Reconceptualizing the Automation–Augmentation Tension in AI-Enabled Talent Selection: A Stage-Embedded Theory

Reconceptualizing the Automation–Augmentation Tension in AI-Enabled Talent Selection: A Stage-Embedded Theory | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Article Reconceptualizing the Automation–Augmentation Tension in AI-Enabled Talent Selection: A Stage-Embedded Theory Hochan Lee This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-9222046/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted 4 You are reading this latest preprint version Abstract Artificial intelligence (AI) is increasingly embedded in talent selection, yet its deployment remains characterized by a persistent tension between automation and augmentation. While existing research often treats this choice as an organizational-level strategy or a temporal progression, it frequently overlooks how institutional sensitivities and task characteristics vary across the recruitment lifecycle. This research examines how AI is configured as automation, augmentation, or hybrid collaboration across specific recruitment stages and why organizations repeatedly transition between these configurations. Adopting a comparative, multi-case theory-building approach, the paper analyzes documented AI deployments across standardized selection stages, including applicant screening, assessment, and final selection. The findings identify three recurring stage-embedded mechanisms. Early recruitment stages stabilize efficiency-driven automation under conditions of high task analyzability and low immediate accountability. Intermediate stages intensify paradoxical tensions, producing unstable hybrid configurations as organizations attempt to balance scalability, validity, and fairness. Late stages institutionalize accountability-driven augmentation, structurally constraining AI autonomy due to normative and legal requirements for human responsibility, regardless of technical performance. Across these stages, organizations engage in "re-augmentation" as a legitimacy repair mechanism to restore trust and institutional alignment following perceived algorithmic failures or ethical concerns. By reconceptualizing the automation–augmentation paradox as a stage-embedded phenomenon, this work demonstrates that contradictory logics are distributed spatially across the hiring process. These insights refine task complementarity theory by establishing accountability as a fundamental boundary condition that limits human–AI collaboration in high-stakes human resource decisions. Humanities/Complex networks Social science/Complex networks Physical sciences/Mathematics and computing Social science/Science technology and society artificial intelligence talent selection automation augmentation case-based theory building Figures Figure 1 1. Introduction Recruitment is one of the most consequential and institutionally sensitive functions in human resource management (HRM). Hiring decisions shape organizational performance, workforce diversity, and long-term capability development, while simultaneously exposing organizations to legal, reputational, and ethical scrutiny (Breaugh, 2013 ; Gatewood, Feild and Barrick, 2020 ). Because recruitment determines who enters the organization—and on what grounds—it remains a focal site for debates about fairness, accountability, and legitimacy (Rynes, Bretz and Gerhart, 1991 ). Against this backdrop, artificial intelligence (AI) has become deeply embedded in recruitment and selection processes. Organizations increasingly rely on algorithmic systems to screen applicants, administer assessments, analyze interviews, and generate predictive insights regarding future performance or retention (Tambe, Cappelli and Yakubovich, 2019 ). These technologies promise substantial efficiency and scalability gains, particularly in contexts characterized by high applicant volumes and time pressure (Vrontis et al., 2023 ). At the same time, AI-enabled recruitment remains highly contested, with recurring concerns about bias, transparency, explainability, and accountability (Hunkenschroer & Luetge, 2022 ). Recent integrative research further demonstrates that AI bias in recruitment cannot be reduced to technical malfunction but reflects the interaction of data histories, organizational routines, and governance arrangements, thereby requiring sustained human oversight and institutional alignment (Soleimani et al., 2025 ). Much of the debate surrounding AI in recruitment has been framed through a distinction between automation and augmentation. Automation emphasizes delegating decision tasks to algorithms in pursuit of efficiency, consistency, and standardization, whereas augmentation frames AI as decision support that enhances human judgment while preserving human discretion and responsibility (Raisch and Krakowski, 2021 ). This distinction mirrors a broader debate in management research regarding whether AI substitutes for or complements human labor (Faraj, Pachidi and Sayegh, 2018 ). Recent scholarship challenges this binary framing by conceptualizing automation and augmentation as a paradoxical relationship rather than mutually exclusive alternatives. From a paradox perspective, automation and augmentation are contradictory yet interdependent: automation can generate efficiency gains that enable higher-order human judgment, while augmentation can foster learning and trust that stabilize future automation (Smith and Lewis, 2011 ; Raisch and Krakowski, 2021 ). Task complementarity theory further suggests that performance benefits depend on how humans and AI are configured at the task level rather than on automation or augmentation per se (Fügener, Walzner and Gupta, 2025). Despite these advances, existing theorizing remains insufficiently grounded in the institutional realities of recruitment. Hiring differs from many organizational domains in that decision outcomes are highly consequential, socially sensitive, and externally scrutinized (Dineen, Yu and Stevenson, 2023). Responsibility for selection decisions is normatively and legally assigned to human actors, even when algorithmic systems play a substantial role (Meijerink et al., 2021 ). As a result, AI use in recruitment cannot be understood solely in terms of technical performance or cognitive complementarity; it is also shaped by legitimacy expectations and accountability structures (Newman, Fast and Harmon, 2020 ). More broadly, research on algorithms at work demonstrates that algorithmic systems do not merely replace tasks but reorganize authority structures, accountability relations, and worker autonomy within organizations (Kellogg, Valentine and Christin, 2020 ). Recruitment therefore represents not only a technical deployment of AI, but a reconfiguration of institutional authority in high-stakes decision-making. A further limitation of prior research is its tendency to treat recruitment as a monolithic process. Personnel selection research has long emphasized that hiring unfolds through a sequence of stages—ranging from applicant attraction and screening to assessment, final decision-making, and validation—each designed to manage trade-offs among predictive validity, efficiency, and risk (Schmidt and Hunter, 1998 ; Sackett et al., 2022 ). Early stages prioritize efficiency and scalability, while later stages involve judgment-intensive evaluations and heightened accountability. Intermediate stages combine both, making them especially prone to tension and instability. This study argues that the automation–augmentation paradox in recruitment is not merely temporal or organizational but structurally embedded across recruitment stages. Organizations rarely adopt a single, stable AI configuration. Instead, they deploy automation and augmentation simultaneously across stages and repeatedly reconfigure AI roles over time. Notably, retreats from automation often reflect legitimacy and accountability concerns rather than technical failure, echoing evidence on algorithm aversion and declining trust following perceived errors (Dietvorst, Simmons and Massey, 2015 ; Logg, Minson and Moore, 2019 ). More broadly, empirical research demonstrates that trust in AI is highly sensitive to perceptions of reliability, transparency, and contextual accountability, particularly in consequential decision settings (Glikson and Woolley, 2020 ). To capture these dynamics, this study adopts a comparative, case-based theory-building approach to examine how AI is configured across recruitment stages and why organizations transition between automation, augmentation, and hybrid configurations. The study addresses the following research question: How are automation and augmentation dynamically configured across recruitment stages, and why do organizations transition between these configurations over time? 2. Theoretical Background 2.1 Recruitment as a Stage-Based and Institutionally Sensitive Decision System Personnel selection research has consistently emphasized that recruitment unfolds as a sequence of interdependent stages rather than as a single evaluative decision (Breaugh, 2013 ; Schmidt and Hunter, 1998 ). Each stage is designed to manage trade-offs among predictive validity, efficiency, and risk, while decisions made at earlier stages constrain the candidate pool available at later ones (Sackett et al., 2022 ). Recruitment is therefore inherently path-dependent. Beyond technical considerations, recruitment stages differ systematically in their institutional exposure. Early stages involve high volumes, standardized information, and relatively low immediate decision stakes, allowing for greater tolerance of error and experimentation (Gatewood, Feild and Barrick, 2020 ). Later stages, by contrast, are characterized by judgment-intensive evaluations, heightened legal and reputational risk, and strong normative expectations that humans remain accountable for outcomes (Rynes, Bretz and Gerhart, 1991 ). Intermediate stages combine elements of both, positioning them as sites of heightened tension between efficiency and legitimacy. This stage-based differentiation implies that recruitment is best understood as an institutionally embedded decision system. Task characteristics and accountability expectations vary across stages, shaping not only how decisions are made but also what forms of delegation to AI are considered legitimate and defensible ex post. Table 1 summarizes the standardized selection process and delineates stage-specific differences in core purpose, evaluation logic, and institutional exposure, thereby providing the analytical foundation for the stage-embedded perspective developed in this study. Table 1 Standardized Selection Process and Stage-Specific Task and Institutional Characteristics Stage Selection Stage Core Purpose Key Tools & Techniques Evaluation Characteristics Institutional & Task Characteristics Reference 1 Applicant Pool Formation To attract and assemble a sufficiently large and qualified pool of applicants, thereby setting the upper bound of potential selection quality Job advertisements, employer branding, employee referrals, online recruitment platforms, labor market intermediaries Indirect predictor of selection quality; determines base rates and range restriction; susceptible to self-selection bias High volume; highly analyzable; high error tolerance; low accountability exposure; low institutional sensitivity Breaugh ( 2013 ); Rynes et al. (1990) 2 Minimum Qualification Screening To eliminate applicants who do not meet essential job-related requirements prior to resource-intensive assessment Degree and certification checks, experience thresholds, legal eligibility verification, automated applicant tracking systems (ATS) Low predictive validity; high efficiency; high risk of false negatives; often rule-based Very high analyzability; high automation potential; low immediate accountability; legitimacy risks largely latent Gatewood et al. ( 2020 ) 3 Resume / CV Screening To evaluate applicants’ job-relevant background and experiences as an initial indicator of person–job fit Manual resume review, competency checklists, keyword matching, AI-based resume parsing and ranking Moderate reliability; limited criterion-related validity; vulnerable to human and algorithmic bia`s Pattern-recognition task; moderate ambiguity; increasing legitimacy scrutiny; moderate accountability Cole et al. ( 2004 ) 4 Preliminary Assessment To assess applicants’ general mental ability, personality traits, and basic competencies linked to future job performance Cognitive ability tests, personality inventories (e.g., Big Five), integrity tests, situational judgment tests (SJTs) High predictive validity (especially cognitive ability); standardized; scalable; potential adverse impact Moderate analyzability; growing ethical and fairness concerns; moderate–high institutional sensitivity Schmidt et al. (1998); Sackett et al. ( 2022 ) 5 Structured Interview To systematically evaluate applicants’ job-relevant knowledge, skills, abilities, and motivation using standardized questions Behavioral interviews, situational interviews, structured scoring rubrics, interviewer training High reliability and validity when structured; reduced bias vs. unstructured interviews; moderate cost Low analyzability; judgment-intensive; high accountability; strong legitimacy expectations Campion et al. ( 1997 ) 6 Work Sample / Job Simulation To directly observe applicants’ ability to perform tasks that closely resemble actual job duties Work sample tests, job simulations, coding tasks, case analyses, in-basket exercises High content and criterion-related validity; strong face validity; limited scalability; high development cost Medium analyzability; low error tolerance; high legitimacy and defensibility requirements Roth et al. ( 2005 ) 7 Assessment Center To evaluate multiple competencies through simulations and exercises assessed by multiple raters Group discussions, role plays, leaderless group tasks, presentations, multi-rater assessments Multidimensional assessment; high developmental value; resource-intensive; validity design-dependent Complex social evaluation; very high institutional sensitivity; high accountability diffusion Arthur et al. ( 2003 ) 8 Reference and Background Checks To verify applicants’ prior performance, qualifications, and potential risk factors Reference interviews, employment verification, criminal background checks, credential validation Low incremental validity; confirmatory rather than predictive; legal and ethical constraints Low analyzability; high legal exposure; primarily legitimacy- and compliance-driven Gatewood et al. ( 2020 ) 9 Final Selection Decision & Job Offer To integrate information from all prior stages and select the candidate with the highest expected utility Compensatory decision models, weighted scoring algorithms, human judgment panels, decision-support systems Decision quality depends on weighting strategy; human–algorithm interaction critical; risk of judgmental bias Very low analyzability; minimal error tolerance; very high accountability; human authority expected Schneider & Schmitt (1976) 10 Validation and Evaluation To assess and improve the effectiveness, fairness, and legal defensibility of the selection system Criterion-related validation studies, utility analysis, adverse impact analysis, ongoing performance monitoring Essential for long-term effectiveness; ensures compliance and continuous improvement; often underutilized Meta-level governance task; institutional learning mechanism; accountability reinforcement Sackett et al. ( 2022 ) 2.2 Automation and Augmentation in AI Research Research on AI in organizations has increasingly framed automation and augmentation as a paradoxical relationship rather than as mutually exclusive strategies (Raisch and Krakowski, 2021 ). From a paradox perspective, automation and augmentation are contradictory yet interdependent: automation emphasizes efficiency, standardization, and scalability, whereas augmentation preserves human judgment, discretion, and responsibility. These logics cannot be fully reconciled but must be continuously managed (Smith and Lewis, 2011 ). Recent sociotechnical scholarship further argues that algorithmic systems operate within layered social infrastructures, where technical optimization and institutional legitimacy frequently diverge (Selbst et al., 2019). This divergence is particularly salient in recruitment, where predictive accuracy alone cannot secure normative acceptance. Most prior work conceptualizes this paradox at the organizational or temporal level, examining how firms shift between automation-oriented and augmentation-oriented strategies over time. However, such approaches implicitly assume that automation and augmentation are resolved or balanced at the firm level. In recruitment, this assumption is problematic. Because recruitment consists of multiple stages with sharply differing task structures and institutional expectations, automation and augmentation are rarely adopted uniformly. Instead, contradictory logics are distributed across stages within the same hiring system. Automation may be stabilized in some stages while augmentation is institutionalized in others, allowing organizations to manage paradox spatially rather than temporally. This distributed configuration suggests that the automation–augmentation paradox in recruitment is not a transient managerial dilemma but a structural feature of the selection process itself. 2.3 Task Complementarity and Its Limits in Recruitment Task complementarity theory provides a task-level explanation for how humans and AI can jointly generate performance benefits (Fügener, Walzner and Gupta, 2025). However, prior formulations of complementarity largely emphasize cognitive allocation efficiency between humans and algorithms (Jarrahi, 2018 ). Such accounts understate how institutional responsibility structures shape the feasible boundaries of delegation in consequential organizational decisions. The theory distinguishes between between-task complementarity—where humans and AI perform different tasks according to their comparative advantages—and within-task complementarity—where humans and AI jointly perform the same task. Automation typically creates value through between-task complementarity in highly analyzable tasks with high error tolerance. Augmentation, by contrast, relies on within-task complementarity, requiring humans to interpret, contextualize, and appropriately weight algorithmic outputs. However, existing formulations of task complementarity largely treat complementarity as a cognitive or technical condition. In recruitment, complementarity is also institutionally constrained. Responsibility for selection decisions is normatively and legally assigned to human actors, even when algorithmic systems demonstrate strong predictive performance (Meijerink et al., 2021 ). This allocation of responsibility limits the extent to which AI can assume autonomous decision authority, particularly in later recruitment stages where accountability exposure is greatest. Consequently, within-task complementarity in recruitment is bounded not only by human interpretive capacity but also by accountability requirements that restrict AI autonomy regardless of technical capability. 2.4 Re-Augmentation and Legitimacy Repair Prior research often interprets retreats from automation as evidence of technological failure or managerial resistance. However, emerging evidence from AI-enabled decision contexts suggests that reversals may instead reflect legitimacy dynamics. Algorithmic systems are often penalized disproportionately after perceived errors, leading to declining trust and acceptance (Dietvorst, Simmons and Massey, 2015 ), even though algorithmic judgment may be preferred in principle under some conditions (Logg, Minson and Moore, 2019 ). Experimental research further demonstrates that perceptions of procedural justice and explainability strongly mediate acceptance of algorithmic decisions, especially when outcomes affect identity-relevant domains such as employment (Lee, 2018 ). These findings suggest that legitimacy shocks in recruitment are socially constructed responses rather than purely performance-driven reactions. In recruitment, where decisions affect identity, dignity, and future opportunity, legitimacy threats are particularly salient. Algorithmic decision-making can undermine perceptions of human consideration and moral accountability, making organizations appear less humane (Newman, Fast and Harmon, 2020 ). In response, organizations may deliberately reintroduce human oversight—not to abandon AI, but to restore trust, defensibility, and institutional alignment. This study conceptualizes such responses as re-augmentation: the deliberate reassertion of human involvement following periods of increased automation. Re-augmentation functions as a legitimacy repair mechanism that stabilizes AI-enabled recruitment systems without fully relinquishing efficiency gains. Unlike human-in-the-loop designs, which assume stable human oversight from the outset, re-augmentation refers to a deliberate organizational reconfiguration following periods of expanded automation. It is triggered not by performance failure, but by legitimacy threats that render existing AI configurations socially or institutionally untenable. 2.5 Theoretical Integration and Expectations Synthesizing these perspectives, this study conceptualizes recruitment as a stage-embedded automation–augmentation system. Automation and augmentation are not alternative strategies to be chosen or sequenced at the organizational level. Rather, they are configured differently across recruitment stages in response to task analyzability, accountability exposure, and legitimacy pressures. This perspective yields three expectations that guide the subsequent analysis. First, automation will stabilize in early recruitment stages characterized by high analyzability and low accountability. Second, intermediate stages will exhibit unstable or hybrid configurations as organizations struggle to balance scalability, validity, and legitimacy. Third, late recruitment stages will institutionalize augmentation, constraining AI autonomy regardless of technical performance. These expectations provide the theoretical foundation for the comparative, case-based analysis that follows. 3. Methodology 3.1 Research Design: Case-Based Theory Building This study adopts a comparative, multi-case study design to theorize how artificial intelligence (AI) is configured as automation, augmentation, or hybrid human–AI collaboration across recruitment stages. Case-based methodology is particularly well suited to this research because the phenomenon of interest—the dynamic reconfiguration of AI roles—is complex, processual, and insufficiently theorized in the HRM literature. Consistent with established traditions of case-based theory building (Eisenhardt, 1989; Eisenhardt & Graebner, 2007 ; Siggelkow, 2007), the objective of this study is not statistical generalization but analytic generalization. Rather than testing predefined hypotheses, I use cases to inductively surface patterns, mechanisms, and boundary conditions that refine existing theories of the automation–augmentation paradox and task complementarity (Eisenhardt & Graebner, 2007 ). Recruitment provides a theoretically fertile context for this endeavor because it is a multi-stage decision system in which task characteristics, accountability structures, and legitimacy requirements vary systematically across stages. Importantly, this study conceptualizes cases not as illustrative examples of theory, but as theory-generating devices. By comparing how AI is deployed, adjusted, and sometimes partially withdrawn across recruitment stages and organizational contexts, I identify recurring configurations and transitions that enable theoretical extension beyond any single organization. 3.2 Case Selection and Theoretical Sampling Cases were selected using theoretical sampling, guided by the goal of maximizing conceptual insight rather than representativeness (Yin, 2018 ). Following Eisenhardt’s (1989) logic, organizations were included if they met three criteria. First, the organization had to demonstrate documented and sustained use of AI technologies in recruitment, rather than isolated pilots. This ensured that observed AI configurations reflected organizational routines rather than experimental anomalies. Second, sufficient publicly available and credible documentation had to exist to allow systematic reconstruction of AI deployment, adjustment, and outcomes across recruitment stages. Because AI-enabled recruitment is highly institutionalized and subject to public scrutiny, organizational disclosures, regulatory responses, and third-party investigations constitute an integral part of the phenomenon itself rather than a limitation of access. Third, cases needed to exhibit variation in AI configurations and observable transitions over time, including instances of automation, augmentation, and hybrid reconfiguration. This criterion was critical for theorizing the automation–augmentation paradox as a dynamic process rather than a static choice. Based on these criteria, the final case set includes large multinational organizations such as Unilever, Amazon, Hilton, and IBM, as well as specialized HR technology providers whose systems are embedded within organizational recruitment processes. This combination enables theoretical replication, as similar stage-level patterns recur across different organizational and technological contexts. Such a multiple-case design typically yields more robust, generalizable, and testable theory than single-case research (Eisenhardt & Graebner, 2007 ). 3.3 Data Sources and Case Materials The study relies on multiple secondary data sources, which are both appropriate and necessary given the institutionalized and publicly contested nature of AI in recruitment. Data sources include peer-reviewed academic studies, practitioner research reports, corporate disclosures, regulatory filings, technology vendor documentation, and credible journalistic investigations. Using secondary data is not a compromise but a methodological strength in this context. High-profile recruitment AI systems are routinely subject to public debate, regulatory intervention, and reputational risk, making organizational responses and adjustments visible through external documentation. These materials capture not only technical design choices but also legitimacy concerns, governance structures, and organizational sensemaking surrounding AI use. Data triangulation across diverse sources enhances construct validity and reduces reliance on single narratives. Where discrepancies emerged across sources, these tensions were treated as analytically informative, often signaling contested interpretations or shifting organizational priorities. 3.4 Analytical Framework: Stage-Based Mapping of AI Configurations To ensure analytical consistency, all cases were mapped onto a standardized recruitment framework derived from established personnel selection literature. This framework conceptualizes recruitment as a sequence of interdependent stages, ranging from applicant pool formation to post-hire validation. For each recruitment stage in each case, I analyzed AI deployment along four analytical dimensions. First, I identified the primary role of AI, classifying it as automation, augmentation, or hybrid human–AI collaboration. Second, I examined task characteristics, including analyzability, ambiguity, and error tolerance, drawing on the stage-based framework outlined earlier. Third, I analyzed organizational responses and outcomes, such as efficiency gains, changes in decision quality, emerging bias concerns, or legitimacy challenges. Fourth, I traced temporal dynamics, documenting whether and why organizations transitioned from automation to augmentation, from augmentation to automation, or toward hybrid reconfiguration. These transitions are central to theorizing the automation–augmentation paradox as a cyclical, stage-embedded process. Coding proceeded iteratively for transparency and trustworthy (Gioia, Corley and Hamilton, 2013 ; Pratt, Kaplan and Whittington, 2020 ). Initial open coding identified references to AI functionality, decision processes, organizational intent, and reported consequences. These codes were then consolidated into theory-informed categories drawing on paradox theory and task complementarity. Emerging interpretations were continuously compared across cases and stages to identify recurring patterns and deviations. This recursive process between data and emerging theory ensures that the resulting constructs and propositions are deeply grounded in empirical evidence (Eisenhardt & Graebner, 2007 ). 3.5 Ensuring Rigor and Analytical Validity Several strategies were employed to enhance the rigor of the study. Construct validity was strengthened through triangulation across multiple data sources and by anchoring analysis in a well-established recruitment framework. Internal validity was addressed by explicitly linking observed patterns to theoretical mechanisms rather than relying on post hoc interpretation. Reliability was supported through transparent documentation of case selection criteria, analytical dimensions, and coding logic (Yin, 2018 ). Although the study does not aim for statistical generalization, analytic generalization is achieved by demonstrating how recruitment-stage dynamics refine and extend broader theories of human–AI collaboration. By making the analytical framework explicit, the study enables future research to apply a similar approach to other HR functions characterized by institutional sensitivity. 4. Findings The comparative analysis reveals that AI configurations in recruitment vary systematically across stages. These variations are not driven by differences in technological maturity alone, but by recurring organizational mechanisms through which efficiency demands, legitimacy pressures, and accountability requirements are reconciled. Three stage-embedded mechanisms emerge consistently across cases. Table 2 provides a structured overview of representative organizational cases and illustrates how AI applications are configured as automation, augmentation, or hybrid arrangements across recruitment stages. Rather than serving as standalone case descriptions, the cases summarized in Table 2 function as empirical anchors that substantiate the stage-embedded mechanisms identified in the analysis. The findings reported below therefore abstract from firm-specific narratives and focus on recurring mechanisms evidenced across these cases. Table 2 Representative Organizational Cases of AI Application Across Selection Stages Selection Stage Firm Case AI Application Application Type Quantified Outcomes Key Implications Limitations Applicant Pool Formation Atlassian NLP-based job advertisement optimization (Textio) to reduce gendered language and broaden applicant pool Augmentation + 57% female technical hires within two years; significant increase in applicant diversity AI can shape who applies by influencing language and signaling inclusiveness at the attraction stage Effects limited to wording; does not address downstream screening or organizational bias Minimum Qualification Screening Unilever Automated rule-based and ML screening for minimum criteria (education, work eligibility, availability) Automation Recruiter screening time reduced by ~ 75%; processing over 250,000 applicants annually AI automation is highly effective for objective, low-ambiguity eligibility thresholds Risk of false negatives if minimum criteria are poorly specified or overly rigid Resume / CV Screening Amazon ML-based resume ranking trained on historical hiring data Automation System discontinued after systematic gender bias detected in rankings Demonstrates dangers of fully automated resume screening trained on biased historical data Lack of explainability and bias auditing undermined system legitimacy Preliminary Assessment Unilever AI-driven gamified cognitive and behavioral assessments (Pymetrics) Automation Time-to-hire reduced from ~ 4 months to ~ 4 weeks; ~70,000 recruiter hours saved annually Scalable early assessment can replace resumes and substantially reduce human workload Construct validity and cross-cultural fairness of gamified assessments remain debated Structured Interview Hilton Asynchronous AI video interviews (HireVue) for structured first-round interviews Automation Time-to-hire reduced from 42 days to ~ 5 days; interview completion rates increased AI enables standardization and scalability in structured interview administration Use of facial and voice analytics raises ethical, legal, and transparency concerns Work Sample / Job Simulation HackerRank AI-assisted coding challenges with automated scoring and plagiarism detection Automation Screening capacity scaled to tens of thousands of candidates; improved hiring manager satisfaction AI-scored work samples offer high job relevance with scalable evaluation May disadvantage candidates unfamiliar with platform-based or time-pressured testing Assessment Center Unilever AI-supported virtual assessment centers with behavioral scoring support Augmentation Assessment center costs reduced by ~ 50%; increased consistency across assessors AI can support evaluators by reducing cognitive load and increasing standardization Final judgments still rely on human assessors; limited transparency in scoring models Reference and Background Checks Checkr Automated AI-based background and compliance checks Automation Background check turnaround time reduced by ~ 30–50% AI excels in compliance-heavy, rule-based verification tasks Risk of outdated, incomplete, or erroneous records affecting candidates Final Selection Decision & Job Offer Eightfold AI AI-generated predictions of candidate fit, retention, and internal mobility Augmentation Client firms report improved quality-of-hire and internal fill rates (case evidence) AI is most appropriate as decision support—not decision maker—at final selection stage Predictive accuracy depends on data quality and organizational stability Validation and Evaluation IBM Continuous AI auditing of hiring models for bias, validity, and performance Augmentation Ongoing bias detection and model recalibration embedded in HR analytics Continuous validation closes the AI hiring loop and sustains organizational legitimacy Requires sustained data governance capability and analytical resources 4.1 Early Recruitment Stages: Efficiency-Driven Automation and Latent Risk Mechanism: Efficiency-driven automation sustained by high analyzability and deferred accountability. In early recruitment stages—such as applicant pool formation, minimum qualification screening, and initial résumé filtering—organizations consistently configure AI as automation. Tasks at these stages are highly standardized, rule-based, and scalable, enabling AI systems to substitute for human effort with minimal contestation. This configuration generates substantial efficiency gains through between-task complementarity: AI assumes routine screening functions, allowing human recruiters to be redeployed to downstream evaluative tasks where human judgment is more critical. Importantly, the legitimacy of automation at this stage is rarely challenged at the point of deployment. This is because errors are often perceived as reversible and decision stakes remain diffuse, leading to a state of deferred accountability. However, as illustrated by the Amazon case (Table 2 ), automation at this stage is not inherently stable or free from conflict. The analysis reveals two critical theoretical nuances regarding this fragile stability: Latent Legitimacy Risks : While automation appears stable due to low immediate accountability, it harbors latent risks. Automated systems trained on historical data may encode prior organizational biases—such as the gender bias found in Amazon’s ML-based ranking tool. These biases often remain hidden until they are surfaced by external scrutiny, regulatory attention, or retrospective audits. Conditional Stability based on Visibility : The perceived stability of early-stage automation is often a function of low error visibility rather than technical perfection. When an algorithmic failure becomes public, the perceived legitimacy collapses, triggering an immediate transition toward "re-augmentation" or the complete withdrawal of the system. Consequently, these retreats do not necessarily reflect technical breakdowns but rather a sudden shift in legitimacy evaluation. Automation in early recruitment stages must be reconceptualized not as a permanently stable state, but as a configuration of "fragile stability" that remains viable only as long as efficiency gains do not publicly collide with institutional fairness expectations. 4.2 Intermediate Recruitment Stages: Paradox Intensification and Hybrid Instability Mechanism: Paradox intensification arising from simultaneous demands for scalability, validity, and legitimacy. Intermediate recruitment stages—such as preliminary assessments, structured interviews, work samples, and assessment centers—consistently emerge as zones of instability. Tasks at these stages promise high predictive value while also carrying heightened ethical, legal, and experiential scrutiny. Organizations frequently attempt to automate these stages to achieve scale and standardization. However, the analysis shows that such efforts quickly encounter legitimacy challenges related to construct validity, bias, transparency, and candidate experience. As a result, organizations rarely sustain fully automated configurations at intermediate stages. Instead, hybrid configurations proliferate. AI systems generate standardized scores or rankings, while humans retain responsibility for interpretation and integration. Yet these hybrids remain unstable. Humans struggle to consistently determine when algorithmic outputs should be trusted or overridden, particularly in socially complex evaluative contexts. This difficulty undermines within-task complementarity and triggers repeated recalibration. Across cases, intermediate stages exhibit oscillation rather than convergence. Organizations alternate between automation and augmentation, adjusting AI autonomy in response to emerging concerns without settling into a stable equilibrium. Paradox intensification thus reflects not managerial indecision, but the structural incompatibility of competing demands at this stage. 4.3 Late Recruitment Stages: Accountability-Driven Augmentation Mechanism: Institutionalized augmentation anchored by accountability and responsibility attribution. In late recruitment stages—such as final selection decisions, job offers, and validation—AI is consistently configured as augmentation rather than automation. Even when AI systems demonstrate strong predictive capability, decision authority remains explicitly human. The analysis shows that this pattern is driven by accountability rather than performance considerations. Late-stage decisions are highly visible, contestable, and consequential, making organizations acutely sensitive to legal defensibility and moral responsibility. Delegating final authority to AI is widely perceived as illegitimate, regardless of technical adequacy. This institutional stabilization occurs even when AI predictive accuracy is reported to exceed human judgment, as seen in cases like Eightfold AI. The persistence of human authority in these settings is driven by the fundamental requirement for moral responsibility and legal defensibility, ensuring that high-stakes outcomes remain attributable to human actors rather than opaque algorithmic processes. As a result, AI systems at this stage are embedded as decision-support tools that inform but do not determine outcomes. Augmentation at late stages is notably more stable than hybrid configurations observed earlier, reflecting the clarity of institutional expectations regarding human accountability. This finding demonstrates that limits on AI autonomy in recruitment are not merely technical or cognitive, but normative and institutional. 4.4 Re-Augmentation as a Cross-Stage Legitimacy Repair Mechanism Mechanism: Re-augmentation as organizational response to legitimacy threats. Across all recruitment stages, a recurring dynamic emerges: organizations repeatedly reintroduce human oversight following periods of increased automation. This pattern—observed in early and intermediate stages in particular—is triggered less by declining performance than by perceived legitimacy threats. Instances of re-augmentation occur when automated systems generate outcomes that are difficult to justify, explain, or defend. In response, organizations recalibrate AI roles by narrowing autonomy, increasing human review, or reframing AI outputs as advisory rather than determinative. Crucially, re-augmentation does not entail abandoning AI. Instead, it functions as a legitimacy repair mechanism that restores trust and defensibility while preserving selected efficiency gains. Through re-augmentation, organizations stabilize AI-enabled recruitment systems by realigning technological capability with institutional expectations. 4.5 Summary of Stage-Embedded Mechanisms Taken together, these findings demonstrate that AI configurations in recruitment are governed by stage-specific mechanisms rather than uniform organizational strategies. Specifically, early stages stabilize automation under conditions of high task analyzability and low immediate accountability; intermediate stages intensify paradoxical tensions that resist stabilization; and late stages institutionalize augmentation due to stringent accountability constraints. Within this framework, re-augmentation serves as a cross-stage corrective mechanism triggered whenever the legitimacy of a configuration is threatened. Figure 1 illustrates the dynamic interaction between automation and augmentation throughout the AI-enabled recruitment lifecycle. The mechanisms occurring during the early stage (Screening) warrant particular attention, as they reveal both the general tendency toward automation and the potential for exceptional deviations. Typically, the initial recruitment stage is characterized by high task analyzability, favoring the establishment of stable, efficiency-driven automation models. However, as evidenced by the case of Amazon’s AI recruitment tool, this trajectory can shift abruptly upon the detection of "algorithmic bias"—such as the generation of gender-biased results. Such technical flaws not only erode trust in the AI system but also compel the organization to confront an acute demand for institutional accountability. In such instances, the established automation path fractures, triggering a 're-augmentation loop.' The Amazon case underscores that even in a supposedly stabilized early stage, a re-augmentation loop can be activated instantaneously. To restore legitimacy, organizations must make strategic choices to reinforce 'human accountability' (represented by the blue line in Fig. 1 ). This involves forcibly reducing 'algorithmic autonomy' (the red line in Fig. 1 ) and mandating intervention by human experts to correct biases or, in extreme cases, decommission the system entirely. Ultimately, these mechanisms suggest that re-augmentation loops are not confined to the intermediate stages of assessment; rather, they can emerge at any point in the recruitment journey where the legitimacy of automation is compromised. This serves as a powerful restorative force that reverts the system back toward an augmentation model. Consequently, the configuration of AI-enabled recruitment should not be viewed as a static end-state, but as a perpetually reconstructed and dynamic process. These stage-embedded mechanisms explain why organizations repeatedly transition between automation and augmentation without ever converging on a single, permanent configuration. 5. Discussion 5.1 Reconceptualizing the Automation–Augmentation Paradox as Stage-Embedded This study set out to examine how automation and augmentation are configured across recruitment stages and why organizations repeatedly transition between these configurations over time. The findings demonstrate that the automation–augmentation paradox in recruitment is neither a temporal sequence nor an organizational-level choice. Instead, it is structurally embedded across recruitment stages. Prior research has predominantly conceptualized the automation–augmentation paradox as unfolding over time at the firm level, suggesting cycles of automation followed by augmentation or vice versa. While this temporal perspective captures important dynamics, it obscures how contradictory logics are simultaneously enacted within organizational processes. By shifting the unit of analysis from organizations to recruitment stages, this study shows that automation and augmentation coexist within a single hiring system, configured differently across stages according to task characteristics and institutional exposure. In doing so, the study aligns with broader organizational research showing that algorithmic governance reshapes control and coordination structures rather than simply substituting human labor (Kellogg et al., 2020 ). Recruitment stages therefore function as micro-sites of algorithmic governance where authority is continuously negotiated. This stage-embedded view extends paradox theory by demonstrating that paradoxical tensions need not be resolved or balanced temporally. Instead, organizations manage paradox spatially by distributing automation and augmentation across interdependent stages. Recruitment thus becomes a site where paradox is continuously enacted rather than episodically addressed. 5.2 Re-Augmentation as a Legitimacy Repair Mechanism A second contribution concerns how this study interprets organizational retreats from automation. Existing research often treats reversals from automation as evidence of technological failure, implementation error, or resistance to change. The findings challenge this interpretation. Across cases, organizations repeatedly reintroduced human oversight after periods of increased automation, particularly in response to concerns about bias, transparency, or accountability. Importantly, these reversals did not entail abandoning AI altogether. Instead, organizations recalibrated AI’s role to restore legitimacy while retaining selected efficiency gains. This pattern is conceptualized here as re-augmentation. Re-augmentation functions as a legitimacy repair mechanism through which organizations reassert human responsibility and moral accountability when algorithmic systems threaten institutional expectations. Rather than representing regression, re-augmentation stabilizes AI-enabled recruitment by aligning technological capability with social and legal norms. By foregrounding legitimacy as a central driver of AI reconfiguration, this study extends paradox theory beyond efficiency–learning trade-offs and highlights the institutional work organizations perform to sustain AI use in high-stakes HR contexts. 5.3 Refining Task Complementarity Theory: Accountability-Constrained Complementarity This study also refines task complementarity theory by demonstrating that complementarity in recruitment is institutionally bounded. Consistent with task complementarity theory, the findings show that automation generates value through between-task complementarity in early recruitment stages characterized by high analyzability and error tolerance. However, within-task complementarity—central to augmentation—proves far more fragile. At intermediate stages, hybrid configurations emerge, but organizations struggle to stabilize them. Humans face persistent difficulty in consistently determining when algorithmic recommendations should be trusted or overridden in socially complex evaluative tasks. As a result, organizations oscillate between automation and augmentation rather than converging on a stable hybrid equilibrium. Most critically, the findings demonstrate that within-task complementarity is structurally constrained at late recruitment stages. Even when AI systems demonstrate strong predictive performance, decision authority remains explicitly human. This constraint is not primarily cognitive or technical, but institutional: responsibility for hiring decisions is legally and normatively assigned to human actors. Accountability therefore limits AI autonomy regardless of performance. Recent HRM research similarly argues that reducing algorithmic bias requires sustained human accountability, cross-stage governance integration, and continuous oversight, reinforcing the structural limits of fully autonomous AI in recruitment (Soleimani et al., 2025 ). This finding resonates with sociotechnical critiques of 'abstraction traps'—the notion that attempting to isolate technical performance from institutional contexts obscures the social allocation of responsibility (Selbst et al., 2019). In high-stakes settings like recruitment, algorithmic autonomy is constrained by the threshold of socially acceptable responsibility transfer. While regulatory frameworks such as GDPR Article 22 or New York City Local Law 144 impose transparency and audit requirements, their primary influence on AI configuration is exercised through the mandate for human oversight (Roig, 2017 ). By introducing accountability as a boundary condition, this study extends task complementarity theory beyond cognitive interaction to encompass institutional responsibility as a central determinant of human–AI collaboration. 5.4 Propositions for Case-Based Theory Building Synthesizing the findings, this study advances four propositions that extend existing research on AI, HRM, and organizational decision-making: Proposition 1 (Stage-Embedded Configuration) : Automation and augmentation in recruitment are configured at the stage level rather than uniformly at the organizational level, resulting in the simultaneous coexistence of multiple human–AI configurations within a single hiring system. Proposition 2 (Legitimacy-Bounded Automation) : Automation in recruitment stabilizes only in stages where task analyzability and legitimacy requirements remain aligned; misalignment triggers reconfiguration toward augmentation. Proposition 3 (Re-Augmentation as Legitimacy Repair) : Re-augmentation emerges as a legitimacy repair mechanism through which organizations reassert human oversight in response to ethical, legal, or reputational challenges arising from automated recruitment systems. Proposition 4 (Accountability-Constrained Complementarity) : Within-task human–AI complementarity is structurally constrained in high-accountability recruitment decisions, limiting AI autonomy regardless of technical performance. These propositions integrate paradox theory and task complementarity within a stage-embedded framework, offering a process-sensitive account of AI governance in institutionally exposed HR systems. offering a more nuanced account of human–AI collaboration in institutionally sensitive organizational processes. 5.5 Implications The findings of this study offer critical insights for human resource management (HRM), human resource development (HRD), and the formulation of policies for AI governance. From an academic and managerial perspective, the results underscore the necessity of a stage-sensitive approach that views recruitment not as a single event, but as an adaptive system where human–AI configurations evolve based on specific task characteristics and temporal dynamics. AI-enabled recruitment should be repositioned as a system for legitimacy management and institutional learning rather than a mere deployment of efficiency-driven technology. Consequently, the benefits of augmentation depend heavily on HRD investments—specifically in recruiter training, the development of interpretive skills, ethical literacy, and the establishment of governance routines that facilitate meaningful human engagement with AI systems. Furthermore, by integrating paradox theory and task complementarity within the HRM context, this study demonstrates how fine-grained analysis of HR processes can refine macro-level theories concerning AI and the future of work. Beyond organizational practice, this stage-embedded theory provides concrete guidance for policymakers and regulatory bodies overseeing AI in the labor market. Rather than applying a monolithic regulatory framework to all AI applications, oversight should be differentiated based on the institutional sensitivity of each stage. In late-stage decisions where legal and normative accountability is paramount, regulations must ensure that "meaningful human control" is substantively exercised, regardless of the AI's predictive accuracy. For intermediate stages, which are characterized by high instability between efficiency and fairness, regulators should prioritize monitoring whether human evaluators possess the actual authority and capability to critically review and override algorithmic outputs. In early recruitment stages, where automation is prevalent, policy should focus on institutionalizing periodic external audits and data lineage verification to detect latent biases before they manifest as systemic exclusion. Moreover, technical interventions must move beyond simple data collection toward methods that mitigate discrimination without relying on sensitive attributes, thereby aligning technical performance with evolving human rights standards. Such 'fairness-by-design' approaches are essential for sustaining the long-term legitimacy of AI-enabled recruitment systems. Finally, "re-augmentation" should be recognized by regulators as a formal safety valve and a legitimate mechanism for restoring institutional trust, encouraging firms to adopt proactive protocols for reintroducing human oversight when algorithmic outcomes threaten social or ethical norms. 6. Conclusion This study argues that the central challenge of AI in recruitment is not choosing between automation and augmentation, nor progressing linearly toward greater algorithmic autonomy. Instead, the challenge lies in governing their configuration across recruitment stages and over time. By conceptualizing recruitment as a stage-embedded automation–augmentation system, this study advances theory in three ways. First, it reconceptualizes the automation–augmentation paradox as spatially embedded within organizational processes rather than temporally resolved at the firm level. Second, it refines task complementarity theory by demonstrating that accountability, not just task analyzability or cognitive interaction, constrains human–AI collaboration in high-stakes HR decisions. Third, it introduces re-augmentation as a theoretically meaningful legitimacy mechanism through which organizations stabilize AI-enabled recruitment systems. From a practical perspective, the findings caution against one-size-fits-all AI strategies in recruitment. Organizations that pursue aggressive automation across all stages risk triggering cycles of legitimacy loss and trust erosion. Conversely, organizations that rely exclusively on augmentation may forgo scalability and efficiency gains. Effective recruitment AI strategies therefore require deliberate differentiation across stages and continuous recalibration in response to institutional pressures. Beyond recruitment, the stage-embedded configuration identified here may illuminate AI governance challenges in other high-stakes domains such as healthcare triage, credit scoring, and judicial decision-making, where accountability cannot be meaningfully automated. By theorizing accountability as a structural ceiling on delegation, this study contributes to ongoing debates on responsible AI and institutional trust. This study has limitations, including its reliance on secondary data and publicly documented cases. Future research could extend this work through longitudinal field studies, cross-national comparisons, or experimental designs examining how candidates and decision-makers perceive stage-specific AI use (Pettigrew, 1990 ). As AI continues to reshape HRM, understanding how automation and augmentation are governed—not merely adopted—will be essential for building legitimate and sustainable hiring systems. Declarations Ethics Statement Ethical approval was not required as the study did not involve human participants. Informed consent This article does not contain any studies with human participants performed by any of the authors. Conflict of interest The author declare that there is no conflict of interest in this paper. Author Contribution H.L.: Conceptualization; Formal Analysis; Data Curation; Writing – Original Draft; Methodology; Supervision; Validation; Writing – Review & Editing. Data Availability Data used in this study consist only of publicly available secondary data cited in the references. No primary or proprietary datasets were used, and all sources are accessible via the references. References Arthur W Jr, Day EA, McNelly TL, Edens PS (2003) A meta-analysis of the criterion‐related validity of assessment center dimensions. Pers Psychol 56:125–153 Baron J (2018) New continuous background checks will take a deeper look at your personal life. Forbes (10 December Breaugh JA (2013) Employee recruitment. Ann Rev Psychol 64:389–416 Campion MA, Palmer DK, Campion JE (1997) A review of structure in the selection interview. Pers Psychol 50:655–702 Cole MS, Feild HS, Giles WF (2004) Interaction of recruiter and applicant gender in resume evaluation: a field study. Sex roles 51:597–608 Dastin J (2022) Ethics of data and analytics 296–299. Auerbach Dietvorst BJ, Simmons JP, Massey C (2015) Algorithm aversion: people erroneously avoid algorithms after seeing them err. J Exp Psychol Gen 144:114 Dineen BR, Yu KYT, Stevenson-Street J (2023) Recruitment in personnel psychology and beyond: Where we've been working, and where I might work next. Pers Psychol 76:617–650 Eisenhardt KM, Graebner ME (2007) Theory building from cases: Opportunities and challenges. Acad Manag J 50:25–32 Faraj S, Pachidi S, Sayegh K (2018) Working and organizing in the age of the learning algorithm. Inf Organ 28:62–70 Fügener A, Walzner DD, Gupta A (2026) Roles of artificial intelligence in collaboration with humans: Automation, augmentation, and the future of work. Manage Sci 72:538–557 Gatewood RD, Feild HS, Barrick M (2020) Human Resource Selection, 9th edn. Cengage Learning Gioia DA, Corley KG, Hamilton AL (2013) Seeking qualitative rigor in inductive research: Notes on the Gioia methodology. Organizational Res methods 16:15–31 Glikson E, Woolley AW (2020) Human trust in artificial intelligence: Review of empirical research. Acad Manag Ann 14:627–660 HackerRank (2023) How HackerRank catches AI-generated code by deploying advanced ML plagiarism detection. HackerRank Blog (7 June Hunkenschroer AL, Luetge C (2022) Ethics of AI-enabled recruiting and selection: A review and research agenda. J Bus Ethics 178:977–1007 IBM AI in recruitment. IBM Think https://www.ibm.com/think/topics/ai-in-recruitment Jarrahi MH (2018) Artificial intelligence and the future of work: Human-AI symbiosis in organizational decision making. Bus Horiz 61:577–586 Kellogg KC, Valentine MA, Christin A (2020) Algorithms at work: The new contested terrain of control. Acad Manag Ann 14:366–410 Kurter H How Hilton reduced their time to hire from 43 days down to 5. Forbes (19 September 2019) Lee MK (2018) Understanding perception of algorithmic decisions: Fairness, trust, and emotion in response to algorithmic management. Big data Soc 5:2053951718756684 Logg JM, Minson JA, Moore DA (2019) Algorithm appreciation: People prefer algorithmic to human judgment. Organ Behav Hum Decis Process 151:90–103 Marino DL et al in 2020 13th International Conference on Human System Interaction (HSI). 155–161 (IEEE) Marr B (2019) Artificial intelligence in practice: how 50 successful companies used AI and machine learning to solve problems. Wiley Meijerink J, Boons M, Keegan A, Marler J (2021) Algorithmic human resource management: Synthesizing developments and cross-disciplinary insights on digital HRM. Int J Hum Resour Manag 32:2605–2621 Newman DT, Fast NJ, Harmon DJ (2020) When eliminating bias isn’t fair: Algorithmic reductionism and procedural justice in human resource decisions. Organ Behav Hum Decis Process 160:149–167 Pettigrew AM (1990) Longitudinal field research on change: Theory and practice. Organ Sci 1:267–292 Pratt MG, Kaplan S, Whittington R (2020) Editorial essay: The tumult over transparency: Decoupling transparency from replication in establishing trustworthy qualitative research. Adm Sci Q 65:1–19 Pumpedu The practical application of AI: Unilever reduced recruitment time by 75%. Pumpedu News https://www.pumpedu.cz/en/news/the-practical-application-of-ai-unilever-reduced-recruitment-time-by-75-112n Raisch S, Krakowski S (2021) Artificial intelligence and management: The automation–augmentation paradox. Acad Manage Rev 46:192–210 Reuters (2026) AI company Eightfold sued for helping companies secretly score job seekers. Reuters (21 January https://www.reuters.com/sustainability/boards-policy-regulation/ai-company-eightfold-sued-helping-companies-secretly-score-job-seekers-2026-01-21/ Roig A (2017) Safeguards for the right not to be subject to a decision based solely on automated processing (Article 22 GDPR). Eur J Law Technol 8 Roth PL, Bobko P, McFarland LA (2005) A meta-analysis of work sample test validity: Updating and integrating some classic literature. Pers Psychol 58:1009–1037 Rynes SL, Bretz RD Jr, Gerhart B (1991) The importance of recruitment in job choice: A different way of looking. Pers Psychol 44:487–521 Sackett PR, Zhang C, Berry CM, Lievens F (2022) Revisiting meta-analytic estimates of validity in personnel selection: Addressing systematic overcorrection for restriction of range. J Appl Psychol 107:2040 Schmidt FL, Hunter JE (1998) The validity and utility of selection methods in personnel psychology: Practical and theoretical implications of 85 years of research findings. Psychol Bull 124:262 Selbst AD, Boyd D, Friedler SA, Venkatasubramanian S, Vertesi J in Proceedings of the conference on fairness, accountability, and transparency. 59–68 Smith WK, Lewis MW (2011) Toward a theory of paradox: A dynamic equilibrium model of organizing. Acad Manage Rev 36:381–403 Soleimani M, Intezari A, Arrowsmith J, Pauleen DJ, Taskin N (2025) Reducing AI bias in recruitment and selection: an integrative grounded approach. Int J Hum Resource Manage 36:2480–2515 Tambe P, Cappelli P, Yakubovich V (2019) Artificial intelligence in human resources management: Challenges and a path forward. Calif Manag Rev 61:15–42 Vrontis D et al (2023) Artificial intelligence, robotics, advanced technologies and human resource management: a systematic review. Artif Intell Int HRM, 172–201 Yin RK (2018) Case Study Research and Applications: Design and Methods, 6th edn. SAGE Additional Declarations No competing interests reported. Supplementary Files SupplementaryMethodologicalDocumentation.docx Cite Share Download PDF Status: Under Review Version 1 posted Editorial decision: Revision requested 04 Apr, 2026 Editor assigned by journal 04 Apr, 2026 Submission checks completed at journal 01 Apr, 2026 First submitted to journal 25 Mar, 2026 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-9222046","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":611950418,"identity":"affc2765-d4e9-4f47-b466-36cd39694a40","order_by":0,"name":"Hochan Lee","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA60lEQVRIie3OsWrCQBzH8Z8cxCUl6/+Qxle4EFALPszdki59gG4NBM6l7oIv4eimJXBZ7gGULk7ZBEcLKTRFt8qpW4f7Ln/uDx/uD/h8/zABBO0YAV0NyNOS3UIICM0vEfcQys7va2TYLerd15LiiNfqsGsaRJM1S18d5OndDJOppZTPs5KUFiArmbKuwzYvAT1oUovP55xU3h62AfvIrxD+reltta2Ko2wE+reQXvuLFBQYkoFoN2DKSawZ9B41JTObZSOl0zCxqkicpCpqvtfjfjQx6fbYxHFclSV3kT+FQOcu4PP5fL4L/QDmKUhR/DxOywAAAABJRU5ErkJggg==","orcid":"","institution":"Korea Advanced Institute of Science and Technology","correspondingAuthor":true,"prefix":"","firstName":"Hochan","middleName":"","lastName":"Lee","suffix":""}],"badges":[],"createdAt":"2026-03-25 10:54:27","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-9222046/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-9222046/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":105886459,"identity":"dfd38aef-fe61-4aab-a897-7c0ea1b97bf7","added_by":"auto","created_at":"2026-04-01 07:29:35","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":165214,"visible":true,"origin":"","legend":"\u003cp\u003eA Stage-Embedded Model of the Automation–Augmentation Paradox in AI-Enabled Recruitment\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-9222046/v1/b15a5f55bd4409e68176ce37.png"},{"id":105886548,"identity":"d60eec03-9c3b-4a15-8d76-3ce20f8cf2ed","added_by":"auto","created_at":"2026-04-01 07:29:55","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1492520,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-9222046/v1/6366c0e0-ca5b-49b8-b139-6cce8dfaf428.pdf"},{"id":105886494,"identity":"44ecc7b4-7416-4f97-9f14-21662d7b4bf2","added_by":"auto","created_at":"2026-04-01 07:29:47","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"supplement","size":28358,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryMethodologicalDocumentation.docx","url":"https://assets-eu.researchsquare.com/files/rs-9222046/v1/dfcf7046441ebefddcec20ac.docx"}],"financialInterests":"No competing interests reported.","formattedTitle":"Reconceptualizing the Automation–Augmentation Tension in AI-Enabled Talent Selection: A Stage-Embedded Theory","fulltext":[{"header":"1. Introduction","content":"\u003cp\u003eRecruitment is one of the most consequential and institutionally sensitive functions in human resource management (HRM). Hiring decisions shape organizational performance, workforce diversity, and long-term capability development, while simultaneously exposing organizations to legal, reputational, and ethical scrutiny (Breaugh, \u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e2013\u003c/span\u003e; Gatewood, Feild and Barrick, \u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e2020\u003c/span\u003e). Because recruitment determines who enters the organization\u0026mdash;and on what grounds\u0026mdash;it remains a focal site for debates about fairness, accountability, and legitimacy (Rynes, Bretz and Gerhart, \u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e1991\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eAgainst this backdrop, artificial intelligence (AI) has become deeply embedded in recruitment and selection processes. Organizations increasingly rely on algorithmic systems to screen applicants, administer assessments, analyze interviews, and generate predictive insights regarding future performance or retention (Tambe, Cappelli and Yakubovich, \u003cspan citationid=\"CR40\" class=\"CitationRef\"\u003e2019\u003c/span\u003e). These technologies promise substantial efficiency and scalability gains, particularly in contexts characterized by high applicant volumes and time pressure (Vrontis et al., \u003cspan citationid=\"CR41\" class=\"CitationRef\"\u003e2023\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eAt the same time, AI-enabled recruitment remains highly contested, with recurring concerns about bias, transparency, explainability, and accountability (Hunkenschroer \u0026amp; Luetge, \u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e2022\u003c/span\u003e). Recent integrative research further demonstrates that AI bias in recruitment cannot be reduced to technical malfunction but reflects the interaction of data histories, organizational routines, and governance arrangements, thereby requiring sustained human oversight and institutional alignment (Soleimani et al., \u003cspan citationid=\"CR39\" class=\"CitationRef\"\u003e2025\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eMuch of the debate surrounding AI in recruitment has been framed through a distinction between automation and augmentation. Automation emphasizes delegating decision tasks to algorithms in pursuit of efficiency, consistency, and standardization, whereas augmentation frames AI as decision support that enhances human judgment while preserving human discretion and responsibility (Raisch and Krakowski, \u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e2021\u003c/span\u003e). This distinction mirrors a broader debate in management research regarding whether AI substitutes for or complements human labor (Faraj, Pachidi and Sayegh, \u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e2018\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eRecent scholarship challenges this binary framing by conceptualizing automation and augmentation as a paradoxical relationship rather than mutually exclusive alternatives. From a paradox perspective, automation and augmentation are contradictory yet interdependent: automation can generate efficiency gains that enable higher-order human judgment, while augmentation can foster learning and trust that stabilize future automation (Smith and Lewis, \u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e2011\u003c/span\u003e; Raisch and Krakowski, \u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e2021\u003c/span\u003e). Task complementarity theory further suggests that performance benefits depend on how humans and AI are configured at the task level rather than on automation or augmentation per se (F\u0026uuml;gener, Walzner and Gupta, 2025).\u003c/p\u003e \u003cp\u003eDespite these advances, existing theorizing remains insufficiently grounded in the institutional realities of recruitment. Hiring differs from many organizational domains in that decision outcomes are highly consequential, socially sensitive, and externally scrutinized (Dineen, Yu and Stevenson, 2023). Responsibility for selection decisions is normatively and legally assigned to human actors, even when algorithmic systems play a substantial role (Meijerink et al., \u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e2021\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eAs a result, AI use in recruitment cannot be understood solely in terms of technical performance or cognitive complementarity; it is also shaped by legitimacy expectations and accountability structures (Newman, Fast and Harmon, \u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e2020\u003c/span\u003e). More broadly, research on algorithms at work demonstrates that algorithmic systems do not merely replace tasks but reorganize authority structures, accountability relations, and worker autonomy within organizations (Kellogg, Valentine and Christin, \u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e2020\u003c/span\u003e). Recruitment therefore represents not only a technical deployment of AI, but a reconfiguration of institutional authority in high-stakes decision-making.\u003c/p\u003e \u003cp\u003eA further limitation of prior research is its tendency to treat recruitment as a monolithic process. Personnel selection research has long emphasized that hiring unfolds through a sequence of stages\u0026mdash;ranging from applicant attraction and screening to assessment, final decision-making, and validation\u0026mdash;each designed to manage trade-offs among predictive validity, efficiency, and risk (Schmidt and Hunter, \u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e1998\u003c/span\u003e; Sackett et al., \u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e2022\u003c/span\u003e). Early stages prioritize efficiency and scalability, while later stages involve judgment-intensive evaluations and heightened accountability. Intermediate stages combine both, making them especially prone to tension and instability.\u003c/p\u003e \u003cp\u003eThis study argues that the automation\u0026ndash;augmentation paradox in recruitment is not merely temporal or organizational but structurally embedded across recruitment stages. Organizations rarely adopt a single, stable AI configuration. Instead, they deploy automation and augmentation simultaneously across stages and repeatedly reconfigure AI roles over time. Notably, retreats from automation often reflect legitimacy and accountability concerns rather than technical failure, echoing evidence on algorithm aversion and declining trust following perceived errors (Dietvorst, Simmons and Massey, \u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e2015\u003c/span\u003e; Logg, Minson and Moore, \u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e2019\u003c/span\u003e). More broadly, empirical research demonstrates that trust in AI is highly sensitive to perceptions of reliability, transparency, and contextual accountability, particularly in consequential decision settings (Glikson and Woolley, \u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e2020\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eTo capture these dynamics, this study adopts a comparative, case-based theory-building approach to examine how AI is configured across recruitment stages and why organizations transition between automation, augmentation, and hybrid configurations. The study addresses the following research question: How are automation and augmentation dynamically configured across recruitment stages, and why do organizations transition between these configurations over time?\u003c/p\u003e"},{"header":"2. Theoretical Background","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003e2.1 Recruitment as a Stage-Based and Institutionally Sensitive Decision System\u003c/h2\u003e \u003cp\u003ePersonnel selection research has consistently emphasized that recruitment unfolds as a sequence of interdependent stages rather than as a single evaluative decision (Breaugh, \u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e2013\u003c/span\u003e; Schmidt and Hunter, \u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e1998\u003c/span\u003e). Each stage is designed to manage trade-offs among predictive validity, efficiency, and risk, while decisions made at earlier stages constrain the candidate pool available at later ones (Sackett et al., \u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e2022\u003c/span\u003e). Recruitment is therefore inherently path-dependent.\u003c/p\u003e \u003cp\u003eBeyond technical considerations, recruitment stages differ systematically in their institutional exposure. Early stages involve high volumes, standardized information, and relatively low immediate decision stakes, allowing for greater tolerance of error and experimentation (Gatewood, Feild and Barrick, \u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e2020\u003c/span\u003e). Later stages, by contrast, are characterized by judgment-intensive evaluations, heightened legal and reputational risk, and strong normative expectations that humans remain accountable for outcomes (Rynes, Bretz and Gerhart, \u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e1991\u003c/span\u003e). Intermediate stages combine elements of both, positioning them as sites of heightened tension between efficiency and legitimacy.\u003c/p\u003e \u003cp\u003eThis stage-based differentiation implies that recruitment is best understood as an institutionally embedded decision system. Task characteristics and accountability expectations vary across stages, shaping not only how decisions are made but also what forms of delegation to AI are considered legitimate and defensible ex post. Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e summarizes the standardized selection process and delineates stage-specific differences in core purpose, evaluation logic, and institutional exposure, thereby providing the analytical foundation for the stage-embedded perspective developed in this study.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eStandardized Selection Process and Stage-Specific Task and Institutional Characteristics\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"7\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eStage\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eSelection Stage\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eCore Purpose\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eKey Tools \u0026amp; Techniques\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eEvaluation Characteristics\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eInstitutional \u0026amp; Task Characteristics\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c7\"\u003e \u003cp\u003eReference\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eApplicant Pool Formation\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eTo attract and assemble a sufficiently large and qualified pool of applicants, thereby setting the upper bound of potential selection quality\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eJob advertisements, employer branding, employee referrals, online recruitment platforms, labor market intermediaries\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eIndirect predictor of selection quality; determines base rates and range restriction; susceptible to self-selection bias\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eHigh volume; highly analyzable; high error tolerance; low accountability exposure; low institutional sensitivity\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eBreaugh (\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e2013\u003c/span\u003e); Rynes et al. (1990)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMinimum Qualification Screening\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eTo eliminate applicants who do not meet essential job-related requirements prior to resource-intensive assessment\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eDegree and certification checks, experience thresholds, legal eligibility verification, automated applicant tracking systems (ATS)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eLow predictive validity; high efficiency; high risk of false negatives; often rule-based\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eVery high analyzability; high automation potential; low immediate accountability; legitimacy risks largely latent\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eGatewood et al. (\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e2020\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eResume / CV Screening\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eTo evaluate applicants\u0026rsquo; job-relevant background and experiences as an initial indicator of person\u0026ndash;job fit\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eManual resume review, competency checklists, keyword matching, AI-based resume parsing and ranking\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eModerate reliability; limited criterion-related validity; vulnerable to human and algorithmic bia`s\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003ePattern-recognition task; moderate ambiguity; increasing legitimacy scrutiny; moderate accountability\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eCole et al. (\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e2004\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003ePreliminary Assessment\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eTo assess applicants\u0026rsquo; general mental ability, personality traits, and basic competencies linked to future job performance\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eCognitive ability tests, personality inventories (e.g., Big Five), integrity tests, situational judgment tests (SJTs)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eHigh predictive validity (especially cognitive ability); standardized; scalable; potential adverse impact\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eModerate analyzability; growing ethical and fairness concerns; moderate\u0026ndash;high institutional sensitivity\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eSchmidt et al. (1998); Sackett et al. (\u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e2022\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eStructured Interview\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eTo systematically evaluate applicants\u0026rsquo; job-relevant knowledge, skills, abilities, and motivation using standardized questions\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eBehavioral interviews, situational interviews, structured scoring rubrics, interviewer training\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eHigh reliability and validity when structured; reduced bias vs. unstructured interviews; moderate cost\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eLow analyzability; judgment-intensive; high accountability; strong legitimacy expectations\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eCampion et al. (\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e1997\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e6\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eWork Sample / Job Simulation\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eTo directly observe applicants\u0026rsquo; ability to perform tasks that closely resemble actual job duties\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eWork sample tests, job simulations, coding tasks, case analyses, in-basket exercises\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eHigh content and criterion-related validity; strong face validity; limited scalability; high development cost\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eMedium analyzability; low error tolerance; high legitimacy and defensibility requirements\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eRoth et al. (\u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e2005\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e7\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAssessment Center\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eTo evaluate multiple competencies through simulations and exercises assessed by multiple raters\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eGroup discussions, role plays, leaderless group tasks, presentations, multi-rater assessments\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eMultidimensional assessment; high developmental value; resource-intensive; validity design-dependent\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eComplex social evaluation; very high institutional sensitivity; high accountability diffusion\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eArthur et al. (\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e2003\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e8\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eReference and Background Checks\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eTo verify applicants\u0026rsquo; prior performance, qualifications, and potential risk factors\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eReference interviews, employment verification, criminal background checks, credential validation\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eLow incremental validity; confirmatory rather than predictive; legal and ethical constraints\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eLow analyzability; high legal exposure; primarily legitimacy- and compliance-driven\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eGatewood et al. (\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e2020\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e9\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eFinal Selection Decision \u0026amp; Job Offer\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eTo integrate information from all prior stages and select the candidate with the highest expected utility\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eCompensatory decision models, weighted scoring algorithms, human judgment panels, decision-support systems\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eDecision quality depends on weighting strategy; human\u0026ndash;algorithm interaction critical; risk of judgmental bias\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eVery low analyzability; minimal error tolerance; very high accountability; human authority expected\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eSchneider \u0026amp; Schmitt (1976)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e10\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eValidation and Evaluation\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eTo assess and improve the effectiveness, fairness, and legal defensibility of the selection system\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eCriterion-related validation studies, utility analysis, adverse impact analysis, ongoing performance monitoring\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eEssential for long-term effectiveness; ensures compliance and continuous improvement; often underutilized\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eMeta-level governance task; institutional learning mechanism; accountability reinforcement\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eSackett et al. (\u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e2022\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec4\" class=\"Section2\"\u003e \u003ch2\u003e2.2 Automation and Augmentation in AI Research\u003c/h2\u003e \u003cp\u003eResearch on AI in organizations has increasingly framed automation and augmentation as a paradoxical relationship rather than as mutually exclusive strategies (Raisch and Krakowski, \u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e2021\u003c/span\u003e). From a paradox perspective, automation and augmentation are contradictory yet interdependent: automation emphasizes efficiency, standardization, and scalability, whereas augmentation preserves human judgment, discretion, and responsibility. These logics cannot be fully reconciled but must be continuously managed (Smith and Lewis, \u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e2011\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eRecent sociotechnical scholarship further argues that algorithmic systems operate within layered social infrastructures, where technical optimization and institutional legitimacy frequently diverge (Selbst et al., 2019). This divergence is particularly salient in recruitment, where predictive accuracy alone cannot secure normative acceptance.\u003c/p\u003e \u003cp\u003eMost prior work conceptualizes this paradox at the organizational or temporal level, examining how firms shift between automation-oriented and augmentation-oriented strategies over time. However, such approaches implicitly assume that automation and augmentation are resolved or balanced at the firm level.\u003c/p\u003e \u003cp\u003eIn recruitment, this assumption is problematic. Because recruitment consists of multiple stages with sharply differing task structures and institutional expectations, automation and augmentation are rarely adopted uniformly. Instead, contradictory logics are distributed across stages within the same hiring system. Automation may be stabilized in some stages while augmentation is institutionalized in others, allowing organizations to manage paradox spatially rather than temporally.\u003c/p\u003e \u003cp\u003eThis distributed configuration suggests that the automation\u0026ndash;augmentation paradox in recruitment is not a transient managerial dilemma but a structural feature of the selection process itself.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec5\" class=\"Section2\"\u003e \u003ch2\u003e2.3 Task Complementarity and Its Limits in Recruitment\u003c/h2\u003e \u003cp\u003eTask complementarity theory provides a task-level explanation for how humans and AI can jointly generate performance benefits (F\u0026uuml;gener, Walzner and Gupta, 2025). However, prior formulations of complementarity largely emphasize cognitive allocation efficiency between humans and algorithms (Jarrahi, \u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e2018\u003c/span\u003e). Such accounts understate how institutional responsibility structures shape the feasible boundaries of delegation in consequential organizational decisions.\u003c/p\u003e \u003cp\u003eThe theory distinguishes between between-task complementarity\u0026mdash;where humans and AI perform different tasks according to their comparative advantages\u0026mdash;and within-task complementarity\u0026mdash;where humans and AI jointly perform the same task. Automation typically creates value through between-task complementarity in highly analyzable tasks with high error tolerance. Augmentation, by contrast, relies on within-task complementarity, requiring humans to interpret, contextualize, and appropriately weight algorithmic outputs.\u003c/p\u003e \u003cp\u003eHowever, existing formulations of task complementarity largely treat complementarity as a cognitive or technical condition. In recruitment, complementarity is also institutionally constrained. Responsibility for selection decisions is normatively and legally assigned to human actors, even when algorithmic systems demonstrate strong predictive performance (Meijerink et al., \u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e2021\u003c/span\u003e). This allocation of responsibility limits the extent to which AI can assume autonomous decision authority, particularly in later recruitment stages where accountability exposure is greatest. Consequently, within-task complementarity in recruitment is bounded not only by human interpretive capacity but also by accountability requirements that restrict AI autonomy regardless of technical capability.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec6\" class=\"Section2\"\u003e \u003ch2\u003e2.4 Re-Augmentation and Legitimacy Repair\u003c/h2\u003e \u003cp\u003ePrior research often interprets retreats from automation as evidence of technological failure or managerial resistance. However, emerging evidence from AI-enabled decision contexts suggests that reversals may instead reflect legitimacy dynamics. Algorithmic systems are often penalized disproportionately after perceived errors, leading to declining trust and acceptance (Dietvorst, Simmons and Massey, \u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e2015\u003c/span\u003e), even though algorithmic judgment may be preferred in principle under some conditions (Logg, Minson and Moore, \u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e2019\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eExperimental research further demonstrates that perceptions of procedural justice and explainability strongly mediate acceptance of algorithmic decisions, especially when outcomes affect identity-relevant domains such as employment (Lee, \u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e2018\u003c/span\u003e). These findings suggest that legitimacy shocks in recruitment are socially constructed responses rather than purely performance-driven reactions.\u003c/p\u003e \u003cp\u003eIn recruitment, where decisions affect identity, dignity, and future opportunity, legitimacy threats are particularly salient. Algorithmic decision-making can undermine perceptions of human consideration and moral accountability, making organizations appear less humane (Newman, Fast and Harmon, \u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e2020\u003c/span\u003e). In response, organizations may deliberately reintroduce human oversight\u0026mdash;not to abandon AI, but to restore trust, defensibility, and institutional alignment.\u003c/p\u003e \u003cp\u003eThis study conceptualizes such responses as re-augmentation: the deliberate reassertion of human involvement following periods of increased automation. Re-augmentation functions as a legitimacy repair mechanism that stabilizes AI-enabled recruitment systems without fully relinquishing efficiency gains. Unlike human-in-the-loop designs, which assume stable human oversight from the outset, re-augmentation refers to a deliberate organizational reconfiguration following periods of expanded automation. It is triggered not by performance failure, but by legitimacy threats that render existing AI configurations socially or institutionally untenable.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec7\" class=\"Section2\"\u003e \u003ch2\u003e2.5 Theoretical Integration and Expectations\u003c/h2\u003e \u003cp\u003eSynthesizing these perspectives, this study conceptualizes recruitment as a stage-embedded automation\u0026ndash;augmentation system. Automation and augmentation are not alternative strategies to be chosen or sequenced at the organizational level. Rather, they are configured differently across recruitment stages in response to task analyzability, accountability exposure, and legitimacy pressures.\u003c/p\u003e \u003cp\u003eThis perspective yields three expectations that guide the subsequent analysis. First, automation will stabilize in early recruitment stages characterized by high analyzability and low accountability. Second, intermediate stages will exhibit unstable or hybrid configurations as organizations struggle to balance scalability, validity, and legitimacy. Third, late recruitment stages will institutionalize augmentation, constraining AI autonomy regardless of technical performance. These expectations provide the theoretical foundation for the comparative, case-based analysis that follows.\u003c/p\u003e \u003c/div\u003e"},{"header":"3. Methodology","content":"\u003cdiv id=\"Sec9\" class=\"Section2\"\u003e \u003ch2\u003e3.1 Research Design: Case-Based Theory Building\u003c/h2\u003e \u003cp\u003eThis study adopts a comparative, multi-case study design to theorize how artificial intelligence (AI) is configured as automation, augmentation, or hybrid human\u0026ndash;AI collaboration across recruitment stages. Case-based methodology is particularly well suited to this research because the phenomenon of interest\u0026mdash;the dynamic reconfiguration of AI roles\u0026mdash;is complex, processual, and insufficiently theorized in the HRM literature.\u003c/p\u003e \u003cp\u003eConsistent with established traditions of case-based theory building (Eisenhardt, 1989; Eisenhardt \u0026amp; Graebner, \u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e2007\u003c/span\u003e; Siggelkow, 2007), the objective of this study is not statistical generalization but analytic generalization. Rather than testing predefined hypotheses, I use cases to inductively surface patterns, mechanisms, and boundary conditions that refine existing theories of the automation\u0026ndash;augmentation paradox and task complementarity (Eisenhardt \u0026amp; Graebner, \u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e2007\u003c/span\u003e). Recruitment provides a theoretically fertile context for this endeavor because it is a multi-stage decision system in which task characteristics, accountability structures, and legitimacy requirements vary systematically across stages.\u003c/p\u003e \u003cp\u003eImportantly, this study conceptualizes cases not as illustrative examples of theory, but as theory-generating devices. By comparing how AI is deployed, adjusted, and sometimes partially withdrawn across recruitment stages and organizational contexts, I identify recurring configurations and transitions that enable theoretical extension beyond any single organization.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec10\" class=\"Section2\"\u003e \u003ch2\u003e3.2 Case Selection and Theoretical Sampling\u003c/h2\u003e \u003cp\u003eCases were selected using theoretical sampling, guided by the goal of maximizing conceptual insight rather than representativeness (Yin, \u003cspan citationid=\"CR42\" class=\"CitationRef\"\u003e2018\u003c/span\u003e). Following Eisenhardt\u0026rsquo;s (1989) logic, organizations were included if they met three criteria.\u003c/p\u003e \u003cp\u003eFirst, the organization had to demonstrate documented and sustained use of AI technologies in recruitment, rather than isolated pilots. This ensured that observed AI configurations reflected organizational routines rather than experimental anomalies.\u003c/p\u003e \u003cp\u003eSecond, sufficient publicly available and credible documentation had to exist to allow systematic reconstruction of AI deployment, adjustment, and outcomes across recruitment stages. Because AI-enabled recruitment is highly institutionalized and subject to public scrutiny, organizational disclosures, regulatory responses, and third-party investigations constitute an integral part of the phenomenon itself rather than a limitation of access.\u003c/p\u003e \u003cp\u003eThird, cases needed to exhibit variation in AI configurations and observable transitions over time, including instances of automation, augmentation, and hybrid reconfiguration. This criterion was critical for theorizing the automation\u0026ndash;augmentation paradox as a dynamic process rather than a static choice.\u003c/p\u003e \u003cp\u003eBased on these criteria, the final case set includes large multinational organizations such as Unilever, Amazon, Hilton, and IBM, as well as specialized HR technology providers whose systems are embedded within organizational recruitment processes. This combination enables theoretical replication, as similar stage-level patterns recur across different organizational and technological contexts. Such a multiple-case design typically yields more robust, generalizable, and testable theory than single-case research (Eisenhardt \u0026amp; Graebner, \u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e2007\u003c/span\u003e).\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec11\" class=\"Section2\"\u003e \u003ch2\u003e3.3 Data Sources and Case Materials\u003c/h2\u003e \u003cp\u003eThe study relies on multiple secondary data sources, which are both appropriate and necessary given the institutionalized and publicly contested nature of AI in recruitment. Data sources include peer-reviewed academic studies, practitioner research reports, corporate disclosures, regulatory filings, technology vendor documentation, and credible journalistic investigations.\u003c/p\u003e \u003cp\u003eUsing secondary data is not a compromise but a methodological strength in this context. High-profile recruitment AI systems are routinely subject to public debate, regulatory intervention, and reputational risk, making organizational responses and adjustments visible through external documentation. These materials capture not only technical design choices but also legitimacy concerns, governance structures, and organizational sensemaking surrounding AI use.\u003c/p\u003e \u003cp\u003eData triangulation across diverse sources enhances construct validity and reduces reliance on single narratives. Where discrepancies emerged across sources, these tensions were treated as analytically informative, often signaling contested interpretations or shifting organizational priorities.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec12\" class=\"Section2\"\u003e \u003ch2\u003e3.4 Analytical Framework: Stage-Based Mapping of AI Configurations\u003c/h2\u003e \u003cp\u003eTo ensure analytical consistency, all cases were mapped onto a standardized recruitment framework derived from established personnel selection literature. This framework conceptualizes recruitment as a sequence of interdependent stages, ranging from applicant pool formation to post-hire validation.\u003c/p\u003e \u003cp\u003eFor each recruitment stage in each case, I analyzed AI deployment along four analytical dimensions. First, I identified the primary role of AI, classifying it as automation, augmentation, or hybrid human\u0026ndash;AI collaboration. Second, I examined task characteristics, including analyzability, ambiguity, and error tolerance, drawing on the stage-based framework outlined earlier. Third, I analyzed organizational responses and outcomes, such as efficiency gains, changes in decision quality, emerging bias concerns, or legitimacy challenges. Fourth, I traced temporal dynamics, documenting whether and why organizations transitioned from automation to augmentation, from augmentation to automation, or toward hybrid reconfiguration. These transitions are central to theorizing the automation\u0026ndash;augmentation paradox as a cyclical, stage-embedded process.\u003c/p\u003e \u003cp\u003eCoding proceeded iteratively for transparency and trustworthy (Gioia, Corley and Hamilton, \u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e2013\u003c/span\u003e; Pratt, Kaplan and Whittington, \u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e2020\u003c/span\u003e). Initial open coding identified references to AI functionality, decision processes, organizational intent, and reported consequences. These codes were then consolidated into theory-informed categories drawing on paradox theory and task complementarity. Emerging interpretations were continuously compared across cases and stages to identify recurring patterns and deviations. This recursive process between data and emerging theory ensures that the resulting constructs and propositions are deeply grounded in empirical evidence (Eisenhardt \u0026amp; Graebner, \u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e2007\u003c/span\u003e).\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec13\" class=\"Section2\"\u003e \u003ch2\u003e3.5 Ensuring Rigor and Analytical Validity\u003c/h2\u003e \u003cp\u003eSeveral strategies were employed to enhance the rigor of the study. Construct validity was strengthened through triangulation across multiple data sources and by anchoring analysis in a well-established recruitment framework. Internal validity was addressed by explicitly linking observed patterns to theoretical mechanisms rather than relying on post hoc interpretation. Reliability was supported through transparent documentation of case selection criteria, analytical dimensions, and coding logic (Yin, \u003cspan citationid=\"CR42\" class=\"CitationRef\"\u003e2018\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eAlthough the study does not aim for statistical generalization, analytic generalization is achieved by demonstrating how recruitment-stage dynamics refine and extend broader theories of human\u0026ndash;AI collaboration. By making the analytical framework explicit, the study enables future research to apply a similar approach to other HR functions characterized by institutional sensitivity.\u003c/p\u003e \u003c/div\u003e"},{"header":"4. Findings","content":"\u003cp\u003eThe comparative analysis reveals that AI configurations in recruitment vary systematically across stages. These variations are not driven by differences in technological maturity alone, but by recurring organizational mechanisms through which efficiency demands, legitimacy pressures, and accountability requirements are reconciled. Three stage-embedded mechanisms emerge consistently across cases.\u003c/p\u003e \u003cp\u003eTable\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e provides a structured overview of representative organizational cases and illustrates how AI applications are configured as automation, augmentation, or hybrid arrangements across recruitment stages. Rather than serving as standalone case descriptions, the cases summarized in Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e function as empirical anchors that substantiate the stage-embedded mechanisms identified in the analysis. The findings reported below therefore abstract from firm-specific narratives and focus on recurring mechanisms evidenced across these cases.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eRepresentative Organizational Cases of AI Application Across Selection Stages\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"7\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSelection Stage\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eFirm Case\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eAI Application\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eApplication Type\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eQuantified Outcomes\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eKey Implications\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c7\"\u003e \u003cp\u003eLimitations\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eApplicant Pool Formation\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAtlassian\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eNLP-based job advertisement optimization (Textio) to reduce gendered language and broaden applicant pool\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eAugmentation\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e+\u0026thinsp;57% female technical hires within two years; significant increase in applicant diversity\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eAI can shape who applies by influencing language and signaling inclusiveness at the attraction stage\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eEffects limited to wording; does not address downstream screening or organizational bias\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMinimum Qualification Screening\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eUnilever\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eAutomated rule-based and ML screening for minimum criteria (education, work eligibility, availability)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eAutomation\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eRecruiter screening time reduced by ~\u0026thinsp;75%; processing over 250,000 applicants annually\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eAI automation is highly effective for objective, low-ambiguity eligibility thresholds\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eRisk of false negatives if minimum criteria are poorly specified or overly rigid\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eResume / CV Screening\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAmazon\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eML-based resume ranking trained on historical hiring data\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eAutomation\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eSystem discontinued after systematic gender bias detected in rankings\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eDemonstrates dangers of fully automated resume screening trained on biased historical data\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eLack of explainability and bias auditing undermined system legitimacy\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePreliminary Assessment\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eUnilever\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eAI-driven gamified cognitive and behavioral assessments (Pymetrics)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eAutomation\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eTime-to-hire reduced from ~\u0026thinsp;4 months to ~\u0026thinsp;4 weeks; ~70,000 recruiter hours saved annually\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eScalable early assessment can replace resumes and substantially reduce human workload\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eConstruct validity and cross-cultural fairness of gamified assessments remain debated\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eStructured Interview\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eHilton\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eAsynchronous AI video interviews (HireVue) for structured first-round interviews\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eAutomation\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eTime-to-hire reduced from 42 days to ~\u0026thinsp;5 days; interview completion rates increased\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eAI enables standardization and scalability in structured interview administration\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eUse of facial and voice analytics raises ethical, legal, and transparency concerns\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eWork Sample / Job Simulation\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eHackerRank\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eAI-assisted coding challenges with automated scoring and plagiarism detection\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eAutomation\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eScreening capacity scaled to tens of thousands of candidates; improved hiring manager satisfaction\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eAI-scored work samples offer high job relevance with scalable evaluation\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eMay disadvantage candidates unfamiliar with platform-based or time-pressured testing\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAssessment Center\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eUnilever\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eAI-supported virtual assessment centers with behavioral scoring support\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eAugmentation\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eAssessment center costs reduced by ~\u0026thinsp;50%; increased consistency across assessors\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eAI can support evaluators by reducing cognitive load and increasing standardization\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eFinal judgments still rely on human assessors; limited transparency in scoring models\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eReference and Background Checks\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eCheckr\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eAutomated AI-based background and compliance checks\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eAutomation\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eBackground check turnaround time reduced by ~\u0026thinsp;30\u0026ndash;50%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eAI excels in compliance-heavy, rule-based verification tasks\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eRisk of outdated, incomplete, or erroneous records affecting candidates\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eFinal Selection Decision \u0026amp; Job Offer\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eEightfold AI\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eAI-generated predictions of candidate fit, retention, and internal mobility\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eAugmentation\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eClient firms report improved quality-of-hire and internal fill rates (case evidence)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eAI is most appropriate as decision support\u0026mdash;not decision maker\u0026mdash;at final selection stage\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003ePredictive accuracy depends on data quality and organizational stability\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eValidation and Evaluation\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eIBM\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eContinuous AI auditing of hiring models for bias, validity, and performance\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eAugmentation\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eOngoing bias detection and model recalibration embedded in HR analytics\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eContinuous validation closes the AI hiring loop and sustains organizational legitimacy\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eRequires sustained data governance capability and analytical resources\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cdiv id=\"Sec15\" class=\"Section2\"\u003e \u003ch2\u003e4.1 Early Recruitment Stages: Efficiency-Driven Automation and Latent Risk\u003c/h2\u003e \u003cp\u003eMechanism: Efficiency-driven automation sustained by high analyzability and deferred accountability.\u003c/p\u003e \u003cp\u003eIn early recruitment stages\u0026mdash;such as applicant pool formation, minimum qualification screening, and initial r\u0026eacute;sum\u0026eacute; filtering\u0026mdash;organizations consistently configure AI as automation. Tasks at these stages are highly standardized, rule-based, and scalable, enabling AI systems to substitute for human effort with minimal contestation. This configuration generates substantial efficiency gains through between-task complementarity: AI assumes routine screening functions, allowing human recruiters to be redeployed to downstream evaluative tasks where human judgment is more critical.\u003c/p\u003e \u003cp\u003eImportantly, the legitimacy of automation at this stage is rarely challenged at the point of deployment. This is because errors are often perceived as reversible and decision stakes remain diffuse, leading to a state of deferred accountability. However, as illustrated by the Amazon case (Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e), automation at this stage is not inherently stable or free from conflict. The analysis reveals two critical theoretical nuances regarding this fragile stability:\u003c/p\u003e \u003cp\u003e \u003cul\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eLatent Legitimacy Risks\u003c/b\u003e: While automation appears stable due to low immediate accountability, it harbors latent risks. Automated systems trained on historical data may encode prior organizational biases\u0026mdash;such as the gender bias found in Amazon\u0026rsquo;s ML-based ranking tool. These biases often remain hidden until they are surfaced by external scrutiny, regulatory attention, or retrospective audits.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eConditional Stability based on Visibility\u003c/b\u003e: The perceived stability of early-stage automation is often a function of low error visibility rather than technical perfection. When an algorithmic failure becomes public, the perceived legitimacy collapses, triggering an immediate transition toward \"re-augmentation\" or the complete withdrawal of the system.\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e \u003c/p\u003e \u003cp\u003eConsequently, these retreats do not necessarily reflect technical breakdowns but rather a sudden shift in legitimacy evaluation. Automation in early recruitment stages must be reconceptualized not as a permanently stable state, but as a configuration of \"fragile stability\" that remains viable only as long as efficiency gains do not publicly collide with institutional fairness expectations.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec16\" class=\"Section2\"\u003e \u003ch2\u003e4.2 Intermediate Recruitment Stages: Paradox Intensification and Hybrid Instability\u003c/h2\u003e \u003cp\u003eMechanism: Paradox intensification arising from simultaneous demands for scalability, validity, and legitimacy.\u003c/p\u003e \u003cp\u003eIntermediate recruitment stages\u0026mdash;such as preliminary assessments, structured interviews, work samples, and assessment centers\u0026mdash;consistently emerge as zones of instability. Tasks at these stages promise high predictive value while also carrying heightened ethical, legal, and experiential scrutiny.\u003c/p\u003e \u003cp\u003eOrganizations frequently attempt to automate these stages to achieve scale and standardization. However, the analysis shows that such efforts quickly encounter legitimacy challenges related to construct validity, bias, transparency, and candidate experience. As a result, organizations rarely sustain fully automated configurations at intermediate stages.\u003c/p\u003e \u003cp\u003eInstead, hybrid configurations proliferate. AI systems generate standardized scores or rankings, while humans retain responsibility for interpretation and integration. Yet these hybrids remain unstable. Humans struggle to consistently determine when algorithmic outputs should be trusted or overridden, particularly in socially complex evaluative contexts. This difficulty undermines within-task complementarity and triggers repeated recalibration.\u003c/p\u003e \u003cp\u003eAcross cases, intermediate stages exhibit oscillation rather than convergence. Organizations alternate between automation and augmentation, adjusting AI autonomy in response to emerging concerns without settling into a stable equilibrium. Paradox intensification thus reflects not managerial indecision, but the structural incompatibility of competing demands at this stage.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec17\" class=\"Section2\"\u003e \u003ch2\u003e4.3 Late Recruitment Stages: Accountability-Driven Augmentation\u003c/h2\u003e \u003cp\u003eMechanism: Institutionalized augmentation anchored by accountability and responsibility attribution.\u003c/p\u003e \u003cp\u003eIn late recruitment stages\u0026mdash;such as final selection decisions, job offers, and validation\u0026mdash;AI is consistently configured as augmentation rather than automation. Even when AI systems demonstrate strong predictive capability, decision authority remains explicitly human.\u003c/p\u003e \u003cp\u003eThe analysis shows that this pattern is driven by accountability rather than performance considerations. Late-stage decisions are highly visible, contestable, and consequential, making organizations acutely sensitive to legal defensibility and moral responsibility. Delegating final authority to AI is widely perceived as illegitimate, regardless of technical adequacy.\u003c/p\u003e \u003cp\u003eThis institutional stabilization occurs even when AI predictive accuracy is reported to exceed human judgment, as seen in cases like Eightfold AI. The persistence of human authority in these settings is driven by the fundamental requirement for moral responsibility and legal defensibility, ensuring that high-stakes outcomes remain attributable to human actors rather than opaque algorithmic processes.\u003c/p\u003e \u003cp\u003eAs a result, AI systems at this stage are embedded as decision-support tools that inform but do not determine outcomes. Augmentation at late stages is notably more stable than hybrid configurations observed earlier, reflecting the clarity of institutional expectations regarding human accountability.\u003c/p\u003e \u003cp\u003eThis finding demonstrates that limits on AI autonomy in recruitment are not merely technical or cognitive, but normative and institutional.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec18\" class=\"Section2\"\u003e \u003ch2\u003e4.4 Re-Augmentation as a Cross-Stage Legitimacy Repair Mechanism\u003c/h2\u003e \u003cp\u003eMechanism: Re-augmentation as organizational response to legitimacy threats.\u003c/p\u003e \u003cp\u003eAcross all recruitment stages, a recurring dynamic emerges: organizations repeatedly reintroduce human oversight following periods of increased automation. This pattern\u0026mdash;observed in early and intermediate stages in particular\u0026mdash;is triggered less by declining performance than by perceived legitimacy threats.\u003c/p\u003e \u003cp\u003eInstances of re-augmentation occur when automated systems generate outcomes that are difficult to justify, explain, or defend. In response, organizations recalibrate AI roles by narrowing autonomy, increasing human review, or reframing AI outputs as advisory rather than determinative.\u003c/p\u003e \u003cp\u003eCrucially, re-augmentation does not entail abandoning AI. Instead, it functions as a legitimacy repair mechanism that restores trust and defensibility while preserving selected efficiency gains. Through re-augmentation, organizations stabilize AI-enabled recruitment systems by realigning technological capability with institutional expectations.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec19\" class=\"Section2\"\u003e \u003ch2\u003e4.5 Summary of Stage-Embedded Mechanisms\u003c/h2\u003e \u003cp\u003eTaken together, these findings demonstrate that AI configurations in recruitment are governed by stage-specific mechanisms rather than uniform organizational strategies. Specifically, early stages stabilize automation under conditions of high task analyzability and low immediate accountability; intermediate stages intensify paradoxical tensions that resist stabilization; and late stages institutionalize augmentation due to stringent accountability constraints. Within this framework, re-augmentation serves as a cross-stage corrective mechanism triggered whenever the legitimacy of a configuration is threatened.\u003c/p\u003e \u003cp\u003eFigure \u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e illustrates the dynamic interaction between automation and augmentation throughout the AI-enabled recruitment lifecycle. The mechanisms occurring during the early stage (Screening) warrant particular attention, as they reveal both the general tendency toward automation and the potential for exceptional deviations.\u003c/p\u003e \u003cp\u003eTypically, the initial recruitment stage is characterized by high task analyzability, favoring the establishment of stable, efficiency-driven automation models. However, as evidenced by the case of Amazon\u0026rsquo;s AI recruitment tool, this trajectory can shift abruptly upon the detection of \"algorithmic bias\"\u0026mdash;such as the generation of gender-biased results. Such technical flaws not only erode trust in the AI system but also compel the organization to confront an acute demand for institutional accountability. In such instances, the established automation path fractures, triggering a 're-augmentation loop.'\u003c/p\u003e \u003cp\u003eThe Amazon case underscores that even in a supposedly stabilized early stage, a re-augmentation loop can be activated instantaneously. To restore legitimacy, organizations must make strategic choices to reinforce 'human accountability' (represented by the blue line in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e). This involves forcibly reducing 'algorithmic autonomy' (the red line in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e) and mandating intervention by human experts to correct biases or, in extreme cases, decommission the system entirely.\u003c/p\u003e \u003cp\u003eUltimately, these mechanisms suggest that re-augmentation loops are not confined to the intermediate stages of assessment; rather, they can emerge at any point in the recruitment journey where the legitimacy of automation is compromised. This serves as a powerful restorative force that reverts the system back toward an augmentation model. Consequently, the configuration of AI-enabled recruitment should not be viewed as a static end-state, but as a perpetually reconstructed and dynamic process. These stage-embedded mechanisms explain why organizations repeatedly transition between automation and augmentation without ever converging on a single, permanent configuration.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e"},{"header":"5. Discussion","content":"\u003cdiv id=\"Sec21\" class=\"Section2\"\u003e \u003ch2\u003e5.1 Reconceptualizing the Automation\u0026ndash;Augmentation Paradox as Stage-Embedded\u003c/h2\u003e \u003cp\u003eThis study set out to examine how automation and augmentation are configured across recruitment stages and why organizations repeatedly transition between these configurations over time. The findings demonstrate that the automation\u0026ndash;augmentation paradox in recruitment is neither a temporal sequence nor an organizational-level choice. Instead, it is structurally embedded across recruitment stages.\u003c/p\u003e \u003cp\u003ePrior research has predominantly conceptualized the automation\u0026ndash;augmentation paradox as unfolding over time at the firm level, suggesting cycles of automation followed by augmentation or vice versa. While this temporal perspective captures important dynamics, it obscures how contradictory logics are simultaneously enacted within organizational processes. By shifting the unit of analysis from organizations to recruitment stages, this study shows that automation and augmentation coexist within a single hiring system, configured differently across stages according to task characteristics and institutional exposure.\u003c/p\u003e \u003cp\u003eIn doing so, the study aligns with broader organizational research showing that algorithmic governance reshapes control and coordination structures rather than simply substituting human labor (Kellogg et al., \u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e2020\u003c/span\u003e). Recruitment stages therefore function as micro-sites of algorithmic governance where authority is continuously negotiated.\u003c/p\u003e \u003cp\u003eThis stage-embedded view extends paradox theory by demonstrating that paradoxical tensions need not be resolved or balanced temporally. Instead, organizations manage paradox spatially by distributing automation and augmentation across interdependent stages. Recruitment thus becomes a site where paradox is continuously enacted rather than episodically addressed.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec22\" class=\"Section2\"\u003e \u003ch2\u003e5.2 Re-Augmentation as a Legitimacy Repair Mechanism\u003c/h2\u003e \u003cp\u003eA second contribution concerns how this study interprets organizational retreats from automation. Existing research often treats reversals from automation as evidence of technological failure, implementation error, or resistance to change. The findings challenge this interpretation.\u003c/p\u003e \u003cp\u003eAcross cases, organizations repeatedly reintroduced human oversight after periods of increased automation, particularly in response to concerns about bias, transparency, or accountability. Importantly, these reversals did not entail abandoning AI altogether. Instead, organizations recalibrated AI\u0026rsquo;s role to restore legitimacy while retaining selected efficiency gains.\u003c/p\u003e \u003cp\u003eThis pattern is conceptualized here as re-augmentation. Re-augmentation functions as a legitimacy repair mechanism through which organizations reassert human responsibility and moral accountability when algorithmic systems threaten institutional expectations. Rather than representing regression, re-augmentation stabilizes AI-enabled recruitment by aligning technological capability with social and legal norms.\u003c/p\u003e \u003cp\u003eBy foregrounding legitimacy as a central driver of AI reconfiguration, this study extends paradox theory beyond efficiency\u0026ndash;learning trade-offs and highlights the institutional work organizations perform to sustain AI use in high-stakes HR contexts.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec23\" class=\"Section2\"\u003e \u003ch2\u003e5.3 Refining Task Complementarity Theory: Accountability-Constrained Complementarity\u003c/h2\u003e \u003cp\u003eThis study also refines task complementarity theory by demonstrating that complementarity in recruitment is institutionally bounded.\u003c/p\u003e \u003cp\u003eConsistent with task complementarity theory, the findings show that automation generates value through between-task complementarity in early recruitment stages characterized by high analyzability and error tolerance. However, within-task complementarity\u0026mdash;central to augmentation\u0026mdash;proves far more fragile.\u003c/p\u003e \u003cp\u003eAt intermediate stages, hybrid configurations emerge, but organizations struggle to stabilize them. Humans face persistent difficulty in consistently determining when algorithmic recommendations should be trusted or overridden in socially complex evaluative tasks. As a result, organizations oscillate between automation and augmentation rather than converging on a stable hybrid equilibrium.\u003c/p\u003e \u003cp\u003eMost critically, the findings demonstrate that within-task complementarity is structurally constrained at late recruitment stages. Even when AI systems demonstrate strong predictive performance, decision authority remains explicitly human. This constraint is not primarily cognitive or technical, but institutional: responsibility for hiring decisions is legally and normatively assigned to human actors.\u003c/p\u003e \u003cp\u003eAccountability therefore limits AI autonomy regardless of performance. Recent HRM research similarly argues that reducing algorithmic bias requires sustained human accountability, cross-stage governance integration, and continuous oversight, reinforcing the structural limits of fully autonomous AI in recruitment (Soleimani et al., \u003cspan citationid=\"CR39\" class=\"CitationRef\"\u003e2025\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eThis finding resonates with sociotechnical critiques of 'abstraction traps'\u0026mdash;the notion that attempting to isolate technical performance from institutional contexts obscures the social allocation of responsibility (Selbst et al., 2019). In high-stakes settings like recruitment, algorithmic autonomy is constrained by the threshold of socially acceptable responsibility transfer. While regulatory frameworks such as GDPR Article 22 or New York City Local Law 144 impose transparency and audit requirements, their primary influence on AI configuration is exercised through the mandate for human oversight (Roig, \u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e2017\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eBy introducing accountability as a boundary condition, this study extends task complementarity theory beyond cognitive interaction to encompass institutional responsibility as a central determinant of human\u0026ndash;AI collaboration.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec24\" class=\"Section2\"\u003e \u003ch2\u003e5.4 Propositions for Case-Based Theory Building\u003c/h2\u003e \u003cp\u003eSynthesizing the findings, this study advances four propositions that extend existing research on AI, HRM, and organizational decision-making:\u003c/p\u003e \u003cp\u003e \u003cul\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eProposition 1 (Stage-Embedded Configuration)\u003c/b\u003e: Automation and augmentation in recruitment are configured at the stage level rather than uniformly at the organizational level, resulting in the simultaneous coexistence of multiple human\u0026ndash;AI configurations within a single hiring system.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eProposition 2 (Legitimacy-Bounded Automation)\u003c/b\u003e: Automation in recruitment stabilizes only in stages where task analyzability and legitimacy requirements remain aligned; misalignment triggers reconfiguration toward augmentation.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eProposition 3 (Re-Augmentation as Legitimacy Repair)\u003c/b\u003e: Re-augmentation emerges as a legitimacy repair mechanism through which organizations reassert human oversight in response to ethical, legal, or reputational challenges arising from automated recruitment systems.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eProposition 4 (Accountability-Constrained Complementarity)\u003c/b\u003e: Within-task human\u0026ndash;AI complementarity is structurally constrained in high-accountability recruitment decisions, limiting AI autonomy regardless of technical performance.\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e \u003c/p\u003e \u003cp\u003eThese propositions integrate paradox theory and task complementarity within a stage-embedded framework, offering a process-sensitive account of AI governance in institutionally exposed HR systems. offering a more nuanced account of human\u0026ndash;AI collaboration in institutionally sensitive organizational processes.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec25\" class=\"Section2\"\u003e \u003ch2\u003e5.5 Implications\u003c/h2\u003e \u003cp\u003eThe findings of this study offer critical insights for human resource management (HRM), human resource development (HRD), and the formulation of policies for AI governance. From an academic and managerial perspective, the results underscore the necessity of a stage-sensitive approach that views recruitment not as a single event, but as an adaptive system where human\u0026ndash;AI configurations evolve based on specific task characteristics and temporal dynamics. AI-enabled recruitment should be repositioned as a system for legitimacy management and institutional learning rather than a mere deployment of efficiency-driven technology.\u003c/p\u003e \u003cp\u003eConsequently, the benefits of augmentation depend heavily on HRD investments\u0026mdash;specifically in recruiter training, the development of interpretive skills, ethical literacy, and the establishment of governance routines that facilitate meaningful human engagement with AI systems. Furthermore, by integrating paradox theory and task complementarity within the HRM context, this study demonstrates how fine-grained analysis of HR processes can refine macro-level theories concerning AI and the future of work.\u003c/p\u003e \u003cp\u003eBeyond organizational practice, this stage-embedded theory provides concrete guidance for policymakers and regulatory bodies overseeing AI in the labor market. Rather than applying a monolithic regulatory framework to all AI applications, oversight should be differentiated based on the institutional sensitivity of each stage. In late-stage decisions where legal and normative accountability is paramount, regulations must ensure that \"meaningful human control\" is substantively exercised, regardless of the AI's predictive accuracy.\u003c/p\u003e \u003cp\u003eFor intermediate stages, which are characterized by high instability between efficiency and fairness, regulators should prioritize monitoring whether human evaluators possess the actual authority and capability to critically review and override algorithmic outputs. In early recruitment stages, where automation is prevalent, policy should focus on institutionalizing periodic external audits and data lineage verification to detect latent biases before they manifest as systemic exclusion.\u003c/p\u003e \u003cp\u003eMoreover, technical interventions must move beyond simple data collection toward methods that mitigate discrimination without relying on sensitive attributes, thereby aligning technical performance with evolving human rights standards. Such 'fairness-by-design' approaches are essential for sustaining the long-term legitimacy of AI-enabled recruitment systems.\u003c/p\u003e \u003cp\u003eFinally, \"re-augmentation\" should be recognized by regulators as a formal safety valve and a legitimate mechanism for restoring institutional trust, encouraging firms to adopt proactive protocols for reintroducing human oversight when algorithmic outcomes threaten social or ethical norms.\u003c/p\u003e \u003c/div\u003e"},{"header":"6. Conclusion","content":"\u003cp\u003eThis study argues that the central challenge of AI in recruitment is not choosing between automation and augmentation, nor progressing linearly toward greater algorithmic autonomy. Instead, the challenge lies in governing their configuration across recruitment stages and over time.\u003c/p\u003e \u003cp\u003eBy conceptualizing recruitment as a stage-embedded automation\u0026ndash;augmentation system, this study advances theory in three ways. First, it reconceptualizes the automation\u0026ndash;augmentation paradox as spatially embedded within organizational processes rather than temporally resolved at the firm level. Second, it refines task complementarity theory by demonstrating that accountability, not just task analyzability or cognitive interaction, constrains human\u0026ndash;AI collaboration in high-stakes HR decisions. Third, it introduces re-augmentation as a theoretically meaningful legitimacy mechanism through which organizations stabilize AI-enabled recruitment systems.\u003c/p\u003e \u003cp\u003eFrom a practical perspective, the findings caution against one-size-fits-all AI strategies in recruitment. Organizations that pursue aggressive automation across all stages risk triggering cycles of legitimacy loss and trust erosion. Conversely, organizations that rely exclusively on augmentation may forgo scalability and efficiency gains. Effective recruitment AI strategies therefore require deliberate differentiation across stages and continuous recalibration in response to institutional pressures.\u003c/p\u003e \u003cp\u003eBeyond recruitment, the stage-embedded configuration identified here may illuminate AI governance challenges in other high-stakes domains such as healthcare triage, credit scoring, and judicial decision-making, where accountability cannot be meaningfully automated. By theorizing accountability as a structural ceiling on delegation, this study contributes to ongoing debates on responsible AI and institutional trust.\u003c/p\u003e \u003cp\u003eThis study has limitations, including its reliance on secondary data and publicly documented cases. Future research could extend this work through longitudinal field studies, cross-national comparisons, or experimental designs examining how candidates and decision-makers perceive stage-specific AI use (Pettigrew, \u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e1990\u003c/span\u003e). As AI continues to reshape HRM, understanding how automation and augmentation are governed\u0026mdash;not merely adopted\u0026mdash;will be essential for building legitimate and sustainable hiring systems.\u003c/p\u003e"},{"header":"Declarations","content":" \u003cp\u003e \u003cb\u003eEthics Statement\u003c/b\u003e \u003c/p\u003e \u003cp\u003e Ethical approval\u003c/strong\u003e was not required as the study did not involve human participants.\u003c/p\u003e\u003ch2\u003eInformed consent\u003c/h2\u003e \u003cp\u003eThis article does not contain any studies with human participants performed by any of the authors.\u003c/p\u003e \u003ch2\u003eConflict of interest\u003c/h2\u003e \u003cp\u003eThe author declare that there is no conflict of interest in this paper.\u003c/p\u003e \u003ch2\u003eAuthor Contribution\u003c/h2\u003e\u003cp\u003eH.L.: Conceptualization; Formal Analysis; Data Curation; Writing \u0026ndash; Original Draft; Methodology; Supervision; Validation; Writing \u0026ndash; Review \u0026amp; Editing.\u003c/p\u003e\u003ch2\u003eData Availability\u003c/h2\u003e\u003cp\u003eData used in this study consist only of publicly available secondary data cited in the references. No primary or proprietary datasets were used, and all sources are accessible via the references.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eArthur W Jr, Day EA, McNelly TL, Edens PS (2003) A meta-analysis of the criterion‐related validity of assessment center dimensions. Pers Psychol 56:125\u0026ndash;153\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBaron J (2018) New continuous background checks will take a deeper look at your personal life. Forbes (10 December\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBreaugh JA (2013) Employee recruitment. Ann Rev Psychol 64:389\u0026ndash;416\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCampion MA, Palmer DK, Campion JE (1997) A review of structure in the selection interview. Pers Psychol 50:655\u0026ndash;702\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCole MS, Feild HS, Giles WF (2004) Interaction of recruiter and applicant gender in resume evaluation: a field study. Sex roles 51:597\u0026ndash;608\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDastin J (2022) Ethics of data and analytics 296\u0026ndash;299. Auerbach\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDietvorst BJ, Simmons JP, Massey C (2015) Algorithm aversion: people erroneously avoid algorithms after seeing them err. J Exp Psychol Gen 144:114\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDineen BR, Yu KYT, Stevenson-Street J (2023) Recruitment in personnel psychology and beyond: Where we've been working, and where I might work next. Pers Psychol 76:617\u0026ndash;650\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eEisenhardt KM, Graebner ME (2007) Theory building from cases: Opportunities and challenges. Acad Manag J 50:25\u0026ndash;32\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFaraj S, Pachidi S, Sayegh K (2018) Working and organizing in the age of the learning algorithm. Inf Organ 28:62\u0026ndash;70\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eF\u0026uuml;gener A, Walzner DD, Gupta A (2026) Roles of artificial intelligence in collaboration with humans: Automation, augmentation, and the future of work. Manage Sci 72:538\u0026ndash;557\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGatewood RD, Feild HS, Barrick M (2020) Human Resource Selection, 9th edn. Cengage Learning\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGioia DA, Corley KG, Hamilton AL (2013) Seeking qualitative rigor in inductive research: Notes on the Gioia methodology. Organizational Res methods 16:15\u0026ndash;31\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGlikson E, Woolley AW (2020) Human trust in artificial intelligence: Review of empirical research. Acad Manag Ann 14:627\u0026ndash;660\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHackerRank (2023) How HackerRank catches AI-generated code by deploying advanced ML plagiarism detection. HackerRank Blog (7 June\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHunkenschroer AL, Luetge C (2022) Ethics of AI-enabled recruiting and selection: A review and research agenda. J Bus Ethics 178:977\u0026ndash;1007\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eIBM AI in recruitment. IBM Think \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://www.ibm.com/think/topics/ai-in-recruitment\u003c/span\u003e\u003cspan address=\"https://www.ibm.com/think/topics/ai-in-recruitment\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJarrahi MH (2018) Artificial intelligence and the future of work: Human-AI symbiosis in organizational decision making. Bus Horiz 61:577\u0026ndash;586\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKellogg KC, Valentine MA, Christin A (2020) Algorithms at work: The new contested terrain of control. Acad Manag Ann 14:366\u0026ndash;410\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKurter H How Hilton reduced their time to hire from 43 days down to 5. Forbes (19 September 2019)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLee MK (2018) Understanding perception of algorithmic decisions: Fairness, trust, and emotion in response to algorithmic management. Big data Soc 5:2053951718756684\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLogg JM, Minson JA, Moore DA (2019) Algorithm appreciation: People prefer algorithmic to human judgment. Organ Behav Hum Decis Process 151:90\u0026ndash;103\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMarino DL et al in 2020 13th International Conference on Human System Interaction (HSI). 155\u0026ndash;161 (IEEE)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMarr B (2019) Artificial intelligence in practice: how 50 successful companies used AI and machine learning to solve problems. Wiley\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMeijerink J, Boons M, Keegan A, Marler J (2021) Algorithmic human resource management: Synthesizing developments and cross-disciplinary insights on digital HRM. Int J Hum Resour Manag 32:2605\u0026ndash;2621\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNewman DT, Fast NJ, Harmon DJ (2020) When eliminating bias isn\u0026rsquo;t fair: Algorithmic reductionism and procedural justice in human resource decisions. Organ Behav Hum Decis Process 160:149\u0026ndash;167\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePettigrew AM (1990) Longitudinal field research on change: Theory and practice. Organ Sci 1:267\u0026ndash;292\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePratt MG, Kaplan S, Whittington R (2020) Editorial essay: The tumult over transparency: Decoupling transparency from replication in establishing trustworthy qualitative research. Adm Sci Q 65:1\u0026ndash;19\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePumpedu The practical application of AI: Unilever reduced recruitment time by 75%. Pumpedu News \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://www.pumpedu.cz/en/news/the-practical-application-of-ai-unilever-reduced-recruitment-time-by-75-112n\u003c/span\u003e\u003cspan address=\"https://www.pumpedu.cz/en/news/the-practical-application-of-ai-unilever-reduced-recruitment-time-by-75-112n\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRaisch S, Krakowski S (2021) Artificial intelligence and management: The automation\u0026ndash;augmentation paradox. Acad Manage Rev 46:192\u0026ndash;210\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eReuters (2026) AI company Eightfold sued for helping companies secretly score job seekers. Reuters (21 January \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://www.reuters.com/sustainability/boards-policy-regulation/ai-company-eightfold-sued-helping-companies-secretly-score-job-seekers-2026-01-21/\u003c/span\u003e\u003cspan address=\"https://www.reuters.com/sustainability/boards-policy-regulation/ai-company-eightfold-sued-helping-companies-secretly-score-job-seekers-2026-01-21/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRoig A (2017) Safeguards for the right not to be subject to a decision based solely on automated processing (Article 22 GDPR). Eur J Law Technol 8\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRoth PL, Bobko P, McFarland LA (2005) A meta-analysis of work sample test validity: Updating and integrating some classic literature. Pers Psychol 58:1009\u0026ndash;1037\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRynes SL, Bretz RD Jr, Gerhart B (1991) The importance of recruitment in job choice: A different way of looking. Pers Psychol 44:487\u0026ndash;521\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSackett PR, Zhang C, Berry CM, Lievens F (2022) Revisiting meta-analytic estimates of validity in personnel selection: Addressing systematic overcorrection for restriction of range. J Appl Psychol 107:2040\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSchmidt FL, Hunter JE (1998) The validity and utility of selection methods in personnel psychology: Practical and theoretical implications of 85 years of research findings. Psychol Bull 124:262\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSelbst AD, Boyd D, Friedler SA, Venkatasubramanian S, Vertesi J in Proceedings of the conference on fairness, accountability, and transparency. 59\u0026ndash;68\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSmith WK, Lewis MW (2011) Toward a theory of paradox: A dynamic equilibrium model of organizing. Acad Manage Rev 36:381\u0026ndash;403\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSoleimani M, Intezari A, Arrowsmith J, Pauleen DJ, Taskin N (2025) Reducing AI bias in recruitment and selection: an integrative grounded approach. Int J Hum Resource Manage 36:2480\u0026ndash;2515\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTambe P, Cappelli P, Yakubovich V (2019) Artificial intelligence in human resources management: Challenges and a path forward. Calif Manag Rev 61:15\u0026ndash;42\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eVrontis D et al (2023) Artificial intelligence, robotics, advanced technologies and human resource management: a systematic review. Artif Intell Int HRM, 172\u0026ndash;201\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYin RK (2018) Case Study Research and Applications: Design and Methods, 6th edn. SAGE\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":false,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"humanities-and-social-sciences-communications","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"palcomms","sideBox":"Learn more about [Humanities \u0026 Social Sciences Communications](http://www.nature.com/palcomms/)","snPcode":"41599","submissionUrl":"https://submission.springernature.com/new-submission/41599/3","title":"Humanities and Social Sciences Communications","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Nature AJ","inReviewEnabled":true,"inReviewRevisionsEnabled":false},"keywords":"artificial intelligence, talent selection, automation, augmentation, case-based theory building","lastPublishedDoi":"10.21203/rs.3.rs-9222046/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-9222046/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eArtificial intelligence (AI) is increasingly embedded in talent selection, yet its deployment remains characterized by a persistent tension between automation and augmentation. While existing research often treats this choice as an organizational-level strategy or a temporal progression, it frequently overlooks how institutional sensitivities and task characteristics vary across the recruitment lifecycle. This research examines how AI is configured as automation, augmentation, or hybrid collaboration across specific recruitment stages and why organizations repeatedly transition between these configurations. Adopting a comparative, multi-case theory-building approach, the paper analyzes documented AI deployments across standardized selection stages, including applicant screening, assessment, and final selection. The findings identify three recurring stage-embedded mechanisms. Early recruitment stages stabilize efficiency-driven automation under conditions of high task analyzability and low immediate accountability. Intermediate stages intensify paradoxical tensions, producing unstable hybrid configurations as organizations attempt to balance scalability, validity, and fairness. Late stages institutionalize accountability-driven augmentation, structurally constraining AI autonomy due to normative and legal requirements for human responsibility, regardless of technical performance. Across these stages, organizations engage in \"re-augmentation\" as a legitimacy repair mechanism to restore trust and institutional alignment following perceived algorithmic failures or ethical concerns. By reconceptualizing the automation\u0026ndash;augmentation paradox as a stage-embedded phenomenon, this work demonstrates that contradictory logics are distributed spatially across the hiring process. These insights refine task complementarity theory by establishing accountability as a fundamental boundary condition that limits human\u0026ndash;AI collaboration in high-stakes human resource decisions.\u003c/p\u003e","manuscriptTitle":"Reconceptualizing the Automation–Augmentation Tension in AI-Enabled Talent Selection: A Stage-Embedded Theory","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-04-01 07:29:03","doi":"10.21203/rs.3.rs-9222046/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Revision requested","date":"2026-04-04T17:16:20+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2026-04-04T17:06:16+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2026-04-01T07:09:28+00:00","index":"","fulltext":""},{"type":"submitted","content":"Humanities and Social Sciences Communications","date":"2026-03-25T10:49:07+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"humanities-and-social-sciences-communications","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"palcomms","sideBox":"Learn more about [Humanities \u0026 Social Sciences Communications](http://www.nature.com/palcomms/)","snPcode":"41599","submissionUrl":"https://submission.springernature.com/new-submission/41599/3","title":"Humanities and Social Sciences Communications","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Nature AJ","inReviewEnabled":true,"inReviewRevisionsEnabled":false}}],"origin":"","ownerIdentity":"4aba5297-b324-4acd-91bd-7df95954975a","owner":[],"postedDate":"April 1st, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"under-review","subjectAreas":[{"id":65110925,"name":"Humanities/Complex networks"},{"id":65110927,"name":"Social science/Complex networks"},{"id":65110929,"name":"Physical sciences/Mathematics and computing"},{"id":65110931,"name":"Social science/Science technology and society"}],"tags":[],"updatedAt":"2026-04-12T08:53:14+00:00","versionOfRecord":[],"versionCreatedAt":"2026-04-01 07:29:03","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-9222046","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-9222046","identity":"rs-9222046","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}