Algorithmic Gatekeeping

When should routing authority in tiered service systems transfer from human gatekeepers to algorithms? The answer hinges on an inseparable bundle. A human gatekeeper who interacts with customers can elicit contextual cues outside the algorithmic data pipeline, but the same unverifiability that makes these cues inaccessible to algorithms makes them impossible to fully contract on, exposing the system to agency distortion. Transferring authority to an algorithm eliminates both. The regime question is therefore whether to accept or reject this bundle. The authors formalize this tradeoff by decomposing customer complexity into hard information (observable by both humans and algorithms) and soft information (residual information that remain with frontline humans). They derive exact cost representations for each regime and a local frontier approximation: algorithmic routing dominates when |a| > c(v) √v, where a measures agency distortion, v measures soft information variance, and c(v) captures how the density of marginal cases differs between regimes. As soft information vanishes, the frontier simplifies to the distribution-free condition |a| > √v. The authors further show that increasing soft information variance can simultaneously raise algorithmic errors and reduce agency costs, a finding that helps explain why human gatekeeping persists in high-uncertainty environments. Hybrid pre-screening exhibits a sharp asymmetry: it mitigates under-referral but not over-referral, reflecting structural limits of partial unbundling. Extensions to heterogeneous gatekeepers, imperfect information processing, signaldependent variance, and human-contact costs show the frontier generalizes with modified parameters. The authors' analysis reframes routing technology adoption from “how accurate is the algorithm?” to “who should hold routing authority?”