Poster #40 - The Diminishing Returns to Human Recruiting in Online Labor Markets

• In Event: Saturday Poster Luncheon

Abstract

As algorithmic screening tools improve, when do human recruiters still add value? The recruiting industry is enormous: employment and staffing services generate roughly $200 billion annually in the US and nearly $900 billion globally. Theory predicts that intermediaries add value by diversifying sampling risk and expanding employer search, yet the empirical evidence on whether recruiting assistance helps employers make better matches is limited.

We study this question using a randomized experiment on a large online labor market. Over six months, 83,017 hourly job posts were randomly assigned to receive human hiring assistance (75%) or to serve as controls with access only to the platform's algorithmic tools (25%). Treated employers were paired with a hiring consultant who communicated directly with them and a recruiter who could invite outside workers to apply and shortlist existing applicants. We complement the experiment with a regression discontinuity design exploiting a discrete algorithmic recommendation label and develop a theoretical model of delegated recruiting to rationalize the patterns we find.

Human consultants were highly active: treated posts received 16% more applications and 35% more interviews, driven by recruiting invitations and shortlisting. Treated employers also increased their own shortlisting effort, suggesting consultants complemented rather than crowded out employer engagement. Despite this activity, treatment had no detectable effect on whether employers made a hire, with fill rates statistically indistinguishable across arms. Match quality declined: treated employers spent significantly less on hires in the first 30 days and their workers completed fewer total hours, with effects concentrated in hourly contracts where hours directly track the duration and intensity of the employment relationship. Even under optimistic assumptions, using the upper bound of the confidence interval on the hiring effect, the program cannot cover its costs.

Effects are heterogeneous in ways consistent with the model. Customer service jobs and the lowest expertise tier, settings where algorithmic scores plausibly capture less of worker quality, show positive hiring effects that survive multiple-hypothesis corrections. The regression discontinuity confirms that the platform's algorithmic recommendation label increases hiring, particularly when applicant pools are thin, and that recruiters respond to the same label, consistent with both parties drawing on a common signal.

We rationalize these patterns with a model in which recruiters and employers observe the same noisy platform profiles: when their assessments are highly correlated, recruited candidates can be positively selected on observable characteristics, yet fit specific jobs worse on dimensions only the employer can assess, and as algorithms improve, the domain where human recruiters can access independent information shrinks.

Author

• John Fallon, Brown University
  Presenting Author