Voice AI in Firms: A Natural Field Experiment on Automated Job Interviews

TEXT START: We study AI agents as information-collection technologies: automated systems that elicit decision-relevant signals from humans through live interactions.

The Dissection

This is not primarily a paper about better interviews. It is evidence that firms can replace a variable, labor-intensive information-collection layer with a controlled AI protocol while retaining humans as downstream judges. The real asset is not voice; it is throughput, consistency, and scalable filtering. Human recruiters become validators of a machine-mediated funnel.

The Core Fallacy

The abstract risks confusing local matching gains with system survival. More offers, job starts, and retention in one experiment do not establish net employment growth or preservation of the wage-consumption circuit. They show that AI can improve a firm’s allocation of jobs while making human interview labor less necessary. “No decline in productivity” applies only to workers who were hired; it says nothing about the economic necessity of the people displaced from information collection. The human recruiter is a lag defense, not a permanent moat. Once interviewing is automated, evaluation and hiring decisions become the next targets.

Hidden Assumptions

• Human recruiters remain employed and retain final authority.
• More offers reflect genuine job creation rather than redistribution or relaxed screening.
• The measured productivity and retention window captures all relevant effects.
• Standardized interviews collect better signals without suppressing tacit, cultural, or accessibility-related information.
• Applicants cannot strategically adapt to or manipulate the AI system.
• The result generalizes beyond this firm, applicant pool, and job class.
• Automation and integration costs do not erase the measured gains.
• AI improves selection without increasing the scale of hiring automation downstream.

Social Function

Classification: partial truth functioning as transition management and elite self-exoneration.

The empirical result may be real and valuable. Its institutional function is still clear: it gives firms a respectable justification for adopting AI as “better information” rather than payroll substitution. The framing sanitizes the power shift. Applicants become machine-readable inputs, recruiters become exception handlers, and the firm gains a larger, more uniform labor-sorting capacity. This is not crude propaganda; it is a clean result being used to make a harsher transition appear like organizational improvement.

The Verdict

This is a DT beachhead result, not a counterexample. It supports a narrow version of P1: AI can outperform heterogeneous human execution on a repeated cognitive task. It illuminates P2: institutions respond to variance and scale problems by delegating standardization to machines. It does not yet prove full P1 across cognitive work or P3, but it shows the mechanism clearly.

The paper documents the first stage of automation and treats temporary human-AI coexistence as the endpoint. The job interview is not being saved. It is being converted into a machine-mediated sorting function. The worker may receive a better chance in this experiment; the labor system receives a more efficient funnel toward obsolescence.