The W3C's new note on user agents says that software standing between a person and the web owes its loyalty to that person. A job-interview bot makes no such promise. The employer buys it, configures it and writes its questions. The applicant on the call is talking to the firm's representative, and the reasonable fear is that whatever the bot learns, it learns for the firm, at the applicant's expense.
A large field experiment complicates that fear, at least in one setting. Brian Jabarian and Luca Henkel worked with PSG Global Solutions, which recruits customer-service staff in the Philippines for the call-center company Teleperformance. Between March and June 2025, about 67,000 eligible applications were randomly assigned to a human recruiter, an AI voice agent, or the applicant's own choice between the two. The agent conducted spoken interviews of 10 to 20 minutes and asked follow-up questions based on each answer. Humans made every hiring decision. The study is a working paper, so independent reviewers have not yet checked it.
Applicants interviewed by the AI received offers more often (9.73 percent against 8.70), started work more often (6.71 against 5.65) and were more likely to still be employed a month later (5.89 against 4.98). Only 14 percent answered a follow-up survey, but those who did rated the AI interview about as fair as the human one, if less natural. In the group allowed to choose, roughly 78 percent of those who responded picked the AI, which offered flexible scheduling.
The likely explanation lies in how the work was divided. Recruiters in both groups had recordings, transcripts and standardized tests. What differed was who did the talking. In the human group, the recruiter who held the conversation also scored it. In the AI group, recruiters scored a conversation that had covered the same ground with every candidate while still following up on individual answers. Their offers tracked language-test results more closely and interview scores less. The authors credit that consistency. Another way to read it is that the agent separated the conversation from the judgment, and gave whoever was judging comparable evidence about every candidate. Auditors work on the same principle, that whoever does the work should not be the only one to grade it.
So whose side was the bot on? Here, plausibly both, because the two sides wanted overlapping things. The firm needed people who would start and stay. Applicants wanted offers, and jobs they would keep. Comparable evidence moved both outcomes in the same direction.
That overlap owes more to the job than to the software. The requirements of customer service can be written down in advance, and the authors confine their claim to qualities of that kind. In roles where interviewers' differing judgments pick up something real, standardizing may discard it. Here it seems to have cost little because the job was narrow to begin with, a smaller claim than the headline numbers suggest.
There are other limits: one firm, one role, no peer review yet. Jabarian became the firm's unpaid chief economist after the data were collected. And where the authors could measure job performance, they found no significant difference between the groups.
What carries over to other settings is a two-part test for any go-between a firm deploys. What does it change about what the decision-maker learns? And would the person on the other side, knowing that, choose it? In this study the researchers came close to asking the second question. Given the option, most applicants who answered picked the AI, though no one told them how it would change what recruiters saw. Most deployments are arranged so that the question never comes up. The firm is the buyer, the reports track the firm's outcomes, and the applicant's preference appears nowhere unless someone decides to measure it.