AI job interviews were quicker to conduct and applicants assigned to them were 12 per cent more likely to receive an offer in a large field experiment. The result is striking because the artificial intelligence did not make the hiring decision. It conducted the conversation, collected information and passed the evidence to human recruiters, who retained control over every offer.
The working paper by Brian Jabarian and Luca Henkel examined 70,884 applications for entry-level customer-service positions in the Philippines. Of those, 67,056 eligible applications were randomised between a human interviewer, an AI voice agent or a condition in which applicants could choose. The experiment covered 48 job postings processed through recruitment outsourcing firm PSG Global Solutions.
Among applicants assigned to a human interviewer, 8.70 per cent received an offer. The rate rose to 9.73 per cent among those assigned to the AI interviewer. That is an increase of 1.03 percentage points, or 12 per cent relative to the human-interview rate. The distinction matters because a 12 per cent relative improvement does not mean that 12 extra offers were made for every 100 applicants.
How the AI Job Interviews Worked
The AI voice agent followed the same structured guidance given to human recruiters. Interviews covered matters such as location, working hours, previous experience, education, salary expectations and willingness to accept the conditions of the role. The system could adapt its follow-up questions while keeping the conversation within that framework.
Applicants were told at the beginning that they were speaking to an AI. They were also informed that a human recruiter would review the interview and make the decision. That separation is central to the finding. The experiment automated information collection, not the final judgement about who should be employed.
After the interview, applicants completed standardised language and analytical tests. Recruiters then reviewed the audio, transcript and test results before deciding whether an applicant met the threshold for an offer. When a human conducted the interview, that recruiter also performed the evaluation. When the AI conducted it, an assigned human recruiter reviewed the evidence afterwards.
The researchers argue that the AI produced what they call controlled variance. Human interviewers were given guidance but differed in which topics they covered, the order they followed and the way they asked questions. The AI was more consistent while remaining responsive to what each applicant said.
Faster Interviews Did Not Mean a Faster Hiring Process
The researchers’ public project summary describes the AI interviews as four times faster. The paper reports a median AI call duration of 9.6 minutes, while full human interviews generally took between 10 and 20 minutes. Automation also allowed applicants to arrange conversations without waiting for a recruiter’s diary.
However, the full recruitment journey was not faster. Among successful remote applicants, the median interval from expressing interest to interview was 0.32 days with AI, compared with 0.51 days with a person. The next stage reversed that advantage. Human recruiters took a median of 7.24 days to make an offer decision after an AI interview, compared with 2.62 days when they had conducted the interview themselves.
Overall, successful applicants in the AI condition took a median of 24 days to move from initial interest to starting work, compared with 20 days in the human condition. The researchers conclude that automation moved the bottleneck. It accelerated access to the interview, but reviewers needed longer afterwards because they had not participated in the conversation.
This qualification does not invalidate the selected headline. The interviews themselves were faster and produced more offers. It does show why organisations should measure the whole process rather than assuming that a quicker automated stage creates a quicker outcome.
More Offers Were Followed by Better Retention
A higher offer rate would be much less impressive if it simply admitted weaker candidates. The researchers therefore followed job starts, retention and a subset of on-the-job performance measures. Across all applicants assigned to either interview condition, the AI group was 18 per cent more likely to start work and 18 per cent more likely to remain employed for at least one month.
The retention advantage continued. Applicants in the AI-interview group were 17 per cent more likely to remain after two months, 16 per cent more likely after three months and 17 per cent more likely after four months. These are relative differences across the original treatment groups, not the proportion of hired workers retained.
Among workers for whom performance data was available, the researchers found no statistically significant or economically meaningful differences in average customer-handling time, customer satisfaction or employer quality-assurance scores. Voluntary and involuntary reasons for leaving also did not differ significantly between the two interview conditions.
Retention is an imperfect measure of job quality, particularly in a high-turnover industry. In this setting it was the firm’s main proxy for whether the worker and role matched, and clients rewarded the recruiter for hires who remained at least one month. The result supports the authors’ conclusion within that market, but it should not be mistaken for a complete assessment of a worker’s long-term career success.
Structure May Have Produced Better Information
Transcript analysis showed that AI interviews adhered more closely to the expected topic order and covered a more consistent number of relevant subjects. The AI covered 45 per cent of guideline topics on average, compared with 38 per cent for human interviewers, and showed stronger adherence to the recommended sequence.
The interviews were not simply identical scripts. The AI varied wording and follow-up questions in response to the applicant. The researchers found more of the conversational features that predicted offers in human-led interviews and fewer features associated with rejection. That suggests structure made the evidence easier to compare without removing every individual response.
Recruiters gave higher interview scores to AI-interviewed applicants, mainly by moving more people from low to medium ratings. They also placed relatively more weight on standardised language scores when reviewing an AI interview. The process changed both the information collected and how the human decision-maker used it.
This provides a useful contrast with LiveAIWire’s broader coverage of AI recruitment tools and opaque automated screening. In the field experiment, the AI did not silently reject applications or generate an unexplained final ranking. It conducted a disclosed interview inside a defined process and left the consequential decision with a named human role.
Applicants Often Preferred the AI Interviewer
In the choice condition, 78.41 per cent of applicants who selected an interviewer chose the AI voice agent. Preference was higher among remote applicants than walk-in applicants. Convenience may explain part of the result because the automated option could be scheduled at the applicant’s preferred time.
Survey responses did not reveal a broad backlash. Offer acceptance and the firm’s Net Promoter Score were similar across conditions. Applicants rated stress, comfort, follow-up fluency and feedback quality similarly, although they considered the AI interaction less natural.
Not every experience was positive. Five per cent of applicants ended the interview because they did not want to speak with an AI, and technical difficulties affected seven per cent of AI interviews. Those failure rates matter at scale. A system that works well on average still needs a human route for people who cannot or do not wish to complete an automated call.
The study also reported that perceived gender discrimination was lower after AI interviews, at 3.30 per cent compared with 5.98 per cent after human interviews. That was a self-reported experience measure, not proof that discrimination had been removed. The researchers found gender differences in offer rates in both conditions, and the design was not a complete audit of every potential source of bias.
What Employers Should Learn From the Trial
The strongest lesson is not that every employer should replace interviewers with voice AI. It is that consistency in information collection can improve a human decision. Organisations could apply that principle by tightening interview structure, whether the conversation is conducted by a person or a machine.
An employer considering automation should measure offer quality, applicant withdrawal, technical failure, reviewer time, job starts, retention and actual performance. It should also keep an accessible human alternative and explain which system records the conversation, how the data is used and who remains accountable for the outcome.
The longer evaluation time in the AI condition is a warning against counting only recruiter minutes saved during the call. Reviewing a conversation conducted by someone else can require more effort. A deployment may reduce interview labour while creating a transcript-review queue, so staffing and workflow design must change together.
The experiment also supports a wider point about AI exposure and the transformation of work. Automation did not remove human recruiters from the process. It moved them away from conducting every initial conversation and towards evaluating standardised evidence. Whether that is better work depends on workload, accountability, judgement quality and the applicant’s ability to challenge a mistake.
Why the Finding Cannot Be Applied to Every Job
This is a working paper submitted to arXiv on 30 July 2026, not a peer-reviewed journal article. The experiment ran from March to June 2025 in the Philippines and focused on entry-level customer-service roles. These were high-volume vacancies with structured requirements, standard tests and measurable early retention.
An executive appointment, specialist engineering post, creative role or job dependent on complex interpersonal judgement presents a different problem. The study does not show that an AI interviewer would collect better evidence in those settings. Nor does it test whether candidates behave differently when the employer, language, labour market or legal context changes.
The partner firm provided the data and operational support. The paper states that it had no role in the analysis, manuscript preparation or publication decision. It also discloses that Jabarian accepted an unpaid chief economist role with PSG Global Solutions almost two months after the preregistered data collection ended, extending the research partnership. That transparency allows readers to consider the relationship when assessing the results.
The experiment is valuable precisely because it separates a popular fear from what was actually tested. An AI did not secretly choose who was hired. A disclosed voice agent conducted a more standardised first-stage conversation, after which people made the decisions. In that setting, interviews were faster to run and more applicants received offers without a measured decline in early worker outcomes.
For applicants, the practical message is neither to fear nor trust an automated interview automatically. Ask how the recording will be judged, prepare for the same role-specific questions you would expect from a person, and request a human alternative when the system fails or creates an accessibility problem. The interviewer may be artificial, but the evidence it collects can still influence a very human decision.
About the Author
Stuart Kerr is Technology Correspondent at LiveAIWire, covering artificial intelligence, cybersecurity and the social impact of emerging technology. LiveAIWire is an independent, human-led technology publication using AI-assisted research, editorial production and original AI-assisted editorial illustrations under his direction.