Algorithmic Recruitment Efficiency, Ethical Paradoxes, and Governance Frameworks: A Cross-Industry Empirical Study of AI-Driven Hiring Practices Under the EU AI Act.

Figure 1: Empirical Longitudinal Trend of Core Performance Indicators in Artificial Intelligence-based Recruitment Benefits and Ethical Concerns (2010–2016)

The tripartite tension embedded in the title—efficiency, ethics, and governance—necessitates a multi-theoretic lens that transcends a singular paradigmatic allegiance. Primarily, we anchor our analysis in Agency Theory, articulated by Jensen and Meckling (1976), reconstituted for the algorithmic milieu. Here, the Principal (the hiring firm) delegates screening to an Agent (the algorithmic system), yet information asymmetry is inverted: the AI’s opaque decision calculus introduces a novel form of moral hazard where the agent’s utility function—optimizing for historical pattern replication—may diverge from the principal’s strategic goal of workforce diversification. This divergence is amplified under the extraterritorial reach of the EU AI Act, which imposes a compliance burden on Indian subsidiaries of EU-based multinationals, thereby activating a distinct agency cost. Complementing this, the Resource-Based View (Barney, 1991) posits that algorithmic recruitment constitutes a VRIO resource only when tacitly embedded within idiosyncratic organizational routines; mere technological appropriation without processual integration yields no sustained competitive advantage. The ethical paradox emerges through the lens of Institutional Theory (DiMaggio & Powell, 1983), wherein Indian firms facing coercive isomorphic pressure from global data protection standards (DPDP Act, 2023) and mimetic pressure from industry peers adopt governance frameworks as ceremonial rituals of legitimacy rather than substantive ethical recalibration. Signalling Theory (Spence, 1973) further illuminates how algorithmic certifications function as costly signals to a discerning applicant pool, particularly within India’s hyper-competitive IT sector where adverse selection dynamics necessitate credible commitments to procedural fairness. The 2025 Indian context, characterised by a demographic dividend coupled with escalating gig-economy precarity, renders these theoretical mechanisms particularly salient, as the EU AI Act serves as an exogenous institutional shock compelling Indian firms to reconcile their indigenous hiring pragmatism with global normative expectations.

Prior empirical scholarship has bifurcated into two largely incommensurable streams. The first, predominantly situated in Western labour markets, celebrates algorithmic recruitment’s capacity to attenuate cognitive biases, drawing upon large-scale audit studies by Raghavan et al. (2020) and Cowgill (2022) which demonstrated substantial reductions in time-to-hire metrics. Conversely, a second corpus—heavily influenced by critical algorithm studies originating from O’Neil (2016) and Eubanks (2018)—has documented the pernicious reproduction of structural inequalities, particularly when training data encode historical discriminatory patterns. However, the emerging market literature, especially concerning India, presents a more ambiguous landscape. Studies by Sharma and Gupta (2023) in the Journal of Indian Business Research found that AI recruitment tools in Indian manufacturing inadvertently penalised candidates from non-metropolitan educational institutions, yet simultaneously reported that these tools exhibited no caste-based disparities—a finding incongruent with Western-centric predictions. Conflict arises between studies that frame technological adoption as developmental modernity (Rana, 2021) versus those portraying it as a novel vector of digital stratification (Kumar & Chandra, 2024). Critically, extant literature undertheorises the regulatory mediation role of the EU AI Act’s Article 22 (automated decision-making safeguards) upon indigenous Indian hiring practices, particularly its effect on cross-industry heterogeneity. Furthermore, most studies rely on cross-sectional data, failing to capture the dynamic adjustment paths as firms iteratively refine their algorithms in response to audit failures. Our paper addresses this gap by deploying dynamic panel econometrics across IT, manufacturing, and services sectors from 2019 to 2025, precisely tracing how governance frameworks moderate the efficiency-ethics trade-off and whether compliance with a foreign regulatory regime induces measurable behavioural shifts in domestic hiring outcomes—a dimension conspicuously absent from the current discourse.

Recruitment is one of the most critical human resource functions, directly influencing organizational success by determining the quality of human capital as observed by Barongo & Mbelwa (2024). Traditionally, recruitment processes have been labor-intensive, involving resume screening, aptitude testing, interviews, and reference checks. With the explosion of job applications in the digital era, especially through online portals and social media, manual methods have proven insufficient.

In India, start-ups and large corporates alike have increasingly adopted AI recruitment tools between 2018 and 2025, driven by the need to hire efficiently in a competitive talent market. Yet, ethical concerns such as algorithmic bias, privacy violations, and transparency gaps remain unresolved, sparking debate on the responsible use of AI in HR.

This investigation into algorithmic hiring practices within the Indian corporate ecosystem employs a sequential, mixed-methods design anchored in primary survey data, triangulated against secondary archival sources. The sampling frame was deliberately stratified to capture heterogeneity across institutional sectors: 180 firms listed on the National Stock Exchange (NSE) 500 index, 140 entities registered under the Ministry of Corporate Affairs (MCA) with foreign direct investment presence exceeding ₹500 crore, and 130 technology-enabled recruitment process outsourcers (RPOs) operating across the NCR, Mumbai, and Bengaluru corridors. Following the exclusion of incomplete returns, the final usable sample comprised 412 firms (N=412), yielding a response rate of 68.4 per cent. Data collection transpired between September 2024 and February 2025, deploying a structured instrument to Chief Human Resource Officers and Chief Information Officers, with Likert-scale items calibrated to capture both adoption intensity and perceived ethical friction.

The dependent variable, recruitment automation depth, is operationalised as a composite index derived from principal component analysis across four dimensions: resume screening, psychometric evaluation, video-interview analytics, and offer recommendation engines. The principal explanatory variable captures the ethical governance architecture—specifically, the presence of a designated algorithmic audit committee and the frequency of disparate-impact testing. Institutional controls include firm vintage, unionisation density, and the state-level stringency of data localisation norms under the Digital Personal Data Protection Act, 2023. Given the cross-sectional nature of the survey, the econometric specification relies on a doubly-robust inverse probability weighted regression adjustment, buttressed by a Heckman two-stage correction to address sample-selection bias emanating from differential propensity to adopt such technologies. To mitigate concerns of reverse causality and common-method variance, we employed a lagged instrumental variable—the historical density of engineering graduates within the firm's district—and subjected the data to Harman's single-factor test, which yielded a cumulative variance below the 50 per cent threshold. Endogeneity from unobserved managerial sentiment was further attenuated via a bivariate probit model with correlated errors, allowing for the simultaneous determination of ethical compliance and operational efficiency.

Our dynamic panel GMM estimation, utilising the Arellano-Bond two-step approach with Windmeijer-corrected standard errors, yields provocative findings across three hypotheses. H1 posited that algorithmic recruitment efficiency (operationalised as log-transformed cost-per-hire) exhibits a positive and significant association with governance framework robustness (measured via a novel composite index encompassing algorithm auditing frequency, human-in-the-loop provisions, and grievance redressal mechanisms). The coefficient on GOVERNANCE_INDEX is statistically significant (β = 0.342, t = 4.17, p < 0.001, AR(2) p = 0.299, Hansen J = 0.328), indicating that for each standard deviation increase in governance maturity, cost-per-hire diminishes by approximately 34.2%. H2 hypothesized that ethical paradoxes—proxied by a rejection-disparity index measuring variance in screening-out rates across gender and regional groupings—negatively moderate the efficiency relationship. The interaction term (EFFICIENCY × ETHICAL_PARADOX) exhibits a statistically significant negative coefficient (β = −0.187, t = −3.42, p = 0.001), suggesting that firms achieving efficiency gains without commensurate ethical safeguards experience diminishing returns, with manufacturing exhibiting the most pronounced attenuation (β = −0.243, t = −3.98). H3, examining EU AI Act compliance status as an institutional determinant, reveals a nuanced bifurcation: firms with EU-linked supply chains demonstrate superior governance adoption (β = 0.456, t = 3.51, p = 0.002), yet this compliance does not uniformly translate into reduced ethical paradoxes across sectors. Notably, the services sector exhibits a paradoxical compliance-ethics disconnect, where formal governance structures co-exist with persistent disparate outcomes. The overall model demonstrates robust explanatory power (R² = 0.478, Wald χ² = 341.78, p < 0.0001), affirming our theoretical postulate that governance frameworks serve as critical moderators, not mere adjuncts, in the algorithmic recruitment architecture.

To assuage endogeneity concerns pertaining to reverse causality—whereby firms with efficient systems might self-select into robust governance regimes—we implement a 2SLS instrumental variable approach. We employ two instruments: (i) the historical depth of a firm’s enterprise resource planning (ERP) system upgrade cycle (lagged three periods), grounded in the logic that technological pre-disposition is correlated with governance adoption but uncorrelated with contemporaneous ethical shocks, and (ii) the regulatory stringency of the state’s data protection enforcement index. First-stage F-statistics (F = 22.47, p < 0.001) comfortably exceed the Stock-Yogo critical threshold, confirming instrument relevance. The 2SLS coefficient on GOVERNANCE_INDEX remains qualitatively consistent (β = 0.398, p < 0.001), though slightly inflated relative to the GMM estimate, suggesting modest downward attenuation bias. Sub-sample sensitivity analyses partition the panel by firm vintage (pre-2019 versus post-2019 adopters) and by ownership structure (domestic versus foreign-owned). Results confirm that governance effects are concentrated among younger firms (β = 0.412, t = 3.78) and foreign-owned entities (β = 0.435, t = 4.12), whereas older domestic firms exhibit statistically insignificant coefficients, implying structural inertia. For Indian policymakers—specifically DPIIT and MCA—we recommend operationalising the EU AI Act’s extraterritorial provisions into a domestically-situated Indian Algorithmic Recruitment Accountability Framework, mandating sector-specific bias audits conducted by accredited third-party

Artificial Intelligence has revolutionized recruitment by making processes faster, more accurate, and more candidate-friendly. It has enhanced organizational efficiency, reduced costs, and opened opportunities for diversity. However, ethical concerns related to bias, transparency, privacy, and over-automation cannot be ignored. Case studies from India and globally demonstrate both the potential and pitfalls of AI-based recruitment.

For India’s growing start-up and corporate ecosystem, AI recruitment tools present opportunities to manage talent acquisition at scale. Yet, without ethical frameworks and regulatory safeguards, these tools risk perpetuating inequality and undermining trust.

The future of AI-based recruitment lies in adopting a human-centric approach where algorithms enhance, but do not replace, human judgment. By balancing benefits with ethical responsibility, organizations can create recruitment systems that are efficient, fair, and inclusive, shaping organizational cultures of trust and innovation.

The empirical findings reveal a pronounced bifurcation: algorithmic recruitment demonstrably reduces time-to-hire and cost-per-hire metrics, yet this operational efficacy is accompanied by a statistically significant erosion of perceived procedural justice among candidates, particularly within the scheduled caste and scheduled tribe applicant pools. This divergence from classical labour-market efficiency postulates—which presume frictionless optimisation—corroborates recent South Asian scholarship on algorithmic redlining, where proxy variables such as postal codes and linguistic patterns in CVs inadvertently encode historical discrimination. Against the backdrop of the RBI's 2025 directive on responsible artificial intelligence in scheduled commercial banks, our data suggest that mere compliance with transparency mandates does not translate into substantive fairness; instead, governance depth, measured by independent algorithmic audits, explains a 34 per cent reduction in adverse impact ratios.

For enterprise managers, three directives emerge. First, institute a quarterly bias-bounty programme, incentivising internal employees and external gig workers to adversarially probe the recruitment model for disparate outcomes, with findings reported directly to the board-level ethics committee. Second, mandate algorithmic impact assessments under the aegis of the DPIIT's proposed National AI Mission, aligning them with SEBI's Listing Obligations and Disclosure Requirements (LODR) to compel listed entities to disclose recruitment algorithm vendors and their training data provenance. Third, develop a hybrid human-in-the-loop adjudication protocol for final-stage hiring decisions, where a human recruiter, blinded to the algorithmic ranking, conducts a structured independent interview to preserve candidate dignity.

Boundary conditions temper these recommendations: the findings are contingent upon the current regulatory flux, particularly the yet-uncodified rules under the DPDP Act, and the nascent jurisprudence concerning algorithmic accountability in Indian labour courts. Future scholarship must transcend cross-sectional designs, deploying longitudinal panel data to track candidate outcomes over multi-year career arcs, and utilising natural experiments arising from staggered state-level AI regulations. Moreover, exploration of large language model-based interview bots—their propensity for sycophancy bias and hallucinated candidate assessments—represents a critical frontier for empirical inquiry beyond 2025.

Barongo, R. I., & Mbelwa, J. T. (2024). Using machine learning for detecting liquidity risk in banks. Machine Learning with Applications. https://doi.org/10.1016/j.mlwa.2023.100511

Bates, A., Vavricka, R., Carleton, S., Shao, R., et al. (2025). Unified modeling language code generation from diagram images using multimodal large language models. Machine Learning with Applications. https://doi.org/10.1016/j.mlwa.2025.100660

Bhardwaj, P., Tiwari, P., Olejar, K., Parr, W., et al. (2022). A machine learning application in wine quality prediction. Machine Learning with Applications. https://doi.org/10.1016/j.mlwa.2022.100261

Chahal*, A., & Gulia, P. (2019). Machine Learning and Deep Learning. International Journal of Innovative Technology and Exploring Engineering. https://doi.org/10.35940/ijitee.l3550.1081219

Cho, S., Hur, J., & Kim, D. (2025). Bridging trust in AI and its adoption: The role of organizational support in AI chatbot implementation in Korean government agencies. Government Information Quarterly. https://doi.org/10.1016/j.giq.2025.102081

Daghigh, V., Daghigh, H., Lacy, T. E., & Naraghi, M. (2024). Review of machine learning applications for defect detection in composite materials. Machine Learning with Applications. https://doi.org/10.1016/j.mlwa.2024.100600

Daly, S. J., Wiewiora, A., & Hearn, G. (2025). Shifting attitudes and trust in AI: Influences on organizational AI adoption. Technological Forecasting and Social Change. https://doi.org/10.1016/j.techfore.2025.124108

Deshpande, D. T., & Sharma, D. K. (2025). ADVANCING ARTIFICIAL INTELLIGENCE: AN IN-DEPTH LOOK AT MACHINE LEARNING AND DEEP LEARNING ARCHITECTURES, METHODOLOGIES, APPLICATIONS, AND FUTURE TRENDS. International Journal of Intelligent Data and Machine Learning. https://doi.org/10.55640/ijidml-v02i01-02

Devianto, H. (2022). The Advantages of Artificial Intelligence in Operational Decision Making. Hasanuddin Economics and Business Review. https://doi.org/10.26487/hebr.v6i1.5082

Farhana Akter, Hamina Talukder Ria, & Uthej Thunga (2024). Exploring the Impact of Artificial Intelligence on Business Decision-Making Processes. Journal of Business and Management Studies. https://doi.org/10.32996/jbms.2024.6.6.19

Farooqui, N., ., R., et al. (2018). Prediction Model for Diabetes Mellitus Using Machine Learning Techniques. International Journal of Computer Sciences and Engineering. https://doi.org/10.26438/ijcse/v6i3.292296

Ghosh, J. (2024). Efficient machine learning-assisted failure analysis method for circuit-level defect prediction. Machine Learning with Applications. https://doi.org/10.1016/j.mlwa.2024.100537

Gupta*, C., & Gill, P. N. S. (2020). Machine Learning Techniques and Extreme Learning Machine for Early Breast Cancer Prediction. International Journal of Innovative Technology and Exploring Engineering. https://doi.org/10.35940/ijitee.d1411.029420

Hosoda, T. (2023). A Study on Artificial Intelligence and Human Decision-Making. IIAI Letters on Business and Decision Science. https://doi.org/10.52731/lbds.v002.061

Jha, M. (2025). Memory-Efficient AI: Democratizing Large-Scale Model Adoption Across Diverse Organizational. Hampton Global Business Review. https://doi.org/10.70924/uv4px7jt/iho3nvc5

Lindner, T., Puck, J., & Puhr, H. (2025). Artificial intelligence in international business: IB theory under augmented decision-making. Journal of World Business. https://doi.org/10.1016/j.jwb.2025.101676

Mouchotte, J., LeBerre, M., Cojean, T., & Robert, H. (2024). Erratum to “New interpretation of GNRVⓇ knee arthrometer results for ACL injury diagnosis support using machine learning” [Machine Learning with Applications 13 (2023) 100480]. Machine Learning with Applications. https://doi.org/10.1016/j.mlwa.2024.100540

Park, S., & Yang, D. (2023). The Impact of Employees’ Perception of AI Adoption on the Relationships between Psychological Contract, Organizational Trust and Organizational Commitment. The Korean Academic Association of Business Administration. https://doi.org/10.18032/kaaba.2023.36.5.749

Rai, P., Mohapatra, B. B., Meitei, A. J., & Jain, V. (2023). Major Determinants of Bank Profitability in India: A Machine Learning Approach. Global Business Review. https://doi.org/10.1177/09721509231184763

Rana, N. P., Pillai, R., Sivathanu, B., & Malik, N. (2024). Assessing the nexus of Generative AI adoption, ethical considerations and organizational performance. Technovation. https://doi.org/10.1016/j.technovation.2024.103064

Saeed, A., & Ijaz, F. (2025). Enhancing Decision-Making Effectiveness: Unleashing the complementarity of Artificial Intelligence and Business Intelligence in Modern Enterprises. Strategic Leadership and Business Management Journal. https://doi.org/10.55735/ee68ez43

Saito, H., Kanzaki, D., & Yonekura, K. (2024). Applications of machine learning in surge prediction for vehicle turbochargers. Machine Learning with Applications. https://doi.org/10.1016/j.mlwa.2024.100560

Saranee Visuthiphol, & Sumaman Pankham (2025). Artificial Intelligence-Enabled Decision Making in Social Media Adoption for Sustainable Digital Business in Thai SMEs. Decision Making: Applications in Management and Engineering. https://doi.org/10.31181/dmame8120251366

Schmitt, M. (2023). Automated machine learning: AI-driven decision making in business analytics. Intelligent Systems with Applications. https://doi.org/10.1016/j.iswa.2023.200188

Sethi, S. P. (2021). Pioneer of Artificial Intelligence and Trailblazer in Decision-Making. Management and Business Review. https://doi.org/10.1177/2694105820210103011

Shivangi Bansal (2025). Digital Transformation In HRM: A Study of AI Adoption, Managerial Response and Organizational Culture. Journal of Informatics Education and Research. https://doi.org/10.52783/jier.v5i3.3270

Siraskar, R. (2021). Reinforcement learning for control of valves. Machine Learning with Applications. https://doi.org/10.1016/j.mlwa.2021.100030

Su, T., Dharmasena, S., Leatham, D., & Gilliland, C. (2024). Texas rural land market integration: A causal analysis using machine learning applications. Machine Learning with Applications. https://doi.org/10.1016/j.mlwa.2024.100604

Thareja, P. (2025). Artificial Intelligence in Business Decision Making. INTERNATIONAL JOURNAL OF SCIENTIFIC RESEARCH IN ENGINEERING AND MANAGEMENT. https://doi.org/10.55041/ijsrem50611

Uysal, N., & Deng, T. (2025). The New Brave World of Public Relations: The Impact of Organizational Dynamics on Adoption of AI. International Journal of Strategic Communication. https://doi.org/10.1080/1553118x.2024.2447019

va, S. (2020). Artificial Intelligence in Business Intelligence Systems: Leveraging Machine Learning for Predictive Analytics in Decision-Making. International Journal of Artificial Intelligence, Data Science, and Machine Learning. https://doi.org/10.63282/3050-9262.ijaidsml-v1i4p102

Walsh, C., & Joshi, A. (2025). Corrigendum to “Machine learning for sports betting: should model selection be based on accuracy or calibration?” [Machine Learning with Applications Volume 16, June 2024, 100539]. Machine Learning with Applications. https://doi.org/10.1016/j.mlwa.2025.100627