Applicants interviewed by an AI voice agent were more likely to receive an offer in a large customer-service hiring experiment in the Philippines. The offer rate was 9.73% for the assigned AI route, compared with 8.70% for assigned human interviews—a 1.03 percentage-point increase, or about 12% in relative terms. Human recruiters reviewed the evidence and decided whom to hire.[1]
Brian Jabarian and Luca Henkel studied recruitment through PSG Global Solutions, a Teleperformance subsidiary. Roughly 70,000 applicants entered human-interviewer, AI-interviewer or choice pathways. The main assigned-format comparison included 13,557 applicants in the human group and 40,103 in the AI group; the choice pathway was separate.[1]
Offers, starts and employment after one month
Outcomes among all applicants assigned to each format
| Outcome | Human interview | AI interview |
|---|---|---|
| Received a job offer | 8.70% | 9.73% |
| Started the job | 5.65% | 6.71% |
| Still employed one month after starting | 4.98% | 5.89% |
The higher offer rate was followed by more job starts and more applicants still employed one month after starting. For the subset of hires with performance data, the authors found no significant or economically meaningful differences in customer handling time, customer satisfaction or quality-assurance scores. That supports the reported employment result in this setting without establishing equivalent productivity in every job.[1]
Consistent questions with responsive follow-up
The authors describe a mechanism they call controlled variance. The voice agent asked questions in a more structured and consistent way across candidates while still responding to individual answers. The resulting interviews, analysed through transcripts, appeared to contain more information relevant to hiring decisions. A human recruiter could then compare candidates using records produced through a more uniform process.[1]
This matters because standardisation has two very different meanings. A rigid questionnaire might treat every applicant identically but fail to explore important experience. A highly improvisational interviewer might discover useful detail but introduce large differences between applicants. The AI system in the trial was intended to preserve relevant follow-up while reducing the incidental variation associated with who conducted the interview.[1]
The transcript analysis supports the authors’ interpretation that more consistent information collection contributed to the result. It does not isolate a single causal feature of the voice agent. Follow-up studies could test which forms of standardisation and adaptation preserve useful detail while reducing irrelevant interviewer variation.[1]
The setting and the research relationship
Random assignment strengthens the comparison within this recruitment process. It does not establish equal benefits across every demographic group, language or accessibility need. The results concern high-volume customer-service hiring in the Philippines, with people reviewing interviews and retaining authority over offers.[1]
Organisations contemplating this approach should also examine transparency. Does the candidate know an agent is conducting the interview? Can they request a human alternative when speech recognition, a disability or a technical failure creates a barrier? How is the recording retained, and who can review an answer that a recruiter misunderstood? An interview process that reduces inconsistency but makes mistakes difficult to challenge would trade one form of unfairness for another.[1]
The paper discloses that PSG supplied the data but did not control analysis or publication. After the prespecified data collection ended on 7 June 2025, Jabarian accepted an unpaid chief-economist role at PSG on 1 August 2025. That relationship is relevant when assessing the work and the value of replication with other employers.[1]
Improving the evidence before the decision
The experiment identifies an organisational use for voice AI: collect information more consistently before a person makes a consequential decision. Its actual offer and employment rates give other employers concrete outcomes to test, alongside candidate experience and accessibility. Replication would show how much of the benefit comes from this particular setting and how much transfers to other recruitment processes.[1]
Sources & further reading
- Voice AI in Firms: A Natural Field Experiment on Automated Job Interviews arXiv · 30 Jul 2026 · Accessed 26 Oct 2026