From Matching Models to Recruiting Agents: A Systematized Narrative Review of AI Recruitment Systems, Evaluation, and Governance

1 Introduction

• The lineage moves from bilateral preferences and platform-scale cascades to neural matching [Qin et al., 2018; Zhu et al., 2018].Later work addressed noisy and sparse interaction data with co-teaching, graphs, and contrastive objectives [Bian et al., 2020; Yu et al., 2024].Source: §1 · p. 2
• LLMs add requirement extraction, shortlist reranking, criterion-level explanations, and web tools [Du et al., 2024; Yu et al., 2025, 2026; Zheng et al., 2026], while agents handle dynamic sourcing tasks [Chen et al., 2025].Some assessment prototypes combine resumes, job descriptions, interview transcripts, and recruiter feedback in rubric-based workflows [Yuksel et al., 2026].Source: §1 · p. 2
• Pair-level classification, ranking, marketplace behavior, and agent success each leave different questions unanswered, so evaluation units must expand with automation scope.Click and application rates conflate exposure, preference, qualification, and interface effects; successful task execution may hide unsafe actions or incomplete coverage.Source: §1 · p. 2
• The review links recruitment stage and automation scope to evaluation units, then maps evidence types to the strongest defensible claim.Its framework separates extraction, recall, filtering, ranking, judgment, action, and outcome, while distinguishing predictive performance, construct validity, business utility, grounding, allocational risk, and outcome evidence.Source: §1 · p. 2, Figure 1 · p. 3
• The review focuses on the evidence chain from pre-LLM search and marketplaces to LLM components, recruiting agents, and profession-aligned evaluation, complementing reviews of individual slices.Its artifacts include a five-layer lineage reconstruction, a triangulated evidence map, and a failure-to-intervention map and research agenda.Source: §1 · pp. 2–3
• As automation expands from profile pairs to ranked lists and case trajectories, evidence must expand from similarity or behavioral labels to coverage, grounding, action safety, and outcomes.The sequence indicates broader system scope and consequence, not a universal ranking of study quality.Source: §1 · p. 3, Figure 1 · p. 3

2 Review Protocol and Evidence Boundaries

• The review is a systematized narrative review, not an exhaustive database census or a PRISMA systematic review.Its protocol distinguishes that scope because proceedings, preprints, product systems, standards, and regulation evolve on different timelines.Source: §2 · p. 5

2.1 Search strategy

• The search combined technical discovery and snowballing with targeted personnel-selection, non-English, industrial, legal, regulatory, and forward-update searches.The primary technical search cutoff was 23 July 2026; targeted updates extended through 2 September 2026, with the forward update not a retrospective rerun of all primary queries.Source: §2.1 · pp. 5–6
• Eligible sources had to concern a recruitment workflow or dependency, be verifiable, and provide enough detail to code task, data, method, metric, and limitation.The representative map retained works changing the system object, learning signal, evaluation unit, or risk surface, rather than near-duplicate variants.Source: §2.1 · p. 6
• The 40 coded works are a structured trace, not a representative sample; judgmental stopping does not establish field-wide saturation.Because no frozen historical export was preserved for every database, the authors do not retrospectively report hit, deduplication, exclusion, screening, or agreement counts.Source: §2.1 · p. 6
• The selected evidence does not support meta-analysis because tasks, labels, outcomes, and designs lack a common effect size.Source-specific appraisal is used instead of one numerical quality score, but it is not a formal risk-of-bias assessment for a future causal-effect review.Source: §2.1 · pp. 6–7

2.2 Coding dimensions

• The coding framework separates system artifact, automation scope, evaluation unit, workflow authority, and employment-decision boundary, while treating several interpretations as author-coded.The companion table is not a complete regional codebook, and translating a benchmark does not reproduce local data conventions, platform behavior, credentials, or legal duties.Source: §2.2 · pp. 7–8

2.3 Evidence boundaries

• An offline gain on private logs supports a claim about the reported task, not improved hiring quality or job performance.Human agreement is not automatically ground truth, and benchmark results are time-stamped observations of a particular environment.Source: §2.3 · p. 8

3 A Unified Framework for Recruitment AI

• The workflow maps a hiring request through requirements, recall, filtering, ranking, evidence, judgment, and policy-guided action, with early-stage errors limiting later recovery.A perfect re-ranker cannot recover a qualified candidate missing from the recalled pool, and a fluent report cannot repair fabricated evidence.Source: §3 · p. 8

3.1 Three axes

• The framework distinguishes five automation scopes, from assistive transformation to authorized action-capable execution, and separates capability from granted permission.A3 requires tool-coverage, freshness, cost, and recovery evidence; A4 additionally requires permission, confirmation, idempotency, audit, and rollback evidence.Source: §3.1 · p. 9
• Figure 3 presents a design synthesis, not a validated deployment, separating context, decision evidence, permission, control, and reciprocal outcomes.Its authority gate is intended to prevent a fluent judgment from silently becoming an external action; connector length represents routing distance, not magnitude.Source: §3 · p. 9, Figure 3 · p. 10
• Evaluation units range from field and pair to list, case, trajectory, and real-world outcome; evidence at one unit does not substitute for another.The paper uses L1–L6 as shorthand for these six units in that order.Source: §3.1 · pp. 10–11

3.2 Illustrative end-to-end case

• Table 2 is a fictional senior-retrieval-engineer design example, not a case study, benchmark result, or validation of its proposed controls.Its caption explicitly reports no empirical results.Source: §3.2 · p. 11, Table 2 · p. 11

3.3 Reciprocity and partial observability

• A reciprocal policy distinguishes candidate willingness to pursue a job from qualification against job-related criteria, using both probabilities and their uncertainty.It can choose to search, ask, defer, recommend, or propose contact while tracking candidate-side measures alongside employer-side qualified yield; acceptance is a downstream outcome, not a proxy for preference probability.Source: §3.3 · pp. 11–12

4.1 Era I: rules, ontologies, and bilateral recommendation

• Early systems made requirements inspectable through Boolean retrieval, curated fields, and taxonomies but were brittle to variation; bilateral recommendation framed matching as two selection processes, while observed job transitions reflected exposure as well as preference.Skill normalization later supported cross-document normalization, hard constraints, and auditability.Source: §4.1 · p. 12

4.2 Era II: industrial retrieval, behavioral learning, and marketplaces

• Industrial recruitment systems separate retrieval from increasingly costly ranking and face multi-sided objectives; clicks are frequent but weak, while offers and hires are sparse and delayed, so conversion gains alone do not validate selection quality.Platform feedback is selection-biased: a missing application may reflect lack of exposure, interest, perceived success probability, or compatibility.Source: §4.2 · pp. 12–13

4.3 Era III: neural semantic matching and structured representations

• Neural matching reduced reliance on exact term overlap, while reciprocal models represented candidate and employer preferences and trajectory models captured temporal development; offline behavioral labels and platform-specific datasets still limit validity.PJFNN learned joint job–resume representations, APJFNN added ability-aware attention, and later work used co-teaching and dual-perspective interaction graphs; predicted next jobs do not evaluate a policy through performance or retention.Source: §4.3 · pp. 13–14

4.4 Retrieval systems require stage-level disclosure

• A role-aware mixture-of-experts ranker with LLM-derived job features reported a 17.29% lift in click-through conversion rate in an online A/B test.The disclosure describes production serving but does not report exact latency or index-update intervals.Source: §4.4 · p. 15
• Retrieval disclosures often omit candidate-pool sizes, negative semantics, latency, refresh cadence, and online metrics, preventing pairwise gains from establishing reachable recall or deployability.Table 3 contrasts fixed-corpus evaluations, production architectures, and open-web agents, whose complete candidate universe may not be enumerable.Source: §4.4 · pp. 14–15

4.5 Era IV: LLM-augmented recruiting pipelines

• LLMs serve as data generators, feature engineers, re-rankers, and document interpreters, but generated biography can be false and model-family reuse can make gains reflect stylistic affinity.LLM-derived job features may be distilled into smaller serving models; resume parsing remains difficult even when outputs are valid JSON, so upstream errors should remain visible downstream.Source: §4.5 · pp. 15–16

4.6 Era V: tool-using recruiting agents

• Agentic recruiting can span role clarification, sourcing, evidence comparison, outreach, and handoff, but adaptive retrieval does not establish complete recall and action-capable systems need separate safeguards.TalentScout elicits and verifies under-specified requests, while TalentTrace evaluates 21 systems over 691 requirements; its pooled multi-model oracle improves open-world comparability without identifying complete recall. Simulated dialogue and recommendation tool-selection are distinct interaction patterns, not a single capability claim.Source: §4.6 · pp. 16–17
• Agents differ operationally from LLMs: they maintain state, use tools, act under a policy, and encounter a changing environment, making the execution trajectory—not just the final report—the evaluation object.Static modular workflows and role simulation are adjacent patterns; A3 adds adaptive tool use, while A4 can initiate consequential external actions that require evidence on permissions, approval, and recovery.Source: §4.6 · pp. 16–17
• Figure 2 frames five historical layers as narrative waypoints, while production systems may combine lexical retrieval, behavioral supervision, neural ranking, LLM evidence modules, and agentic handoff.Comparative analysis uses the three primary axes from Section 3, with model representation, evidence source, supervision signal, market sidedness, decision authority, context, and validity threat as secondary annotations.Source: §4.6 · p. 17

5.1 Requirement understanding and document intelligence

• Document representations should link typed claims to source spans and mark them supported, contradicted, or unknown, so missing evidence is not silently treated as failure.Extraction should preserve provenance and distinguish hard from soft constraints; interactive clarification can improve request formation but does not establish that criteria are job-related or lawful.Source: §5.1 · pp. 17–18

5.2 Candidate and job retrieval

• Retrieval evaluation should use independently judged or otherwise defensible target sets, because observed positive pairs omit unseen interest and candidates outside a platform database.PeopleSearchBench checks returned people against live external evidence and scores relevance, qualified-result yield, and information utility, with temporal and source-availability dependencies.Source: §5.2 · p. 18

5.3 Filtering, re-ranking, and candidate assessment

• Assessment is stronger than ranking: comparisons with recruiter ratings show limited agreement, while controlled tests distinguish qualification ranking from abstention on equivalent candidates.A selective interface can recommend when evidence and separation suffice, or request a fact, show a tie, or defer.Source: §5.3 · p. 19

5.4 Interviewing, communication, and candidate support

• Interview fluency, candidate experience, and selection validity are distinct: dialogue can be irrelevant or overstep consent, and favorable applicant perceptions do not prove predictive validity or equal outcomes.Evaluation should separately examine competency coverage, factual grounding, consistency, accommodation, experience, and score validity.Source: §5.4 · p. 19
• In one firm’s field experiment, AI-interviewer assignment increased offers, starts, and short-horizon retention, while recruiters retained evaluation and hiring authority.Performance records covered only a subset of hires, and the evidence is bounded to one provider, predominantly entry-level service roles, applicant refusal, and technical failures.Source: §5.4 · p. 19
• Human-manager review after AI-based rejection improved applicants’ reactions to the organization, but does not establish that the decision was more accurate or fair.Candidate-facing assistance can homogenize applications; guardrails and human checkpoints reduce different forms of unsupported resume content but do not remove every subtle error.Source: §5.4 · pp. 19–20

5.5 End-to-end agents and human handoff

• A safe recruiting agent is a policy over tools and handoffs, keeping criterion-level evidence and uncertainty visible, separating proposed from executed actions, and producing an audit record.Professional sourcing-task completion measures utility but does not establish that subsequent screening or hiring decisions are job-related, fair, lawful, or causally beneficial.Source: §5.5 · p. 20

• The map preserves source-specific evaluated objects and claim boundaries rather than ranking industrial effects, benchmark scores, and controlled validity audits together.Archival and production studies support claims within reported settings, while 2025–2026 agent rows expose an emerging research surface rather than mature evidence for autonomous employment decisions.Source: §6 · pp. 20–21
• Table 5 maps 40 coded works in 37 display rows; within this purposive set, automation scopes are A0=2, A1=24, A2=8, A3=4, A4=2, and workflow authority is P0=3, P1=27, P2=8, P3=2.These counts describe the selected map, not field prevalence; PJFNN/APJFNN and three ConFit versions share display rows.Source: §6 · pp. 20–21
• MockLLM [Sun et al., 2024] warns that plausible resume completion can add unsupported biography, while modular interview workflows do not validate selection instruments or labor-market outcomes.Source: §6 · p. 22
• A 17-month Jobindex deployment spanning 41,390 jobs reported adoption and task success, but workflow uptake does not establish long-term job performance or general fairness.Source: §6 · p. 22
• Resume-screening validity scores depend on live-web timing, source access, and incomplete ground truth, despite evaluating criterion verification, qualified-result yield, and information utility.The evidence includes controlled qualification differences and equivalence cases.Source: §6 · p. 23
• Assessment stability is construct- and input-dependent: an audit of 94 participants and two commercial systems treats instability as necessary-condition evidence, not criterion validity or a deployed selection effect.The audit varied source, file format, embedded URLs, and time.Source: §6 · p. 23
• Of 70,884 applications received, 67,056 eligible applications were randomized, and the study reports offer, start, and retention outcomes; missing qualification and later-employer outcomes prevent causal attribution.Source: §6 · p. 23

6 Representative Literature and Evidence Map

• Multi-domain benchmark results do not measure production attack prevalence.Source: §6 · p. 23
• Pair- and list-level technical evidence is more developed than trajectory-safety, human-correction, or long-term-outcome evidence in the selected map.This comparison describes the purposively selected set, not field-wide prevalence.Source: §6 · pp. 23–24
• Recent agent benchmarks broaden task scope faster than they establish employment-validity claims, motivating comparisons that connect evaluation units instead of replacing pair metrics with one end-to-end score.The suggested comparisons should link several evaluation units.Source: §6 · p. 24
• Table 6 retains each source’s reported metric, denominator, comparator, uncertainty, and claim boundary rather than pooling unlike tasks or treating their values as directly comparable.Performance deltas, deployment samples, and benchmark scales answer different questions.Source: §6 · p. 24, Table 6 · p. 24

7 Industrial Evidence and Regional Deployment Context

• Industrial evidence answers different questions: production papers may support architecture or online metrics, reports strategic scope, and regulator audits observed compliance failures, but none alone establishes causal hiring quality.Table 7 organizes five public evidence layers; private data and experiments, costs, negative results, and long-term outcomes are often unavailable.Source: §7 · p. 25, Table 7 · p. 25
• Industry papers and product disclosures describe system designs, operational sequences, and selected business metrics, but marketing selection, non-public denominators, and absent independent causal or fairness evaluation limit inference.Model design, slice performance, candidate-level error, and causal contributions to financial outcomes may remain opaque.Source: §7 · pp. 25–26
• Regulator and public disclosures have boundaries: the ICO reported recruitment-tool compliance issues, while its later employer engagement was voluntary and neither source compares predictive quality or generalizes automatically across jurisdictions.A detailed UK pre-deployment record still has rare positive cases, human-derived labels, and undisclosed denominators that constrain inference from perfect sensitivity on one check.Source: §7 · pp. 26–27

7.1 Evidence from non-English and locally regulated markets

• Selected regional rules imply distinct controls and scopes; the review’s comparison is time-stamped and is neither legal advice, a product-specific applicability decision, nor a global census.For example, covered EU recruitment systems are subject to an amended application timetable, while GDPR duties already apply; the UK and EU regimes should not be treated as identical.Source: §7.1 · pp. 27, 29
• Locally situated studies add candidate perspectives and constructs that platform metrics and translated benchmarks omit, but do not estimate prevalence across China.The China–Pakistan convenience samples were retrospective and lacked a randomized human-interview condition, so they do not establish causal country comparisons.Source: §7.1 · p. 27
• External validity depends on language, labor-market practice, platform population, product role, and applicable law, not simply whether a paper is English-language or translated.A credible multi-region study should report all five and avoid extrapolating from one platform’s behavior labels or one jurisdiction’s audit definition.Source: §7.1 · p. 30

8.1 Personnel-selection validity is a use argument

• Selection validity connects job constructs, measures, criteria, and decisions; consistency alone is insufficient, and studied LLM scores showed modest, information-dependent convergence and sensitivity to input perturbations [Rhea et al., 2022].Validation should test future job-relevant criteria, differential prediction, subgroup error, incremental validity, and operational utility beyond a strong existing procedure.Source: §8.1 · pp. 30–31

8.2 Labels are constructs, not facts

• Observed clicks, applications, advances, interviews, offers, and accepted jobs measure distinct stages, so exposure-gated nonobservations are not substantive negatives and behavioral outcomes do not identify latent suitability without additional assumptions.Unobserved pairs are not necessarily unsuitable, and agreement with one recruiter should not be called human-level performance without reporting recruiter variability.Source: §8.2 · pp. 31–32

8.3 Retrieval and ranking

• nDCG@K evaluates ordering against graded relevance, but its interpretation depends on complete, independent judgments; recruitment rankings also need stability checks under equivalent documents, candidate permutations, and changing pools.Reports should describe the candidate universe and include recall, precision or nDCG, qualified-result yield, and slice coverage.Source: §8.3 · p. 32

8.4 Grounding, uncertainty, and selective prediction

• Grounded LLM evaluation must separately measure whether citations support claims and cover material claims, while selective prediction should show risk as acceptance thresholds vary and account for review and evidence-acquisition costs.Recruitment uncertainty is typed: missing information differs from contradictory evidence, and targeted requests such as proof of certification differ from global refusal.Source: §8.4 · p. 32

8.5 Human–AI evaluation

• Deployment studies offer workflow and outcome evidence: the Jobindex recommender was studied across 17 months and 41,390 jobs, while a randomized AI-voice interview study reported offers, starts, retention, and productivity [Jabarian and Henkel, 2026].The interview intervention automated administration but left evaluation and hiring to human recruiters; its causal evidence is specific to that setting, not automated final selection or other labor markets and job families.Source: §8.5 · p. 33

8.6 Agent and profession-aligned benchmarks

• xbench links professional recruiting tasks to workflow value, PeopleSearchBench evaluates factual verification, and TalentTrace tests 691 requirements across 21 systems, shifting evaluation toward usable, verifiable work products [Chen et al., 2025; Shi et al., 2026; He et al., 2026].TalentTrace exposes failures before retrieval, but union-pooled candidates and multi-model judging remain system-dependent approximations to open-world recall.Source: §8.6 · p. 33

8.7 A tool trajectory is not a hiring lifecycle

• A tool trajectory ends with search, evidence collection, shortlist, and handoff, whereas a hiring lifecycle extends through employment outcomes; lifecycle evaluation must use timestamp-available information and address delayed labels, censoring, and market interference.An audit of roughly 497,000 candidate–vacancy records found aggregate binary-gender parity alongside age, salary-level, and intersectional disparities, but missing qualification and employer data limit causal and selection-validity interpretation [Galdón-Clavell, 2026].Source: §8.7 · p. 34

8.8 Evidence requirements by claim level

• The six evidence levels run from field/document through pair, list, case, and trajectory to outcome; conclusions should stop at the highest-consequence unit for which matching evidence is supplied.The levels order system scope and consequence, not methodological strength, and fairness and security apply at each relevant unit.Source: §8.8 · pp. 34–35

9.1 Fairness is stage- and task-dependent

• Fairness findings vary with task, population, prompt, proxy, model, and metric; paired-resume tests found near-zero effects under clean conditions but small effects in some combinations, while a pipeline audit found disparities hidden by aggregate parity [Galdón-Clavell, 2026].The studies ask different questions and do not provide a portable fairness certificate; aggregate parity can hide intersectional harms, and accommodations may still expose candidates to sensitive-data or disability-related risks.Source: §9.1 · pp. 35–36
• Fairness-aware re-ranking in LinkedIn Talent Search post-processes a top-K list toward a specified representation distribution while preserving utility, though the target distribution and protected-attribute policy remain normative choices.Counterfactual audits report name-associated differences in hiring and salary recommendations, disability-related sensitivity in resume screening, ranking changes after small demographic perturbations, and group-difference directions that vary across settings.Source: §9.1 · p. 35

9.2 Privacy, consent, and purpose limitation

• Open-web sourcing aggregates personal facts across contexts, so accountable systems should specify purpose, minimize retention, separate search from storage, record provenance, and enable correction of consequential inaccuracies.The passage notes jurisdiction-specific rules, including GDPR Article 22 safeguards for specified solely automated decisions.Source: §9.2 · p. 36

9.3 Untrusted inputs, identity, and tool security

• Recruiting agents face instruction and ranking attacks, identity and provenance failures, and cross-tool or memory risks; RAPIDS uses a high-recall detector followed by LLM verification to reduce cost and latency while retaining attack coverage.LongPIBench found defenses that appear effective on short inputs can be bypassed more often as context grows; this is benchmark evidence, not a production attack rate [Liu et al., 2026].Source: §9.3 · pp. 36–37

9.4 Regulation and employment validity

• Legal obligations vary by jurisdiction and actor, and this time-stamped technical map is not legal advice or a durable compliance checklist; the review says internal-dataset accuracy alone is insufficient.Covered high-risk systems entail distinct provider and deployer responsibilities, and the cited EU regulation sets 2 December 2027 as the application date for specified provisions and systems [European Parliament and Council of the European Union, 2026].Source: §9.4 · pp. 37–38

9.5 Accountability at system boundaries

• A deployment record should assign who owns requirements, changes scores, approves external actions, and handles candidate challenges; interviews with 21 developers and vendors found diversity and debiasing often described as retrofits.Model cards alone are too narrow for compound systems; system cards should also record retrieval corpus, filter logic, prompts, permissions, retention, and incident procedures.Source: §9.5 · p. 38

10 From Failure Modes to Intervention Paths

• The intervention map separates recruitment-tested approaches from transferred methods and unvalidated design proposals, so concrete solution families should not be mistaken for demonstrated remediation effects.The review says risk diagnosis is stronger than proof of end-to-end remedies; Table 9 makes the evidence states explicit.Source: §10 · p. 38, Table 9 · p. 38
• Exposure-biased clicks and applications can be addressed with exposure propensities, counterfactual or propensity-weighted learning, and separate job- and candidate-exposure models, but propensity misspecification and feedback remain concerns.Hybrid retrieval can combine lexical, dense, behavioral, and graph recall, but the review notes limited open full-corpus targets [Yu et al., 2024, 2026].Source: §10 · p. 38
• For representation disparity in top-K, declared-target re-ranking and stage-specific audits are proposed, but target choice, unknown attributes, intersectional uncertainty, and downstream opportunity still require scrutiny.The intervention map also calls for counterfactual tests, adverse-impact analysis, validity checks, and periodic monitoring.Source: §10 · p. 39
• Unsupported scores or rationales call for atomic criteria linked to source spans, explicit supported, contradicted, and unknown states, and abstention when evidence is insufficient [Buchmann et al., 2024; Castleman et al., 2026].Evidence-first interfaces still need testing for independent expert correctness, omission detection, and recruiter error rather than user preference.Source: §10 · p. 39
• Human review should route cases when expected review utility exceeds automation, but evaluation must address reviewer heterogeneity, automation bias, workload, differential deferral, override quality, and final-action accountability [Mozannar and Sontag, 2020; Verma and Nalisnick, 2022].The proposed routing approach also includes mandatory deferral for sensitive, conflicting, or externally acting cases.Source: §10 · p. 39
• For untrusted inputs and unsafe actions, the proposed path combines high-recall detection, long-context stress tests, source and identity binding, isolated case memory, least-privilege tools, and confirmation before outreach or record mutation [Liu et al., 2026].Identity, memory, and full-workflow defenses remain design synthesis; testing should include collisions, stale evidence, leakage, recovery, and end-to-end attack success.Source: §10 · p. 39

10.1 A testable human-handoff contract

• The proposed handoff contract specifies triggers, an evidence bundle, competent reviewer authority, a boundary between recommendation and execution, and contest and learning procedures.A versioned record should preserve provenance, uncertainty, proposed action, reviewer decision, timestamps, and recovery state.Source: §10.1 · p. 40
• The handoff contract is an outlook, not a validated recruitment architecture; a credible study would compare it with ordinary review under fixed human time and measure quality, correction, latency, automation bias, subgroup deferral, and appeals.The DfE vacancy-QA transparency record offers a pre-deployment pattern, not a validated general recruitment architecture [UK Department for Education, 2026].Source: §10.1 · p. 40

11.1 Separate preference, qualification, and opportunity

• Future datasets should separate candidate preference, employer qualification, exposure, and process opportunity, then test reciprocal objectives under alternative negative-sampling and exposure models.The recruitment-specific challenge is to model both sides’ exposure and response without treating either propensity model as known.Source: §11.1 · p. 40

11.2 Build open, multilingual, full-pipeline benchmarks

• Full-pipeline benchmarks should include heterogeneous documents, multilingual jobs, hard and preferred criteria, independent recall targets, criterion-level evidence, partial candidate orders, and allowable actions.Synthetic control should be distinguished from external validity; translated English benchmarks do not recreate local credentials, labor-market conventions, platform behavior, or regulation.Source: §11.2 · p. 41

11.3 Evaluate counterfactual reachability, not only final accuracy

• Benchmarks should report eligible reach and maximal reachable performance, and use gold replacements for OCR, retrieval pools, or evidence to identify binding stages rather than attributing every failure to the final LLM.Downstream metrics require their own assumptions about ordering, utility, and missing relevance judgments.Source: §11.3 · p. 41

11.4 Move from plausible explanations to verified evidence

• Recruitment explanations should use atomic claims linked to source spans or dated evidence, represent unknown and contradictory information, and test whether evidence-first interfaces improve human error detection.The agenda contrasts error detection with measuring only whether users prefer longer rationales.Source: §11.4 · p. 41

11.5 Design selective and budgeted human collaboration

• Human collaboration should compare uncertainty-based deferral, critical-risk rules, and learned policies under the same review budget, measuring independent judgment, correction, time, automation bias, and differential deferral.Recruitment policies must also route mandatory legal or policy exceptions that cannot be reduced to model uncertainty [Mozannar and Sontag, 2020; Verma and Nalisnick, 2022].Source: §11.5 · p. 41

11.6 Treat temporal validity as a first-class property

• Sourcing benchmarks need time-indexed evidence, freshness thresholds, and conflict handling, while leaderboards should retain stable and rotating task sets to separate model progress from task replacement.Candidate roles, publications, locations, and availability change over time.Source: §11.6 · p. 41

11.7 Measure causal utility and long-term effects

• Deployment studies should track funnel outcomes, appeals, and longer-term criteria with frozen prediction-time information, eligible populations, follow-up windows, delayed labels, censoring, and a longitudinal estimand.When randomization is infeasible, studies should state quasi-experimental identification assumptions; competing candidates, jobs, recruiter policies, and market shocks create interference.Source: §11.7 · p. 42

11.8 Evaluate multi-objective trade-offs jointly; 11.9 Make benchmarks contestable

• In the coded set, quality, utility, cost, fairness, and security occur in different combinations; privacy is not directly evaluated, and no row jointly evaluates all six axes, so compound systems need Pareto analysis.Benchmarks should be contestable by candidates and other stakeholders and report quality–latency–cost curves, slice performance, and attack robustness rather than optimizing one actor’s productivity alone.Source: §11.8 · p. 42, §11.9 · p. 42

12 Limitations

• The 40 coded works are a purposive map, not a statistically representative sample, so the review does not estimate method or finding prevalence.Item-level coding is also an author interpretation, not independently dual-coded; the companion file exposes those judgments without providing inter-rater reliability evidence.Source: §12 · p. 43
• English-language discovery predominates; targeted Chinese-language and China-situated additions improve regional specificity but do not provide broad multilingual or global coverage.Much 2025–2026 recruiting-agent evidence consists of preprints, benchmarks, and demonstrations, establishing emerging artifacts and evaluation problems more than stable deployment effects.Source: §12 · p. 43
• Private data and confidential experiments limit independent validation of production claims and obscure negative results, subgroup errors, costs, and long-term employment outcomes.The regional legal comparison is a technical map checked through 2 September 2026, not legal advice or a durable compliance checklist; these limits constrain generalization without invalidating source-specific evidence.Source: §12 · p. 43

13 Conclusion

• Recruitment AI has progressed from rule-based search and bilateral recommendation through behavioral ranking and neural matching to LLM pipelines and tool-using agents, without a clean replacement sequence.Mature systems still need lexical recall, structured constraints, inexpensive rankers, and human approval; LLMs add flexible interpretation and action alongside uncertainty and control failure.Source: §13 · p. 43
• An upstream result can bound but cannot validate a later decision unit; employment claims require construct validity, temporal controls, stage diagnostics, human collaboration, outcomes, and governance audits.Randomized interview evidence and a public-employment audit show higher-level studies are feasible, while interpretation depends on workflow, population, human authority, and missing downstream labels [Galdón-Clavell, 2026; Jabarian and Henkel, 2026].Source: §13 · p. 43
• Triangulating academic, industrial, corporate, regulator, and regional-rule sources preserves their distinct evidentiary roles; solution work supports intervention families but does not show the full recruitment problem is solved.The cited intervention families include counterfactual learning, hybrid recall, fairness-aware re-ranking, grounded abstention, selective deferral, and privilege-separated action.Source: §13 · p. 43
• The conclusions are bounded by a selected map that is not a field-prevalence estimate and a time-stamped legal comparison that is not a compliance checklist.These boundaries follow the review design described in Section 12.Source: §13 · p. 43
• A credible recruiting agent should find relevant candidates without erasing the long tail, distinguish unknown from negative evidence, ground consequential claims, and preserve preference, correction, selective deferral, and authorized action boundaries.The review frames progress as a more reliable, reciprocal, contestable, and useful recruitment process—not a more persuasive hiring recommendation.Source: §13 · p. 44

14 Materials Availability

• The companion files literature_map_coding.csv and supplementary_search_log.csv are distributed with the manuscript source bundle.The coding boundaries and field definitions are documented in Section 2.Source: §14 · p. 44
• The manuscript release is version 1.13, dated 3 September 2026.The passage provides SHA-256 checksums for both companion files to identify the reviewed source artifacts in this release.Source: §14 · p. 44
• The companion files’ coding boundaries and field definitions are documented in Section 2.The listed checksums identify the reviewed source artifacts for this release.Source: §14 · p. 44

A Recruitment-AI Evidence Landscape

• The inner ring groups system objects, the middle ring separates recurring mechanisms and evidence questions, and the outer ring identifies data, task, metric, and evidence forms.The figure presents a compact topology of the representative literature map.Source: A Recruitment-AI Evidence Landscape · p. 44, Figure 4 · p. 45
• Sector size is an organizing device, not an estimate of field prevalence.Table 5 carries the source-specific evidence and claim boundaries used in the main synthesis.Source: A Recruitment-AI Evidence Landscape · p. 44, Figure 4 · p. 45
• Table 5 maps all 40 coded works to system objects, reported evidence, and claim boundaries using author–year citations.The figure is placed in the appendix because Table 5 carries the source-specific evidence and claim boundaries used in the main synthesis.Source: A Recruitment-AI Evidence Landscape · p. 44, Figure 4 · p. 45