Short answer: Do AI interviews work? The best field evidence says they can, in a narrow sense. In one large experiment, applicants interviewed by an AI voice interviewer received more offers, started more often and were more often still employed a month later, while recruiters made every decision. Another found recorded interviews lost far more applicants than live ones.
Key takeaways
- In a randomised experiment with 67,056 applications for customer service jobs in the Philippines, AI voice interviews produced about one extra offer and one extra start per 100 applicants, with recruiters making every decision (Jabarian and Henkel, version 2, 11 September 2026).
- In a US experiment with 3,296 tech applicants, about 40% completed a recorded interview with nobody present, against about 85% who only confirmed interest and 68% who joined a live video interview. Women dropped out of the recorded format more often (Avery et al., version of 25 March 2026).
- Both are working papers, not peer reviewed. Neither tested Dutch frontline roles or any specific product.
- Interview format changes who stays. Pilot AI interviews next to your current route and measure completion by step, early retention by route and drop out by group.
- Keep a route to a person for applicants who do not want an AI interview.
On this page: What did the largest field experiment of AI interviews find? · Do candidates drop out of AI or recorded interviews? · How do applicants feel about AI interviews? · What does this research not show? · How should you pilot AI interviews for frontline roles? · So, do AI interviews work? · How Radius helps · Frequently asked questions
Most pages that answer this question come from companies selling the tools and rely on opinion surveys, not outcomes. This article looks at the two field experiments that followed real applicants: what they found, and what they did not test.
What did the largest field experiment of AI interviews find?
AI voice interviews led to more offers and starts, while recruiters decided.
The study is by Brian Jabarian and Luca Henkel. We used version 2 of their working paper, posted on arXiv on 11 September 2026 and still the latest on 21 October 2026. Its abstract and most coverage round the sample to "70,000 applicants"; the counts below come from the paper itself.
Between March and June 2025, a recruitment process outsourcing firm hiring in the Philippines received 70,884 applications for entry level customer service jobs. Of these, 67,056 were eligible and were randomly assigned to one of three routes: an interview with a human recruiter (13,557 applications), an interview with an AI voice interviewer (40,103), or a choice between the two (13,396). The AI interviewer held a spoken conversation and asked follow up questions. In every route, a human recruiter reviewed the interview, using the transcript, the recording and test scores, and made the hiring decision.
Source: Jabarian and Henkel (2026), arXiv 2607.28222, version 2. Philippines, entry level customer service roles, applications received March to June 2025.
Three further results matter to a buyer:
- No loss in performance. For the hires whose performance data were available, the paper finds no significant differences between the routes in call handling time, quality assurance scores or customer satisfaction.
- Recruiters did not expect it. Asked to forecast the result, only 15% of the firm's own recruiters expected more offers after an AI interview; 36% expected fewer and 49% no difference.
- More consistent interviews. The authors attribute the gain to the AI asking a more consistent set of questions while still following up on answers, which drew out more relevant information per interview.
What the firm's own recruiters expected
Our view: read the absolute numbers. Coverage, following the paper's abstract, says AI interviews raised offers by 12%. That is a relative change; the absolute gain is about one extra offer and one extra start per 100 applicants. On our own arithmetic from the paper's counts, the share of starters still employed a month later was almost the same on both routes: about 88% (2,363 of 2,691 after an AI interview, 675 of 766 after a human one). The authors report that the AI route's lead in employment held at two, three and four months. So the gain came mainly from getting more people to a first shift; once started, they stayed at about the same rate. In high volume hiring, where the hard part is often getting enough suitable people to a first shift, that matters.
Who supplied the data, and what was disclosed? The paper states that the partner firm supplied the data but had no role in the analysis, the writing or the decision to publish. It also discloses that on 1 August 2025, after data collection, the lead author accepted an unpaid role as chief economist at that firm, PSG Global Solutions, a Teleperformance subsidiary. The paper does not name the company behind the AI interviewer. None of this undoes the random assignment, the study's main strength, but it is a reason to wait for peer review and an independent replication.
Do candidates drop out of AI or recorded interviews?
Some do, and recorded interviews with nobody present lost the most.
The second experiment, by Mallory Avery, Edwin Ip, Andreas Leibbrandt and Joseph Vecci, was registered in advance (AEARCTR-0013356) and published as a University of Exeter discussion paper, in a version dated 25 March 2026. In 2024, 3,296 people living in the United States applied for programmer, web designer and content creator jobs on US job boards. They were randomly asked to do one of three things next:
- confirm their interest with one click (667 applicants): about 85% did;
- complete a recorded audio or video interview, answering set questions with no interviewer present (2,535 applicants): about 40% did;
- join a live online video interview with a person (94 applicants, so a less precise figure): 68% did.
Who completed the next step
3,296 US applicants for tech jobs, given one of three next steps
Men and women continued at the same rate when they only had to confirm interest. In the recorded interview route, men completed 5.1 percentage points more often than women. In a separate survey experiment, people invited to a recorded interview believed more applicants had been invited and saw the process as less fair, and both beliefs predicted a lower chance of completing it. Comfort with technology did not. The study was funded by the Australian Research Council, and the paper does not name the interview platform.
The first experiment adds two numbers on friction. In 7% of AI interviews, the interview stopped because of a technical failure of the AI interviewer, and in 5% the applicant said they did not want to continue speaking with the AI (Jabarian and Henkel, 2026).
"AI interview" covers formats that work very differently for the applicant, which is one reason the headlines conflict:
| Format | What the applicant experiences | What the field evidence shows |
|---|---|---|
| Live AI voice interview | A spoken conversation in which the AI asks questions and follows up | More offers and starts than a human interview in one experiment; 7% technical failures and 5% of applicants unwilling to continue (Jabarian and Henkel) |
| Recorded one way interview | Answering set questions on camera or microphone; nobody responds | About 40% completed, against about 85% who only confirmed interest; larger drop for women (Avery et al.) |
| Live interview with a person | A video or phone conversation with a recruiter | 68% joined a live video interview (Avery et al., small group); the comparison route in Jabarian and Henkel |
Our view: format and speed change who stays in a process, and the effect can go either way depending on the format, the job and the people applying. Evidence about a recorded interview should not be used to judge a live conversation with an AI interviewer, or the other way round.
To see what a structured voice interview looks like in practice, explore the AI Interviewer.
How do applicants feel about AI interviews?
Field and laboratory studies only partly agree, so treat reactions as an open question.
In the field experiment, applicants offered a choice picked the AI interviewer 78.41% of the time. That is not a quality signal: on average, those who chose the AI had lower language and analytical test scores. Applicants rated AI and human interviews similarly for fairness, but rated AI interviews as significantly less natural (Jabarian and Henkel, 2026).
Laboratory studies point the other way on some measures:
- Among 123 participants who watched videos of interviews, highly automated interviews were rated lower on social presence and fairness than videoconference interviews, which reduced acceptance (Langer, König and Papathanasiou, 2019).
- Among 148 participants, highly automated interviews were seen as more consistent but as conveying less social presence (Langer et al., 2020).
- Across two studies, AI interviews were seen as less fair in procedure and in how people are treated than interviews with a person, and two way communication played a particular part in how participants reacted (Acikgoz et al., 2020).
Most laboratory participants judged a process they watched or imagined, not one they went through to get a job. Our view: both lines of evidence support the same safeguards. Tell applicants clearly that the interviewer is automated, explain what happens to their answers and who decides, and offer a person to anyone who asks.
What does this research not show?
It does not show results for Dutch frontline roles or any product.
Before you take either study to a board or a works council, be clear about its limits:
- One setting each. One firm, one country and one type of job per study: customer service in the Philippines, and tech roles in the United States. Neither covers warehouse, logistics, cleaning, security or hospitality work in the Netherlands.
- Working papers. Neither paper had been peer reviewed when we checked on 21 October 2026, and figures can change between versions.
- Fairness across groups was barely tested. In the version we read, Jabarian and Henkel report no outcomes by group. Avery et al. found the gender gap in recorded interviews, and that one commercial AI tool rated transcripts of the answers from women and underrepresented minorities higher than human evaluators did. Neither study tested outcomes by first language, nationality or disability.
- Short time horizon. Retention is reported up to four months after starting, and job performance only for some of the hires, in one firm.
- How well it predicts. In a 2023 review of hiring methods, Sackett and colleagues said they looked forward to a time when a body of validity evidence would accumulate for newer tools such as asynchronous video interviews and automated scoring (Sackett et al., 2023). The two experiments above are early steps, not that body of evidence.
- No product. Neither study names the technology provider. No study of Radius Hire's own interviews has been published, so neither experiment tells you how Radius, or any other product, performs for your roles.
How should you pilot AI interviews for frontline roles?
Run both routes side by side and measure who gets through.
The studies offer a method more than an answer: compare routes on the same applicant flow and count what happens next.
- Choose one or two roles with a steady flow of applicants, and write the requirements and questions first. Asking everyone the same job related questions is what makes interviews comparable (structured and unstructured interviews compared).
- Split the applicant flow. Random assignment is best; alternating weeks is simpler but weaker. Keep the advert, pay and shifts the same on both routes.
- Tell candidates at the start that the interviewer is automated, what happens to their answers, and how to reach a person instead.
- Fix your measures and review dates before you start, and review again one month after the first starts.
- Keep people in the decision. A recruiter reviews every interview and records a reason whenever they reach a different conclusion from the interview summary.
| Measure | How to calculate it | Why it matters |
|---|---|---|
| Completion by step | Applicants who finish each step, divided by those invited to it, per route | Shows where a format loses people, as the recorded interview did in Avery et al. |
| Opt outs and failures | Share who ask for a person; share of interviews that fail technically | Both occurred in the first experiment |
| Offers and starts per 100 applicants | Offers and first shifts, divided by all applicants assigned to the route | The outcome the largest experiment measured |
| Early retention by route | Still employed after four weeks, as a share of starters and of all applicants | Separates "more people start" from "people stay longer" |
| Drop out by group, where lawful | Completion by gender, age band or interview language; agree with your privacy officer what you may record | Avery et al. found a larger drop for women in recorded interviews |
| Recruiter review time | Minutes spent reviewing each interview | Shows whether time saved on calls returns as review work |
| Override rate | Share of candidates where the recruiter's conclusion differed from the interview summary, with a reason | A sign that people are reviewing rather than rubber stamping |
Illustrative example (invented figures). An agency hiring order pickers for one client site sends half its applicants to an AI voice interview and half to its usual phone screen for six weeks. Per 100 applicants, the AI route produces 9 first shifts and the phone route 7. Four weeks later, 8 of the 9 and 6 of the 7 are still working, roughly the same share of starters. The honest summary is "the AI route got more people to a first shift", not "it found people who stay". Only the two denominators together tell those apart.
The legal basics for a pilot in the Netherlands:
- Since 2 August 2026, the EU AI Act has required providers to design AI systems that talk directly with people so that those people are told they are dealing with AI, unless that is obvious from the context (Regulation (EU) 2024/1689, Article 50). For an interview, the safe practice is to say so at the start.
- AI systems intended to evaluate candidates are listed as high risk in Annex III of the AI Act. The main obligations for them apply from 2 December 2027, a date set by the Digital Omnibus on AI (Regulation (EU) 2026/1744).
- The GDPR does not ban automated screening. It gives candidates the right not to be subject to a decision based solely on automated processing that significantly affects them, such as an automatic rejection, unless an exception applies (GDPR, Article 22).
- Using AI to evaluate candidates will in most cases require a data protection impact assessment before it starts (GDPR, Article 35).
- If you have a works council, involve it early and get advice on whether its consent is needed for your set up.
For outcome and fairness measures you can keep using after the pilot, see how to measure quality of hire in frontline recruitment.
So, do AI interviews work?
In one large field test they did, on offers and starts; the rest needs your own data.
In the strongest study so far, an AI voice interviewer got about one more applicant in every 100 through to an offer and a start, with recruiters deciding and no measured loss in performance. A second study shows the risk: a recorded interview with nobody present lost more than half of the applicants who would otherwise have continued, women more than men. Both settings are far from Dutch frontline work, so choose the format carefully, tell candidates what is happening, keep a person in charge and measure by route before you scale.
How Radius helps
Radius Hire is an AI hiring platform for frontline roles in the Netherlands. Its AI Interviewer holds a short, structured voice interview that candidates can take on their phone, whenever it suits them, in the language set for the role. Candidates are told at the start that the interviewer is automated and that a person reviews the result. The interview does not read emotions or assess accent, appearance or background.
Radius never rejects a candidate on its own. A recruiter reviews the evidence, can overrule anything Radius presents and makes every hiring decision (why the tool should never be the one saying no). Each interview can be kept as a recording, a transcript and a summary, and candidate data is hosted in the EU (trust and compliance). See how the platform works.
Radius is new and no study of its interviews has been published, so judge it as you would any AI interview: in a pilot on your own roles, with the measures above.
Frequently asked questions
Are AI interviews legit?
An AI interview from a real employer is a normal screening step, but you should be told it is automated. Under the EU AI Act, providers must design AI systems that talk directly with people so that people know they are dealing with AI, unless that is obvious. Ask who reviews your answers and whether you can speak to a person instead, and treat any "interview" that asks for payment or bank details as a warning sign.
Do AI interviews lead to better hires?
There is no published evidence yet that they pick better workers. In the largest experiment, hires from both routes performed about the same and, on our arithmetic from the paper's counts, stayed at a similar rate once they started. The gain came from more applicants reaching an offer and a start.
Who makes the decision after an AI interview?
In both experiments, people did. In the larger study, a recruiter reviewed the transcript, the recording and test scores and made every hiring decision. Under the GDPR, candidates have the right not to be subject to a solely automated decision with significant effects, such as an automatic rejection, unless an exception applies.
Is an AI voice interview the same as a recorded video interview?
No. In a live AI voice interview the applicant has a spoken conversation and the AI asks follow up questions; in a recorded interview the applicant answers set questions and nobody responds. The field evidence differs for the two formats, so check which one a provider means.
Do AI interviews work for frontline jobs in the Netherlands?
No published field experiment has tested that yet, as far as we could find on 21 October 2026. The closest evidence comes from customer service jobs in the Philippines and tech jobs in the United States, so a measured pilot on your own roles is the most reliable guide for now.
Sources
- Jabarian, B., and Henkel, L. (2026). Voice AI in Firms: A Natural Field Experiment on Automated Job Interviews. arXiv 2607.28222, version 2, 11 September 2026. Working paper, not peer reviewed. https://arxiv.org/abs/2607.28222. Checked 21 October 2026.
- Avery, M., Ip, E., Leibbrandt, A., and Vecci, J. (2026). A Brave New World of Hiring: A Natural Field Experiment on How Asynchronous Interviews and AI Assessment Reshape Recruitment. University of Exeter Department of Economics Discussion Paper 26/02, version of 25 March 2026. Registered in advance (AEARCTR-0013356); not peer reviewed. https://exetereconomics.github.io/RePEc/dpapers/DP2602.pdf. Checked 21 October 2026.
- Langer, M., König, C. J., and Papathanasiou, M. (2019). Highly automated job interviews: Acceptance under the influence of stakes. International Journal of Selection and Assessment, 27(3), 217 to 234. https://doi.org/10.1111/ijsa.12246. Checked 21 October 2026.
- Langer, M., König, C. J., Sanchez, D. R. P., and Samadi, S. (2020). Highly automated interviews: Applicant reactions and the organizational context. Journal of Managerial Psychology, 35(4), 301 to 314. https://doi.org/10.1108/jmp-09-2018-0402. Checked 21 October 2026.
- Acikgoz, Y., Davison, K. H., Compagnone, M., and Laske, M. (2020). Justice perceptions of artificial intelligence in selection. International Journal of Selection and Assessment, 28(4), 399 to 416. https://doi.org/10.1111/ijsa.12306. Checked 21 October 2026.
- Sackett, P. R., Zhang, C., Berry, C. M., and Lievens, F. (2023). Revisiting the design of selection systems in light of new findings regarding the validity of widely used predictors. Industrial and Organizational Psychology, 16(3), 283 to 300. Open access. https://doi.org/10.1017/iop.2023.24. Checked 21 October 2026.
- Regulation (EU) 2024/1689 laying down harmonised rules on artificial intelligence (AI Act), Official Journal of the European Union, 12 July 2024: Article 50(1) and Annex III, point 4. https://eur-lex.europa.eu/eli/reg/2024/1689/oj. Checked 21 October 2026.
- Regulation (EU) 2026/1744 (Digital Omnibus on AI), Official Journal of the European Union, 24 July 2026. https://eur-lex.europa.eu/eli/reg/2026/1744/oj. Checked 21 October 2026.
- Regulation (EU) 2016/679 (General Data Protection Regulation), Articles 22 and 35. https://eur-lex.europa.eu/eli/reg/2016/679/oj. Checked 21 October 2026.
- Radius Hire. AI Interviewer. https://radiushire.com/platform/ai-interviewer. Checked 21 October 2026.
General information, not legal advice. Position as of 21 October 2026.