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

TEXT START: Artificial intelligence in recruitment has shifted the object being automated from profile pairs and ranked lists to multi-stage workflows that retrieve evidence, compare candidates, and support or execute actions.

The Dissection

This is a taxonomy and governance memo for the industrialization of hiring. It documents the conversion of recruitment from a prediction problem into an executable selection pipeline: retrieving evidence, comparing candidates, assessing them, sourcing, interviewing, and taking action. Its real contribution is recognizing that the unit of automation is now the compound workflow, not the isolated model.

The paper also reveals the institutional response to that transition: make agentic selection measurable, auditable, contestable, and legally defensible so it can scale.

The Core Fallacy

It treats recruitment quality as if it were employment-system health. Better matching, reciprocal suitability, and productivity-aligned evaluation can improve the allocation of remaining jobs while reducing the need for recruiters and other cognitive workers.

Under the Discontinuity Thesis, these agents are not merely better gatekeepers. They are substitutes for labor and force multipliers for whoever owns them. Governance that makes selection cheaper, faster, and more accurate can accelerate the break in the mass employment → wage → consumption circuit. The paper is refining the machinery of labor selection while ignoring the possibility that the machinery is making most participants economically unnecessary.

Hidden Assumptions

• A large and stable supply of jobs will remain for the improved system to allocate.
• Human handoff and contestability can remain meaningful once agentic workflows operate at scale.
• Suitability, qualification, and behavioral evidence are sufficiently objective to govern without reproducing power asymmetries.
• Utility, fairness, privacy, and security can be optimized without addressing ownership or control of the systems.
• Better matching benefits candidates rather than primarily empowering employers and AI-capital owners.
• Employment outcomes remain socially positive even when automation is reducing employment itself.
• Technical evaluation can contain the consequences of structural labor displacement.

The abstract is honest about methodological limits—private and synthetic data, pipeline failures, and missing joint evaluation. That caution improves its empirical discipline, but it does not repair its macroeconomic blind spot.

Social Function

Primary classification: transition management with a partial-truth core and an ideological-anesthetic effect.

The review supplies institutions with a vocabulary for deploying recruiting agents responsibly enough to survive scrutiny. Its concerns are real, not fabricated: weak labels, privacy gaps, uncertainty, and hidden pipeline failures matter. But by defining the problem as evidence quality, fairness, auditability, and risk control, it leaves ownership, bargaining power, labor demand, and productive participation outside the frame.

It makes the transition governable without asking whether the old economic order remains viable after the transition.

The Verdict

The paper is a useful autopsy of recruitment technology and an inadequate diagnosis of the system recruitment serves. It correctly identifies the shift from matching models to operational agents. It does not establish that those agents preserve wages, mass employment, or productive participation.

Under DT logic, recruiting agents are instruments for sorting the shrinking human labor residue—and for automating the people who performed the sorting. The paper documents the selection machinery becoming competent while omitting the mass whose economic role that competence makes unnecessary.