A GenAI CV assistant pilot made screening more uniform and pushed recruiters from reading CVs to verifying AI summaries, but the evidence is limited to one enterprise rollout without formal outcome measurement.

Main Finding

A generative-AI (GenAI) CV-analysis assistant, architected with explainability, verifiability and human overridability, can improve screening efficiency and uniformity while reshaping recruiters’ tasks from manual reading to verification, interpretation and judgment. Recruiter acceptance depends more on system design (transparency, traceability, user control, security) than on automation per se. Benefits are largest for high-volume, low-complexity hiring; for specialist/senior roles the assistant functions primarily as a complement to recruiter judgement.

Key Points

• Theoretical contributions
  • Integrates organisational justice theory, role identity theory and contingency theory to frame GenAI deployment in recruitment.
  • Proposes three propositions:
  • P1: Recruiter perceptions of procedural fairness are driven by explainability, verifiability and overridability in system design, not by automation alone.
  • P2: Introduction of GenAI shifts recruiter tasks from manual searching/reading toward verification, interpretation and judgement; experienced as competence-enhancing when meaningful evaluative discretion remains.
  • P3: Efficiency and decision-quality gains are greater for high-volume, lower-complexity roles; for specialist/senior roles GenAI is a complement rather than a substitute.
• Design principles (co-design, explainability/traceability, user control, security/compliance) guided development.
• System architecture
  • Cloud-native, layered design deployed on Microsoft Azure.
  • Core tech: Azure OpenAI Service (GPT-4o), Azure Document Intelligence (CV parsing), Azure Cognitive Search (indexing), Azure SQL/Storage, Key Vault, Entra ID, App Gateway + WAF, telemetry via Application Insights/Azure Monitor.
  • End-to-end workflow: CV ingestion → Document Intelligence extraction → indexing → conversational GenAI queries with links back to source CVs.
• Empirical/implementation observations (pilot, single enterprise)
  • Pilot evidence (usage logs, design docs, informal recruiter feedback) suggests improved uniformity of screening and faster information retrieval.
  • Recruiters appreciated transparency and the ability to query and override AI outputs; this increased trust and acceptance.
  • Recruiter role shifted toward verification and candidate communication; perceived effect on professional identity varied depending on evaluative discretion retained.
• Limitations
  • Single-enterprise case, preliminary and qualitative; no formal pre-post survey, controlled interviews, nor objective causal metrics (time-to-hire, quality-of-hire) were reported.
  • Generative outputs are probabilistic and can vary across runs, posing auditability challenges.

Data & Methods

• Research design: Design-science approach producing a deployable artefact plus a single-organisation case study (large enterprise, multiple business units, high-volume hiring).
• Co-design process: Agile iterative development with HR practitioners; prototyping, internal testing, and refinement based on recruiter feedback.
• Data sources for implementation observations:
  • System design documentation and architecture artefacts.
  • Configuration and usage logs from the pilot rollout.
  • Informal feedback gathered from recruiters and hiring managers during early use.
• Analysis: Thematic analysis structured around the three theoretical propositions (procedural justice/explainability, task-shift/identity, contingency/role-type value).
• Missing from current study: controlled quantitative evaluation, formal qualitative interview protocol, longitudinal HR outcome metrics — proposed as future work.

Implications for AI Economics

• Task-based labor effects
  • Evidence aligns with task-based models of technical change: GenAI substitutes routine, information-extraction tasks while complementing higher-order verification, judgment and interpersonal tasks.
  • Expect job-task reallocation rather than pure displacement for recruiters who retain evaluative control; potential for skill upgrading (interpretation, candidate communication) and changes in hiring job-bundles.
• Productivity and heterogeneity of returns
  • Largest productivity gains for high-volume, well-specified roles (graduates, entry-level, operational hiring).
  • Diminishing marginal returns for idiosyncratic or senior searches where tacit knowledge matters; in those contexts GenAI yields smaller efficiency gains and more value as a drafting/research aid.
  • Heterogeneous firm-level adoption: centralised HR functions and scale economies in applicant volume increase expected ROI.
• Adoption and diffusion considerations
  • Adoption depends on perceived fairness, explainability and control — not only on efficiency gains. Design features that enable traceability and human-in-the-loop control reduce resistance.
  • Firms face adoption costs beyond licensing: secure enterprise integration, compliance, role redesign and training costs.
• Governance, regulation and externalities
  • Explainability and auditability requirements increase governance costs but can be decisive for legal/regulatory compliance and reputational risk management.
  • Probabilistic nature of GenAI outputs creates audit and contestability externalities; incomplete audit trails can raise litigation/reputational risks and potentially negative externalities on applicants.
• Market implications
  • Demand for modular, explainable GenAI HR tools that integrate with enterprise identity and security systems; vendors that prioritize traceability and human control may capture larger enterprise market share.
  • Potential for new complementary markets: auditing tools, provenance-tracing, and recruiter retraining/upskilling services.
• Measurement and empirical agenda for economists
  • Recommended outcome metrics to collect: time-to-screen, time-to-hire, interview-to-hire ratios, cost-per-hire, quality-of-hire (performance/retention), shortlist diversity metrics, recruiter satisfaction/turnover, and contestation/appeal rates.
  • Suggested empirical designs: randomized controlled trials (rollout by vacancy or unit), difference-in-differences with staggered adoption, matched-pair designs across role types, and mixed-methods longitudinal studies combining objective HR metrics with recruiter and candidate surveys/interviews.
  • Causal questions to test: (a) Does GenAI improve hire quality net of selection bias? (b) Are diversity outcomes improved or worsened conditional on design choices? (c) How do task reallocation and wage/income effects unfold for recruiters over time?
• Theoretical implications
  • Supports refinement of skill-biased technical change models to incorporate augmentation via generative models (not just automation of scoring).
  • Suggests modelling adoption decisions as a function of role-complexity composition, firm size, centralization of HR, and governance costs.
• Policy relevance
  • Regulators and standard setters should emphasize requirements for traceability, human-in-the-loop controls, and data protection in GenAI HR tools.
  • Policymakers may need to consider support for retraining and certification pathways for HR professionals as GenAI shifts core tasks.

Overall, the paper highlights that the economic effects of GenAI in recruitment will be strongly mediated by system design and organisational context: benefits are real but heterogeneous, and governance and skill-upgrading investments are crucial for realizing productivity gains while maintaining fairness and legitimacy.