When 70,000 real applicants were randomized between an AI voice agent and a human recruiter for the same screening interview, 78% chose the AI. They were not choosing a faster process or a cheaper employer. They were choosing a format that fit the way they actually live: in transit, between shifts, after 9pm, on a phone with a cracked screen.

The number comes from a 2025 natural field experiment by Brian Jabarian at the University of Chicago Booth and Luca Henkel at Erasmus University Rotterdam, published on SSRN and covered in Bloomberg. The supporting data is what should change how an operator builds their funnel: AI-screened candidates received 12% more offers, started their jobs 18% more often, and were 17% more likely to still be in the role 30 days later (Jabarian & Henkel, August 2025).

This is not the death of the human interviewer. The same datasets show candidates still want humans for the final round, salary negotiation, and culture-fit conversations: 89%, 84%, and 71% respectively (Daily Hire, November 2025). What is dying is the phone tag bottleneck at the front of the funnel.

The phone tag era is over. The question for operators is what replaces it, and how to keep the human step where it actually matters.

The 78% number: where it comes from

The 78% finding is from the Jabarian and Henkel natural field experiment, conducted across 48 roles in healthcare, IT, and industrial sectors, with a sample of approximately 70,000 applicants randomized between an AI voice agent and a human recruiter for the same structured screening script (SSRN, August 2025; Bloomberg, 2025). The choice was offered to every applicant; the default was not imposed.

Three things make this number credible. The sample is large enough that the result is not a methodological artefact — n ≈ 70,000 is several orders of magnitude beyond the typical preference survey. The choice was real: applicants were not asked hypothetically, they picked. And the outcome data is downstream, not self-reported — the 78% choice is paired with 12% more offers, 18% more job starts, and 17% higher 30-day retention for the AI-screened cohort. Candidates who chose the AI did not just prefer it; they performed better against the rubric.

The 78% is not universal. It is a preference for the screening step. For the final round, salary conversation, and culture-fit, the same surveys put the human preference at 89%, 84%, and 71% respectively (Daily Hire, November 2025). The AI belongs at the front of the funnel; the human belongs at the back. Forcing either into the other's slot produces worse outcomes for both.

The three reasons: convenience, time, no bias

Three reasons carry most of the preference signal in the Jabarian and Henkel data.

Convenience. 64% of AI voice interview completions happen outside business hours (Daily Hire, November 2025). The candidate who finishes the screening at 11pm, after her second shift, was never going to take a 10am Tuesday call with a human recruiter. The recruiter's calendar does not bend to the candidate's. The AI's does.

Time. The median time from profile to interview for the AI-screened cohort was 0.32 days, versus 0.51 days for the human-screened cohort. The AI is not faster at deciding — the human review step is where the AI-screened process actually slowed down, producing a 22-day total time-to-hire versus 19 days for the human-led process. The AI is faster at scheduling and running the interview. For the candidate, that is what speed feels like.

Perceived bias. 3.30% of AI-screened candidates reported perceived gender bias, versus 5.98% in the human-screened cohort (Jabarian & Henkel, 2025). 71% of AI-interviewed candidates left positive free-text comments, versus 52% in the human group. Candidates do not always experience human screening as more humane. They experience it as more variable.

The three reasons overlap. A structured, on-her-time, fair-process interview at 11pm wins. A 25-minute hold-time at 2pm loses, and the candidate tells her friends.

After-hours completion: 64% of candidates finish outside 9-5

The single most under-discussed number in the AI voice screening literature is the 64%. 64% of AI voice interview completions happen outside business hours (Daily Hire, November 2025). The same dataset does not exist for human-screened candidates, because a 25-minute live phone screen at 10pm is not a service any recruiter is offering.

The implication is structural. The total addressable candidate pool for a 9-to-5 human phone screen is the subset of applicants who can take a 9-to-5 phone call. The total addressable pool for a 24/7 AI voice screen is every applicant with a phone. Operators who treat the AI as a "faster version of the same screen" miss this — the AI is reaching a different candidate.

The pattern shows up in secondary screens too: 36% of candidates actively chose the AI interviewer over a human recruiter screen when given the option (Ashby / Talent Llama, 2025). The candidate who picks the AI is not making a technology choice; she is making a logistics choice.

The 3-point completion gap between AI voice and async video

The channel war inside high-volume screening is now a two-horse race: AI voice versus async video. The completion gap is large and consistent.

AI phone screens complete at above 80% on mobile. One-way video interviews complete at 40–60% on the same devices (Heymilo / Classet, 2026). 33% of candidates abandon lengthy one-way video interviews (1SecondHire, 2025), concentrated in candidates who do not enjoy being on camera — not the trait you want to be selecting against unless the role is literally on-camera.

Video on a phone requires a stable connection, decent lighting, a quiet room, and a willingness to record yourself. Audio on a phone requires none of those. Audio is the format the candidate can do in transit, on a break, in a car.

The structural read: voice-first is the default for hourly, shift-based, retail, hospitality, BPO, and frontline healthcare. Video-first remains a defensible choice for sales, client-facing, and senior technical roles where on-camera presence is itself a job-relevant skill (IJSA study, 2025; Pin, 2026). The error case is using video-first for a high-volume frontline role and being puzzled by the 50% completion rate.

Where humans still matter: judgment, calibration, final decision

The 78% preference is for the screening step. It is not a preference for AI to make the hiring decision. 89% of candidates want a human involved in the final round; 84% in salary negotiation; 71% in culture-fit (Daily Hire, November 2025). These are the steps where the human is the product, not a cost.

The AI replaced the 25-minute phone screen, not the 60-minute debrief. Operators who push the AI further down the funnel will hit the preference wall at 89%.

The product pattern: AI does the structured, repeatable, loggable step. The human does the contextual, calibrated, judgment step. AI-assisted, human-decided is not a tagline. It is the architecture that produces the 12/18/17 outcome improvements the data actually shows.

The AI is doing the 25-minute phone screen. The human is doing the 60-minute decision. The two are not interchangeable.

The dropped-perceived-gender-bias finding

The most actionable secondary finding from the Booth/Erasmus study is the perceived gender bias drop. 3.30% of AI-screened candidates reported perceived gender bias, versus 5.98% in the human-screened cohort (Jabarian & Henkel, 2025). A 45% relative drop, measured in a randomized controlled design with 70,000 applicants.

The number is not a guarantee the AI is unbiased — any AI screening system can encode the bias of its training data, rubric designers, or prompt. What it shows is that a structured AI interview, with a fixed rubric and fixed question order, produces a more consistent candidate experience than a human recruiter on their 30th screen of the day. Consistency is the variable the candidate experiences as fairness.

The corollary: the bias audit is not optional. NYC Local Law 144 requires an annual independent bias audit, refreshed and posted publicly (NYC DCWP, 2023). Illinois's HB 3773, effective 1 January 2026, requires annual demographic reporting to DCEO (Introl, 2026). The EU AI Act's high-risk employment provisions, in force from 2 August 2026, require continuous bias monitoring (European Commission, 2026). Candidates who experience the AI as more consistent are also more likely to recommend the employer.

What high-volume operators are doing differently in 2026

Five patterns have emerged across the public 2025–2026 case studies, and they are the spec every BPO, contact center, retail, and hospitality operator is now benchmarked against:

AI voice at the front of the funnel, human at the back. A global BPO deployed voice AI across five countries and reported a 42% drop in time-to-hire, 50% fewer unqualified finalists, and 20% higher offer acceptance (Andru.ai 2025 ROI report). AI runs the structured screen; human runs the rubric review and the final round.
Same-day conditional offer at first contact. Amazon's hourly warehouse flow is no-resume, no-interview, conditional offer the same day as application, with I-9 and tax forms completed online before the in-person event (Amazon hiring site, 2026). Walmart extends a verbal offer during the call. McDonald's is at 1–3 days from application to start in most US locations. ADNOC hired 1,100+ UAE nationals in 2025 against a 62.8% Emiratization target (The National, June 2026).
Compliance bundled into the interview. Right-to-work checks, I-9 data capture, E-Verify, HRIS provisioning, and policy acknowledgments are stitched into the same flow as the screening (Klippa, 2026). For a GCC operator, the bundle also covers the Qiwa-documented contract and the Nitaqat-sector quota check.
Voice agents running 24/7 in the candidate's language. Code94 Labs, JXG, and Janashakthi Group in APAC ran 488 resumes to 15 shortlisted in one week, and 150 candidates screened in five days versus a prior 4–5 week manual process (Talvin AI, 2025–2026). Express Package in US logistics cut candidate-to-in-person time from 7 days to 2 days (HappyFleet, 2025–2026).
Async voice, not async video, for hourly. The completion gap between AI phone screens (80%+) and one-way video (40–60%) is the single biggest channel decision an operator makes in 2026 (Heymilo / Classet, 2026).

A 30-day operating plan for a BPO/contact-centre talent lead

A practical sequence a BPO or contact-centre TA lead can run in 30 days:

Days 1–3: instrument the funnel. Capture source, channel, time-to-screen, time-to-offer, and 30-day retention by cohort. If you cannot segment AI-screened from human-screened candidates, you cannot measure the lift.
Days 4–7: pilot AI voice on one high-volume role. Pick the role with the worst phone-tag bottleneck.
Days 8–14: build the rubric with the hiring manager. A 6-to-8 question structured rubric with a defined scoring guide. The rubric is the product; the AI is the messenger.
Days 15–21: switch the interview invite to the candidate's channel. WhatsApp in the GCC and India. Email plus SMS in the US and EU. Open and reply rates are the metric.
Days 22–25: run the bias audit. Document the rubric, the question order, the training data, and the scoring logic.
Days 26–28: build the human handoff. The recruiter reviews a structured scorecard and a 5-minute summary. The hiring manager gets a transcript, a rubric, and a one-line summary for every AI-screened candidate.
Days 29–30: publish the result. Time-to-screen, time-to-offer, completion rate, and 30-day retention, by cohort, internally.

The plan is not a tooling project. It is a discipline. The data is in the public domain and the candidates are already choosing it.

FAQ

Q1. Where does the 78% preference number come from? The Jabarian and Henkel natural field experiment (University of Chicago Booth and Erasmus University Rotterdam, SSRN, August 2025; n ≈ 70,000 applicants, 48 roles in healthcare, IT, and industrial). 78% chose the AI voice interview when given a free choice (Daily Hire, November 2025).

Q2. Does the 78% preference mean AI is better at the whole hiring process? No. Candidates still want humans for the final round (89%), salary negotiation (84%), and culture-fit (71%). The AI wins the screening step; the human wins the decision step.

Q3. What about after-hours candidates who couldn't take a human call anyway? 64% of AI voice interview completions happen outside business hours (Daily Hire, November 2025). The AI is reaching a different candidate population, not just running the same screen faster.

Q4. Does AI voice really complete better than one-way video? Yes. AI voice completes at above 80% on mobile; one-way video at 40–60% (Heymilo / Classet, 2026). 33% of candidates abandon lengthy one-way video interviews (1SecondHire, 2025).

Q5. Is the perceived gender bias drop real? Yes: 3.30% of AI-screened candidates reported perceived gender bias versus 5.98% in the human-screened cohort (Jabarian & Henkel, 2025). It is not a guarantee of AI fairness — bias audits are still required under NYC LL144, Illinois AIVFA, and the EU AI Act.

78% of Candidates Choose AI Voice Over a Human for the First Round