluated on a case-by-case basis but rarely is. Equally, when issues stemming from non-representative data arise, a commonly suggested solution is that more data or more representative data are needed. However, these types

appropriateness of more generalized datasets ought to be evaluated on a case-by-case basis but rarely is. Equally, when issues stemming from non-representative data arise, a commonly suggested solution is that more data or more representative data are needed. However, these types of problems also raise questions about the suitability of the data to meet the stated objective. Those cautious of AI have also pointed to the potential problems associated with the third parameter: programming. Algorithms are a key decision-making feature at the core of AI systems. They can be defined, in their most basic sense, as a set of rules executed through computer programming code with a particular aim or objective. Algorithms function with varied levels of autonomy and human involvement; these characteristics are also determined by their design and deployment (OECD 2021). The most basic algorithms merely execute a list of prescribed instructions or offer data insights or recommendations that can assist human decision making; basic algorithms only change with intervention from those who use or program them. In these instances, bias can be intentionally or unintentionally embedded in the computer code by the programmer. Issues also arise with more autonomous ‘intelligent’ ‘machine learning’ algorithms like those associated with and used in AI systems. Programming can become fraught as the system ‘evolves’. Self-learning systems, although still developed and deployed by humans, often operate in opaque ways and their precise functionalities may elude even those who built them.09 ILO Working Paper 154 X Figure 2. The workings of AI systems

Source: Authors’ elaboration.

For those who warn of the detrimental impact that AI could have in the world of work, how these three parameters of AI systems (objective, data, and programming) are chosen and deployed could jeopardize work quality and workplace fairness. Moreover, the use of third parties’ ‘off-the-shelf’ systems can heighten risks. First, the algorithms embedded in AI systems may be poorly understood by firms and workers using them because of intellectual property concerns; and firms using externally developed systems may have limited ability to modify how the system functions to meet their specific needs. Second, deploying

embedded in AI systems may be poorly understood by firms and workers using them because of intellectual property concerns; and firms using externally developed systems may have limited ability to modify how the system functions to meet their specific needs. Second, deploying these types of technologies in contexts where real-life data differ from training data can lead to discrimination or poor outcomes. Third, concerns about workers’ data privacy have also emerged. While firms have long tracked a host of data about worker performance, the use of third-party systems often involves complex licensing arrangements that may include provisions that require workers’ data to be shared with the AI developer. Such sharing raises important questions about consent, privacy, and even the commodification of workers’ data (Kresge 2020). Within purpose-built systems, like those used in HR, the risks differ slightly according to the intended function. What follows is an analysis of contemporary uses of AI to help fulfil key HR functions, including recruitment, setting compensation, scheduling work, and performance management. Each of these discrete and purpose-built systems for HR functions is assessed in terms of the overall objective that is specified, as well as the data and programming involved. Recruitment Of all human resource functions, recruitment has been most transformed by digitalization, and more recently by AI. The shift to online recruitment began in the mid-1990s and was hailed by industry experts and economists as a revolution that would improve labour market matching by lowering the friction costs associated with job search and worker sourcing (Krueger 2000). Yet the ease of advertising and applying for jobs has increased the volume of applications, making the10 ILO Working Paper 154 task of sorting candidates more cumbersome, thus necessitating technological solutions to assist the process. Recruitment is thus a prime example of the “paradox of automation” (Gray and Suri, 2019), whereby each problem that technology tries to solve, creates a new problem to be solved. The use of AI (or sometimes just algorithmic systems) in recruitment spans the different stages of the recruitment process, including sourcing of applications, but also their screening and even2019), whereby each problem that technology tries to solve, creates a new problem to be solved. The use of AI (or sometimes just algorithmic systems) in recruitment spans the different stag