When Should a Principal Delegate to an Agent in Selection Processes?

Benjamin Fish, Diptangshu Sen, Juba Ziani

TL;DR

This paper provides a theoretical framework to determine when a decision-maker (the principal) should handle selection processes personally versus delegating them to an automated tool or third party (the agent). It reveals that while outsourcing saves review costs, it introduces a misalignment in evaluating candidate "fit." Crucially, the paper demonstrates that retaining control allows humans to correct for known group biases, but can lead to the complete rational exclusion of minority groups when their evaluation data is simply noisier.

What Problem Does This Paper Address?

With the rise of agentic AI and automated software-as-a-service tools, organizations frequently outsource high-stakes selection tasks like hiring, admissions, and talent promotion. This creates a classic principal-agent dilemma. The principal (e.g., a hiring manager or a university professor) cares about both an applicant's core ability and how well they "fit" into a specific team or project. However, evaluating fit is a manual, costly process.

Delegating to an agent (like an AI screening tool or a centralized committee) eliminates evaluation costs for the principal, but the agent typically screens candidates purely on easily quantifiable, noisy signals of ability (like test scores or GPA), ignoring fit entirely. This paper maps out the exact conditions under which delegation benefits the principal, evaluating the trade-offs across three dimensions: the principal's net utility, the actual quality of the selected candidates, and demographic fairness.

Methods

The authors develop a stylized, theoretical principal-agent model rooted in statistical decision theory.

  • Applicant Modeling: Applicants possess two unobservable traits: true ability ($s$) and true fit ($f$), both modeled as independent Gaussian random variables.
  • Signals: The agent only observes a noisy ability signal ($\tilde{s}$). The principal can observe both $\tilde{s}$ and a noisy fit signal ($\tilde{f}$), but must pay a fixed marginal review cost ($c_{rev}$) per application checked.
  • Regimes: The paper mathematically characterizes two environments:
  • Delegated Regime: The agent applies a strict threshold on the ability signal ($\tilde{s} \ge \tau_1$).
  • Non-Delegated Regime: The principal optimizes both the number of applications to review ($n_{rev}$) and a customized quality threshold ($\tilde{\tau}$) to maximize their net utility.

To evaluate fairness, the authors introduce a multi-group setting with an advantaged group ($A$) and a disadvantaged group ($B$). They model disadvantage in two distinct ways: as a negatively biased signal mean (Group B's scores are systematically lower) or as disparate variance (Group B's signals are noisier and less reliable).

Datasets

Not discussed in the paper. This work is entirely theoretical and mathematical; it uses synthetic numerical simulations to visually illustrate the behavior of its derived equations rather than empirical datasets.

Key Insights and Ideas

  • The "Equal Weight" Preference Trap: Intuitively, one might think that the more a principal cares about fit, the less they would want to delegate to an agent who ignores it. However, the paper reveals a non-monotonic (inverted U-shaped) relationship. When a principal places roughly equal weight on ability and fit, finding a candidate who excels at both is incredibly difficult and costly. In this intermediate zone, delegating to an agent is actually more attractive because the agent's high ability standard compensates for the lack of fit evaluation.
  • The Paradox of Noisy Data: Under a group-blind delegated AI agent, if Group B has a noisier signal (higher variance), they actually end up over-represented among selectees at high thresholds due to the fatter tails of their distribution. However, the expected quality of these selectees is lower, which can dangerously reinforce negative societal stereotypes.
  • Rational Exclusion under Human Review: Most strikingly, if the principal retains control and acts as a purely rational utility-maximizer, a noisier signal from Group B leads to complete exclusion (Theorem 4.7). Because the data is unreliable, the principal refuses to waste costly review resources on Group B applicants, opting to hire exclusively from Group A.
  • Bias vs. Noise Mitigation: The nature of a group's disadvantage dictates whether fairness can be salvaged. Constant additive biases (like systemic differences in standardized testing) can be perfectly corrected by a human principal to achieve demographic parity without losing utility. Disparate noise, however, cannot be fixed without sacrificing selection efficiency.

Key Results

  • Viability Threshold: The authors prove that conducting selection personally is only viable for a principal if the ratio of applicant diversity to signal noise exceeds the review cost: $$\frac{\sigma_t^2}{\sigma_{\tilde{t}}} \cdot \frac{1}{\sqrt{2\pi}} > c_{rev}$$
  • The Exclusion Principle (Theorem 4.7): In a mixed pool where Group B suffers purely from a noisier signal, the optimal human strategy sets the number of reviewed applications from Group B to exactly zero ($n_{rev}(B)^* = 0$).
  • Perfect Parity via Correction (Theorem 4.8): If Group B's disadvantage is a known mean bias ($\beta$), the principal can perfectly offset it by shifting their internal threshold ($\tilde{\tau}_B^* = \tilde{\tau}_A^* - (1-\alpha)\beta$), achieving flawless demographic parity.

Limitations and Future Work

The insights are bound to the assumptions of a stylized, one-shot mathematical model. The traits (ability and fit) are assumed to be completely independent and normally distributed, which may not hold true in real-world environments. Additionally, the model assumes the principal perfectly knows the exact parameters of the bias or noise affecting the minority population.

The authors suggest that future extensions should explore multi-stage sequential screening pipelines (e.g., an AI agent doing an initial cheap screen followed by a human evaluating fit) and settings where the parameters of group disparities must be dynamically learned over time.

Summary updated

Abstract

Decision-makers in high-stakes selection processes often face a fundamental choice: whether to make decisions themselves or to delegate authority to another entity whose incentives may only be partially aligned with their own. Such delegation arises naturally in settings like graduate admissions, hiring, or promotion, where a principal (e.g. a professor or worker) either reviews applicants personally or decisions are delegated to an agent (e.g. a committee or boss) that evaluates applicants efficiently, but according to a potentially misaligned objective. We study this trade-off in a stylized selection model with noisy signals. The principal incurs a cost for selecting applicants, but can evaluate applicants based on their fit with a project, team, workplace, etc. In contrast, the agent evaluates applicants solely on the basis of a signal that correlates with the principal's metric, but this comes at no cost to the principal. Our goal is to characterize when delegation is beneficial versus when decision-making should remain with the principal. We compare these regimes along three dimensions: (i) the principal's utility, (ii) the quality of the selected applicants according to the principal's metric, and (iii) the fairness of selection outcomes under disparate signal qualities.
BibTeX
@misc{fish2025when,
  author = {Benjamin Fish and Diptangshu Sen and Juba Ziani},
  title = {{When Should a Principal Delegate to an Agent in Selection Processes?}},
  year = {2025},
  eprint = {2502.07792},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2502.07792},
}
Figure