Do AI First Interviews Work? What the Research Says (2026)
Two findings matter most. Structured interviews, where every candidate gets the same questions and is scored on the same scale, are among the best predictors of job performance. And in one large field experiment at a BPO in the Philippines, applicants interviewed by an AI voice agent were more likely to get an offer, start the job and still be in it a month later, while people made every hiring decision. These are other researchers’ results, not Gahiz results.
Structured interviews predict job performance
Researchers measure this with "validity": how closely scores from a hiring method line up with how people later perform in the job. 0 means no link and 1 means a perfect link. For example, a validity of .42 means interview scores and later performance clearly move together, but far from perfectly.
| Study | Structured interviews | Unstructured interviews | Based on | Status |
|---|---|---|---|---|
| Sackett, Zhang, Berry & Lievens (2022) | Mean validity .42, the top-ranked selection procedure in the review | Mean validity .19 | Structured: 105 studies, 7,864 people. Unstructured: 32 studies, 3,125 people | Peer-reviewed meta-analysis |
| Schmidt & Hunter (1998) | Mean validity .51 | Mean validity .38 | Earlier meta-analytic data (McDaniel et al., 1994) | Peer-reviewed. The older estimate, which Sackett et al. say was inflated |
| Huffcutt & Arthur (1994) | Structure is a major moderator of interview validity | Not reported separately in the abstract | 114 validity coefficients for entry-level jobs | Peer-reviewed. Abstract only |
Sackett and colleagues also found smaller group score differences for structured interviews (d = .23) than for unstructured interviews (.32) or cognitive ability tests (.79), using US Black-White data. That means smaller group differences, not an absence of bias.
Two cautions. Validity varies a lot between studies (a standard deviation of .19 for structured interviews), and Oh, Le and Roth (2023) dispute the method behind the newer, lower estimates; the authors replied. And these studies predict job performance, not how long people stay.
The field experiment on AI voice interviews
Brian Jabarian and Luca Henkel ran a pre-registered field experiment (AEA RCT Registry #15385) with PSG Global Solutions, a Teleperformance subsidiary hiring entry-level customer service staff in the Philippines. Out of 70,884 applications, 67,056 applicants were randomly assigned: 40,103 to an interview with an AI voice agent, 13,557 to a human recruiter, and 13,396 could choose. 131 recruiters took part, and all hiring decisions were made by human recruiters.
12% more likely to get a job offer
Offers went from 8.70% of applicants with a human interviewer to 9.73% with the AI interviewer: 1.03 percentage points, or 12% more in relative terms.
18% more likely to start the job
Job starts went from 5.65% to 6.71% across all randomised applicants.
18% more likely to still be employed after one month
One-month retention went from 4.98% to 5.89%. After two, three and four months the relative gap was 19%, 18% and 19%. Among people who accepted an offer, the gain was smaller: 6% at one month.
No difference in on-the-job performance
The authors found neither statistically significant nor economically meaningful differences in time spent handling customers, customer satisfaction scores or quality assurance scores. This is a null result, not an improvement.
78% chose the AI interviewer when given the choice
In the group that could choose, 78% picked the AI voice agent. On average, those who chose it had lower language and analytical test scores.
A note on the numbers: "12% more" is a relative change. In absolute terms it is about one more offer for every 100 applicants (1.03 percentage points).
The authors also report that the AI interviews were more structured and consistent, while still adapting to each applicant.
What applicants said
- Applicants rated the AI interview as less natural (p = 0.014) and rated its fairness about the same as a human interview (no significant difference, p = 0.68). Only 2,764 of 19,200 applicants answered the survey (14%), so read this with care.
- In the 2025 version of the paper, 5% of applicants in AI interviews stopped because they did not want to talk to an AI, and 7% of AI interviews had technical problems.
- A lab study points the other way on acceptance. Langer, König and Papathanasiou (2019) found that highly automated interviews lowered overall acceptance, through lower social presence and fairness. It used video vignettes with 123 participants, not real applicants.
Speed matters in volume hiring
Context from the US, not from Egypt or the Gulf: the SHRM Talent Access Report (2022, covering 2021) found a median time-to-fill of 44 days for non-executive roles, and an average of 54 days, across 840 organisations. Time-to-fill runs from the day the job requisition opens to the day the offer is accepted.
A first interview that applicants can take as soon as they apply does not fix the whole process, but it removes the wait for a recruiter’s calendar at the very first step.
What AI interviews don’t do
- They don’t make the hiring decision. In the field experiment, people made every decision, and that is how Gahiz works too.
- They don’t replace a certified language exam. An estimated language level is a screening signal, not a certificate.
- They don’t need to score CVs. Gahiz uses a CV, when the employer shares one, only to tailor questions; the CV never produces a score.
- Speech recognition can be less accurate for some accents. Koenecke and colleagues (2020) measured an average word error rate of 0.35 for Black speakers and 0.19 for white speakers, in US English recordings of 73 Black and 42 white speakers; Arabic dialects were not tested. Listen to the recording for anyone close to your cut-off.
- One study is one firm in one country. Results for your roles can differ.
- They don’t remove the need to check fairness in your own hiring results.
Caveats on the evidence
- The field experiment is a working paper (arXiv, version 2, September 2026). It has not been peer-reviewed.
- It covers one firm, in the Philippines, hiring for English-language, entry-level customer service roles.
- Recruiters knew which interviews the AI had run.
- Disclosed in the paper: on 1 August 2025, after data collection ended on 7 June 2025, Brian Jabarian accepted an unpaid role as Chief Economist at PSG Global Solutions.
- Validity estimates for interviews vary widely between studies, and researchers still debate how to correct them.
How Gahiz uses this
We built Gahiz around the two things this research supports: the same structured questions for everyone, scored against the same rubric, and people making the decision. The numbers on this page are other researchers’ results, not Gahiz results.
Common questions
- Do AI interviews work?
- In one large field experiment (67,056 randomised applicants at a BPO in the Philippines), applicants interviewed by an AI voice agent were 12% more likely to get an offer and 18% more likely to start the job and to stay a month, with people making every hiring decision. It is one study in one firm, so test on your own roles.
- Are AI interviews fair to candidates?
- In that experiment, applicants rated the AI interview’s fairness about the same as a human interview, but less natural, and few answered the survey. Structured interviews in general show smaller group score differences than unstructured interviews or cognitive tests. Always check fairness in your own results.
- Is an AI interview the same as a certified English test?
- No. Gahiz shows an estimated CEFR range based on the CEFR Companion Volume descriptors, with the answers behind it. It is not a certified CEFR result.
- Are these Gahiz results?
- No. Every number on this page comes from published research by other people, listed in the sources below.
Interview question banks for these roles
More hiring guides
- How to Spot AI-Written CVs, and What to Do Instead
- High-Volume Hiring in Egypt: How to Screen 500 Applicants for One Role
- Applicant Screening: A Practical Guide to Phone Screening Interviews
- Phone Screening Interview Questions for Customer Service and Sales Roles
- English Level Test for Call Centre Hiring: A Practical Guide
- AI Interview Tools Compared: What to Look For
- AI Interview Tools in Egypt and the Gulf: A 2026 Comparison
- Egypt’s Data Protection Law and Recorded Candidate Interviews
Sources
- Jabarian, B. & Henkel, L. (2026). Voice AI in Firms: A Natural Field Experiment on Automated Job Interviews. arXiv 2607.28222, version 2 (working paper).
- Jabarian, B. & Henkel, L. (2025). Earlier version, SSRN 5395709. doi:10.2139/ssrn.5395709.
- Sackett, P. R., Zhang, C., Berry, C. M. & Lievens, F. (2022). Revisiting meta-analytic estimates of validity in personnel selection. Journal of Applied Psychology, 107(11), 2040-2068.
- Schmidt, F. L. & Hunter, J. E. (1998). The validity and utility of selection methods in personnel psychology. Psychological Bulletin, 124(2), 262-274.
- Huffcutt, A. I. & Arthur, W. (1994). Hunter and Hunter (1984) revisited: Interview validity for entry-level jobs. Journal of Applied Psychology, 79(2), 184-190.
- Langer, M., König, C. J. & Papathanasiou, M. (2019). Highly automated job interviews: Acceptance under the influence of stakes. International Journal of Selection and Assessment, 27(3), 217-234.
- Koenecke, A. et al. (2020). Racial disparities in automated speech recognition. PNAS, 117(14), 7684-7689.
- SHRM (2022). Talent Access Report (2021 data).