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.
The Selection Dilemma: Fit vs. Ability
Selection processes—ranging from university admissions to corporate hiring—are increasingly mediated by automated agents or specialized committees. While these agents offer efficiency, they often operate with a narrower set of criteria than the ultimate decision-maker (the "Principal"). This research investigates the fundamental trade-off between the high cost of evaluating nuanced criteria like "cultural fit" and the efficiency of delegating to an agent that evaluates only "raw ability."
In this stylized model, applicants possess two hidden traits: ability ($s$) and fit ($f$). The Principal seeks to maximize a weighted utility $t = \alpha f + (1-\alpha)s$. However, while ability signals ($\tilde{s}$) are readily available to everyone, evaluating fit requires a costly "review fee" ($c_{rev}$) per applicant. This creates two distinct regimes:
- Delegation: An Agent (e.g., an AI screening tool or HR department) sets a threshold on ability signals and hires those who pass. The Principal avoids all review costs but receives candidates selected without regard for fit.
- No Delegation: The Principal pays the review cost to observe both signals and sets an optimal threshold based on the full utility metric.
The "Fit" Non-Monotonicity
A counterintuitive finding of the research is that the Principal's preference for delegation is not a simple linear relationship with the importance of "fit" ($\alpha$). One might assume that as fit becomes more critical, the Principal would always prefer to handle selection personally. However, the model reveals a non-monotonic utility curve. At intermediate levels of $\alpha$, the difficulty of finding candidates who excel in both ability and fit—combined with the high cost of review—can make the Agent’s "ability-only" selection more net-beneficial than the Principal's manual review.
Fairness: Strategic Humans vs. Naive Agents
The paper provides a critical perspective on algorithmic fairness by comparing "strategic" human decision-makers with "naive" agents. The authors examine two types of demographic disadvantage: bias (a lower mean signal) and noise (higher variance or unreliability in signals).
When a group is disadvantaged by a simple bias (a constant offset in scores), a human Principal can easily correct for it. However, when a group is disadvantaged by noise, a utility-maximizing Principal may engage in "strategic exclusion." Because the signals for this group are unreliable, the risk of a "bad hire" relative to the review cost becomes too high. Consequently, a rational, cost-conscious Principal might stop reviewing candidates from that group entirely to save on costs.
In these specific cases, a "naive" agent—one that applies a uniform threshold to all ability signals without considering the cost of review—may actually be the "lesser evil." While the agent does not correct for the noise, it does not "boycott" the group to save on costs, potentially leading to better representation than a strategic human.
Key Theoretical Results
- Review Cost Thresholds: The authors prove (Theorems 4.1 & 4.7) that if the cost of review is sufficiently low, the Principal will always prefer manual review. However, in scenarios with disparate signal variance, the optimal strategy for a non-delegating Principal often involves setting the review rate for the disadvantaged group to zero.
- Utility Curves: Under delegation, the Principal’s utility strictly decreases as the importance of fit ($\alpha$) increases. Under manual selection, the utility is U-shaped, indicating that the Principal struggles most when trying to balance ability and fit equally.
- Selectivity and Unfairness: Using a fairness metric ($\mathcal{D}$) to quantify deviations from demographic parity, the researchers show that as an Agent becomes more selective (higher thresholds), the gap between advantaged and disadvantaged groups consistently widens.
Limitations and Future Directions
The current model relies on Gaussian (Normal) distributions for applicant traits, which may not capture the "fat tails" or skewed distributions found in real-world data. Additionally, the model treats the choice as binary—either total delegation or total manual review. Future work could explore "human-in-the-loop" pipelines where an agent performs initial filtering followed by a human review of the top candidates. Finally, the authors note that this static model does not account for the long-term feedback loops where strategic exclusion might discourage future applicants from marginalized groups.
Summary updated