TLDR
- Use an AI interviewer for the first conversation with every applicant, and people for the later rounds and the decision. Tracy AI Interviewer is built for that split: everyone who applies is interviewed, at any hour, with follow-ups on what was said.
- The best field evidence is narrow. One experiment found more job offers with an AI voice agent in entry-level call-centre hiring in the Philippines, with humans deciding. It does not cover skilled roles, and it was not about Tracy.
- People are better where the job of the interview is judgment, rapport or the final call.
Fact-checked on 4 Nov 2026 against the original papers, surveys and Tracy's own pages. Sources are listed at the end.
An interview has several jobs: gathering the same evidence from everyone, probing what a person says, judging fit with a team, and making a decision. AI and people do those differently. This page sets out what the evidence supports for each, keeps what people think apart from what was measured, and says where a human interviewer is the better pick. For the basics of the AI side, see how AI interviews work.
What the evidence says
Measured outcome: one field experiment
The strongest study we found is a natural field experiment by Brian Jabarian (Carnegie Mellon University) and Luca Henkel (Erasmus University Rotterdam), a working paper, read here in its arXiv version 2 of 11 September 2026. It covers about 70,000 entry-level customer service applications in the Philippines between March and June 2025. Applicants interviewed by an AI voice agent received job offers 12% more often than those interviewed by a human recruiter: 9.73% versus 8.70%. Human recruiters made every hiring decision.
What to keep in mind:
- It was a voice agent, not video, and not Tracy.
- It was high-volume, entry-level hiring in one country. It says little about skilled or technical roles.
- It is a working paper, not shown to be peer reviewed. The paper discloses that on 1 August 2025, after data collection ended, Brian Jabarian accepted an unpaid role as Chief Economist at PSG Global Solutions, the recruitment firm that supplied the data.
Measured outcome: what applicants chose
In the same experiment, when applicants were allowed to choose, 78% picked an AI voice agent over a human recruiter. It shows what applicants chose in one setting and country; it does not measure general preference. The paper also notes that applicants who choose the AI voice agent have lower language and analytical scores than those who choose a human recruiter.
Measured outcome: who finished
Not every applicant finishes. In a preprint (arXiv version 2, 7 August 2026) on one Junior Frontend Engineer posting that drew 34,493 unique applicants, about one in four of the 25,536 applicants assigned to the AI-interview track (25.07 percent) completed a 30 to 40 minute AI interview. That is a completion rate for invited applicants to one posting, on one platform, from a preprint not shown to be peer reviewed. Two of its four authors are affiliated with micro1, the platform that built the AI interviewer being studied. The paper links the lower completion to the 30 to 40 minute length, and the figure cannot be read as a general completion rate for AI interviews.
Perception: what people say they want
In a Pew Research Center survey of 11,004 US adults in December 2022, 71% opposed AI making a final hiring decision, 7% favored it and 22% were not sure. The survey asked in 2022 about the final decision, so it says little about first-round interviews. It is still a good reason to keep a person on the decision.
The yardstick for any interview
A 1998 meta-analytic summary in Psychological Bulletin (Schmidt and Hunter) put the average validity of structured interviews for predicting job performance at .51, compared with .38 for unstructured interviews. Validity here is a correlation, not a percentage of right calls, and the .51 describes human-run structured interviews, not AI interviewers. A 2022 re-analysis in the Journal of Applied Psychology (Sackett and colleagues) found structured interviews to be the top-ranked selection procedure for predicting job performance. The authors also report that most of the procedures that ranked high in earlier summaries remain high in rank, with mean validity estimates reduced by .10 to .20 points.
The result supports asking everyone the same questions and reading answers against the same evidence. It gives no ground for calling an AI interview accurate because a human-run structured interview was.
What each does well
| Job of the interview | AI interviewer (Tracy) | Human interviewer |
|---|---|---|
| The same first conversation for everyone | Yes: every applicant gets a consistent first round at the same depth, and a conversation that adapts to their answers, whenever they join | Hard to keep identical across many people |
| Follow-up on a vague answer | Yes: each follow-up depends on the answer before it, and a vague answer gets a sharper question | Yes, and can read a room |
| Written result with evidence | Yes: a written assessment per question with its reasoning, beside a transcript timed to the recording | Depends on the interviewer's notes |
| Reading the CV and tailoring questions | Can: with CV check on, 3 to 5 questions tailored to the CV and the role's must-haves | Yes, as far as the interviewer's time allows |
| Reach across everyone who applies | Yes: each applicant starts on click, at any hour, with no scheduling | Limited by calendars |
| Integrity evidence | Signals such as tab switching or an external screen, marked on the recording for a person to review | By observation, not recorded evidence unless notes are kept |
| Fits the tools you use | API and MCP connections, three ways to start an interview, a customised setup for enterprise teams | Depends on the team's process |
| Fit with a specific team | Not the first round's job | Yes |
| Rapport and selling the role | Not its job | Yes |
| The final decision | No: Tracy does not reject anyone, and a person decides | Yes |
The table is our assessment of the jobs, with Tracy's column resting on its own published description. The human column is general reasoning, not a measured comparison.
Where AI falls short
- Skilled roles are not covered by the best evidence. The experiment above was entry-level. If you hire engineers or experts, run your own comparison on a small batch, and read the results next to a human interview. AI technical interviews covers what a technical first round can and cannot show.
- People who have not agreed to be recorded. Tracy records every interview with a transcript, tells people about the recording before it starts, and is not for talking to someone who has not agreed.
- Rapport. A conversation meant to sell the role or judge team fit belongs to a person.
Where people fall short
- Consistency at volume. Ten interviewers asking ten different sets of questions give you ten different data sets.
- Reach. A recruiter's calendar limits how many people get a conversation, which is the problem how to screen candidates when there are too many applicants is about.
These two points are our reasoning, not measurements.
How to combine them
- AI for the first conversation. Set the questions around the must-haves of the role. With Tracy you start from the job description or bring your own questions, and candidates join from a link in the browser with nothing to install, so the first round can cover everyone, not a shortlist.
- A person reads the results. Each interview ends in a written assessment per question with its reasoning, with the recording, transcript and integrity signals behind it. AI interview reports covers how to review them quickly.
- People run the later rounds and decide. Keep the decision with a person, as most US adults in the 2022 Pew survey wanted.
How a first round works with an AI interviewer walks through the steps.
What it costs
With Tracy the first 3 results are free, then $7 per credit, charged per started 20 minutes when the result is ready. A session with no answers is never charged. There are no seats.
Questions
Is an AI interview better than a human interview?
What is an AI interview best at?
What are human interviewers best at?
Do candidates prefer AI or human interviewers?
Should AI make the hiring decision?
Can I use both?
Try it free and interview your first applicants at no cost, or book a demo to see how a result reads.
Sources
All sources retrieved 4 Nov 2026.
- Tracy AI Interviewer, Home page: follow-ups, "You read the words, then you decide", integrity signals, "Teams can also connect through the API".
- Tracy AI Interviewer, Pricing: free results, $7 per credit, billing, no-answer sessions, no seats, Enterprise with a customised setup and API and MCP connections.
- Tracy AI Interviewer, Product: questions from the job description or your own, CV check, three ways in, no install, people told about the recording before it starts, recorded with a transcript timed to the recording, follow-ups that depend on the answer before, assessment per question, integrity signals, "Tracy does not reject anyone. A person decides.".
- Tracy AI Interviewer, Hiring at volume: each applicant starts on click, at any hour, with no calendar; a consistent first round at the same depth for applicant one and applicant five hundred.
- Brian Jabarian and Luca Henkel, Voice AI in Firms: A Natural Field Experiment on Automated Job Interviews, arXiv working paper, version 2, 11 September 2026: offers 9.73% versus 8.70%, 78% choice, the author disclosure.
- Ada Aka, Emil Palikot, Ali Ansari and Nima Yazdani, Better Together: Quantifying the Benefits of AI-Assisted Recruitment, arXiv preprint, version 2, 7 August 2026: 25.07 percent of invited applicants completed a 30 to 40 minute AI interview; two authors are affiliated with micro1.
- Frank L. Schmidt and John E. Hunter, The Validity and Utility of Selection Methods in Personnel Psychology, Psychological Bulletin, 1998: structured interview validity .51 versus .38.
- Sackett, Zhang, Berry and Lievens, Revisiting meta-analytic estimates of validity in personnel selection, Journal of Applied Psychology, 2022: structured interviews top-ranked.
- Pew Research Center, Americans' views on use of AI in hiring, chapter 1 of AI in Hiring and Evaluating Workers: What Americans Think, 20 April 2023: 71% oppose AI making a final hiring decision.