At a glance
- Company
- PSG Global SolutionsRecruitment outsourcing firm, part of Teleperformance
- Kind of work
- Hiring and HR
- Where and when
- Philippines. Study completed June 2025.
- Who produced the figures
- Measured by researchersThe figures come from a working paper by university researchers. It has not yet been peer-reviewed. The firm supplied the data. After the data were collected, one author took an unpaid role as the firm’s chief economist.
See alsoSiemens: inspectors caught more defects when the AI showed what it had seen(Company case) · “Copilot saves UK civil servants 26 minutes a day”(Claim check)
What they did
The firm recruits customer-service staff in the Philippines for large client companies. Applicants who passed a first screening were assigned at random to one of three groups. One group was interviewed by a human recruiter. One was interviewed by an AI voice agent. The third could choose.
In every group a human recruiter then reviewed the interview and decided whether to make an offer.
The numbers
- 70,884 applications in the study.
- 12% more likely to receive a job offer when the AI ran the interview.
- More likely to still be in the job: by 19% after two months, 18% after three and 19% after four.
- No drop in the measured productivity of the people hired.
- 78% of applicants who could choose picked the AI interviewer.
- 131 recruiters evaluated the applicants.
Why the researchers think it worked
The AI’s interviews were more consistent. It covered the planned topics in a steadier way than the human interviewers did, while still responding to each applicant. That gave the recruiters more of the information they needed to decide.
Applicants rated the AI interviews as less natural.
What the case shows
- The AI did the gathering. People did the judging. A decision that changes someone’s working life stayed with a person.
- Consistency was the gain. The benefit came from every interview covering the same ground.
- It was measured properly. Random assignment means the two groups can be compared.
- One firm, one kind of job. These were entry-level customer-service roles recruited in high volume.
- Not every applicant reached the interview. The study covers applicants who had passed a first screening. People screened out before that are not in it. In the United States, a court has let applicants over 40 bring a joint claim against Workday over its AI screening of applicants, in a preliminary ruling. See the court order.
What this says about the practices
Practices the company applied
- Keep a person checking anything that goes to a customer or carries legal weight The AI only ran the interview. In the paper’s words, “all hiring decisions are made by human recruiters”. Do first · Effort: medium (weeks, a few people)
- For each kind of task, test whether a person with AI beats either one alone The firm split the job in two: the AI gathers the information, and a person judges it. Do next · Effort: medium (weeks, a few people)
- Measure time saved per task against a baseline Applicants were assigned at random to an AI or a human interviewer, so the two groups could be compared directly. Do first · Effort: low (days, one person)
- Set the limits before the tool acts The AI voice agent could run the interview and nothing more. In the paper’s words, “all hiring decisions are made by human recruiters”. Do next · Effort: low (days, one person)
What to take from it
If someone cites PSG Global Solutions to argue that AI will pay off for you, the case supports a narrower point. The result was measured in one organization, on one kind of work. Treat the result as a reason to measure your own pilot the same way, not as a forecast of what you will get. A practice applied here, to check in your own plans: Keep a person checking anything that goes to a customer or carries legal weight.
Read next
- Siemens: inspectors caught more defects when the AI showed what it had seenResearchers gave 48 inspectors at a Siemens electronics factory an AI aid for checking circuit boards. Half saw only the AI’s score. Half also saw a picture marking where the AI had found a problem. The second group caught more defects, and most of them did better than the AI working alone.
- “Copilot saves UK civil servants 26 minutes a day”The figure is in the UK government’s report, but it is what users estimated, not time measured. It comes from one survey question with answers in ranges. HMRC, with licences given at random, put the estimate at about an hour a week. The Department for Business and Trade found no evidence that time saved led to higher productivity.
- Ashurst: first drafts 45–80% faster in timed trials, with lawyers still ahead on accuracyThe law firm Ashurst tested generative AI with 411 of its people for five months and published the results. First drafts came 45% to 80% faster in timed tests. In a blind comparison, lawyers’ own drafts were more accurate on average, and one AI draft cited a wrong verdict.
Sources
- Jabarian and Henkel (2026). Voice AI in Firms: A Natural Field Experiment on Automated Job Interviews. Working paper, arXiv 2607.28222. arxiv.org
- Vaccaro, Almaatouq and Malone (2024). When combinations of humans and AI are useful: A systematic review and meta-analysis. Nature Human Behaviour, 8(12), 2293–2303. What this study did and found