Introduction and Background
The ongoing war for talents shapes the future of organisations and is a major challenge for the entire field of talent management (TM). TM thereby involves (1) recruitment, staffing, and succession planning, (2) training and development, and (3) retention management. Nonetheless, organisations struggle substantially in managing the scarcity of talents effectively. The growing capabilities of artificial intelligence (AI) show potential in creating a strategic advantage in this war for talents (Brock & von Wangenheim, 2019). Given the rapidly increasing number of studies examining how AI impacts TM, the research lacks a comprehensive review. To address this void and thus support scholars and practitioners by providing an overview of extant knowledge on this topic, we conducted a systematic literature review (SLR) on the use of AI in TM.
Method
We based our SLR on the PRISMA approach (Page et al., 2021). To identify relevant literature, we searched in different electronic databases. This provided us with 3,714 publications from which 47 were eventually analysed in detail. We took a stepwise approach to select a conclusive final sample (see Figure 1). Based on previously defined criteria, we excluded non-fitting articles. Each step reduced our initial sample and eventually yielded a preliminary sample of 37 relevant articles. Based on these articles, we conducted backward and forward search, which resulted in additional 10 relevant articles. Consequently, in sum we received a final sample of 47 articles published between 2015 and 2021 on the influence of AI on TM.
Findings and Conclusion
We identified three overarching streams in the literature regarding AI in TM. Literature addressed AI use for recruitment, training and development, and retention management whereas the first stream was the most prominent. Regrading recruitment, literature emphasized prerequisites for and barriers to AI adoption. Cost-effectiveness of AI use, employee readiness, top-management support, and technology vendor support (Pillai & Sivathanu, 2020) are found to be crucial prerequisites for AI adoption. In contrast, security risks, privacy concerns, and high in-house technological complexity hinder AI adoption in recruitment (Pan et al., 2021; Pillai & Sivanthanu, 2020). Further, literature showed how AI can benefit recruitment efficiency across various TM activities such as outreaching to candidates, job-candidate matching, screening of candidates, candidate assessment and evaluation (Black & van Esch, 2020). For instance, machine learning (ML) algorithms trained based on video interviews, have been successfully employed to predict applicants’ personality traits (Hickman et al., 2021). However, literature also showed the importance of both, AI expertise and domain expertise for successful AI application (van den Broek et al., 2021). Despite possible use cases in training and development, like using AI to reveal potential skill gaps (Malik et al., 2020) or the introduction of AI based animated characters to provide feedback on learning progress (Vrontis et al., 2021), research in this area is scarce. In relation to retention management and the prediction of employee turnover through AI applications (e.g., ML for patterns detection and neural networks) is a well-researched phenomenon (e.g., Choudhury et al., 2021; Teng et al., 2021; Yuan et al., 2021). Additionally, AI has been employed to understand employees’ mood swings, enabling targeted countermeasures to prevent turnover (Kshetri, 2021).
The second stream comprises literature on individuals’ perception of AI application in TM. Research addressed user acceptance and willingness to use AI applications. Literature on user acceptance showed that AI in the application process is accepted when applicants are informed about its use early in the process. However, AI evaluation and selection decisions are more likely to be accepted by applicants, recruiters, and managers when matched with human decision-making (Laurim et al., 2021). Literature on user’s willingness to use AI showed contradictory results, as known AI use reduced willingness to apply for a position (Mirowska, 2020). Incorporating trendiness, fair treatment, and intrinsic rewards offered for using the AI, mitigated the resentments and increased users’ willingness to use AI (van Esch & Black, 2019).
The third stream targets algorithmic decision making in TM, addressing particularly AI bias and fairness and AI decision’s effects on users. Extant literature showed how biased decisions for example regarding gender manifested in ML training data (Köchling & Wehner, 2020). Particularly, algorithmic decisions were perceived less fair than human decisions and were perceived reductionist (Newman et al., 2020). Regarding algorithmic decision-making, literature identified a potential overconfidence in the algorithms’ objectivity, potentially even devaluing human decision- making, which was perceived subjective and deficient to the algorithms (Giermindl et al., 2021).
With our SLR, we offer a comprehensive overview of the extant studies on AI in TM. Thus, we enable scholars to identify possible future research options and practitioners to build decisions on the adoption and implementation of AI for TM on a scientific basis. Thereby, we contribute to both, research and practice.
Employee retention is becoming increasingly important, given the scarcity of high-skilled employees. Thus, practitioners and scholars are trying to understand and predict employee turnover as accurate as possible (Ben-Gal et al., 2021; Choudhury et al., 2021; Farrell and Rusbult, 1981; Oswald, 2020; Wang and Zhi, 2021; Yuan et al., 2021; Zhao et al., 2018). Complex machine learning (ML) models can offer powerful support in this context, as they enable high accuracy predictions (Hong et al., 2007). However, these models are often based on black-box models and provide limited interpretability of the results (Shrestha et al., 2021). Hence, to use black-box models, more interpretable approaches (Ben-Gal et al., 2021; Mitchell et al., 2001), i.e., explainable AI (XAI) methods (Guidotti et al., 2018; Hamm et al., 2021; Barredo Arrieta et al., 2020) are necessary. Understanding the models for turnover prediction – respectively the underlying patterns- is particularly important for Human Resource Management (HRM) to take appropriate countermeasures (Oswald, 2020). Therefore, we evaluate the use of three XAI methods in the context of turnover. As basis of our exemplary illustration, we use the publicly available IBM HR Analytics Employee Attrition & Performance data set (Subhash, 2017), which is a synthetic data set created by IBM that includes administrative data, performance data, data on job satisfaction, and data on individual characteristics (e.g., age and gender) of 1470 fictional employees (see Appendix for an exhaustive list of features) (Subhash, 2017). The goal of our exemplary use of XAI was to predict turnover (0 = no turnover, 1 = turnover) based on these characteristics and make the prediction interpretable.
For this purpose, we selected a sample of 7 frequently used ML models (e.g., Hastie et al., 2009; Wu et al., 2007) and evaluated their accuracy using 10-fold cross-validation (Berrar, 2018). The results regarding the achieved mean accuracy are shown in Figure 1. The Random Forest (RF) model achieves the highest accuracy. After the final model optimization, we achieve a f1-score1 of 84%. Due to the comparatively high accuracy, this model can provide a solid basis to predict employee turnover.
However, two fundamental questions remain. First, what are the general reasons for turnover in the organization? Second, how to prevent a turnover of an identified employee? Technically, these questions address two different types of explanations: Global interpretability (the average model behavior) and local interpretability (so the model’s prediction for an individual observation – here a particular employee) (Molnar, 2020). To answer these questions, we employ Local Interpretable Model-Agnostic Explanations (LIME) (Ribeiro et al., 2016), Shapley Additive Explanation (SHAP) (Lundberg and Lee, 2017), and Partial Dependence Plots (PDP) (Kamath and Liu, 2021).
For global interpretability, we use a SHAP summary plot (Molnar, 2020), as shown in Figure 2. This ranks the features according to their influence on the model output. Since the influence of the feature value is different for each instance, depending on the value of the other features, SHAP displays this with a scatter plot. For example, the feature StockOptionLevel has the highest average impact on the model, where especially a low level of stock options links to a turnover prediction of the model. In comparison, the feature TotalWorkingYears has a lower impact on the model. By analyzing these features and their model impact, the key drivers of turnover in the organization can be detected and used by the HRM for data-driven decisions (Oswald, 2020). In this case, the following question could be derived: What level of stock options is appropriate to prevent potential turnover?
Therefore, we take a more granular insight with PDP (Kamath and Liu, 2021; Molnar, 2020) to investigate the dependence between the feature StockOptionLevel and the model output (Figure 3). The plot shows that the curve flattens significantly from the StockOptionLevel of 1 onwards and thus a further increase in the level of stock options rarely changes the average response of the model’s decision.
Besides PDP, we used LIME and SHAP as they offer additional insights. Both XAI methods allow the analysis of individual observations (i.e., employees in our case) and thus enable local interpretability (Molnar, 2020). They display, in descending order, the strongest influence of the features on an individual model prediction (Figure 4). Both methods come to fairly close results and thus could be used to improve our understanding of potential reasons of a single turnover decision.
Finally, we demonstrated the use of XAI applications for decision making in HRM. XAI enables to make the decisions of complex ML models understandable. This inside can be used in two ways: First, to get a global understanding of turnover in the organization, and second, to understand possible reasons for the supposed turnover of a particular employee. This leads to the limitations with the presented XAI methods. In addition to the methodological challenges of the correct application (Choudhury et al., 2021), these methods do not provide any information on whether the taken actions actually change the decision of the model and thus, of the employees (Slack et al., 2020; Fernández-Loría et al., 2022). Hence, it is not possible to read off or determine what quantity of change, for example in the level of stock options, changes in the end the decision. However, the presented XAI approaches are a first step to get insights into complex AI models and to a data-driven assessment of which actions are most likely to be successful.
