What is an AI interviewer? A 2026 buyer's guide

An AI interviewer is software that runs a job interview itself: it asks the questions, listens to the answers, follows up, and hands a person an evaluation to decide on. The category is new enough that the label gets stretched over almost anything with “AI” and “interview” on the box: scheduling bots, one-way video tools, online tests. This guide draws the line, walks through how a voice AI interview works from invitation to decision, and ends with the questions every buyer of AI interview software should ask, including the ones you should ask us.

What an AI interviewer is, and what it isn’t

The definition we use: an AI interviewer conducts a structured, two-way interview on its own, adapts its next question to the last answer, and produces an evaluation tied to evidence for a human to decide on. Three kinds of tools sit close to it and get confused with it:

• Recruiting chatbots schedule interviews, answer FAQs, and collect knockout answers (“Are you authorized to work in this country?”). They handle the process around an interview and leave the interview itself to someone else.
• One-way video interviews show fixed prompts and record the candidate’s answers on camera, to be reviewed later. Nobody follows up. We compared the two formats in AI voice interviews vs one-way video.
• Assessments, such as cognitive tests, situational judgment tests, and work samples, measure a specific thing with a fixed instrument. They are often worth pairing with an interview, though none of them holds a conversation.

	Converses with the candidate	Follows up on answers	Produces an evaluation
Recruiting chatbot	Text, mostly logistics	No	Rarely
One-way video	No (records answers)	No	Yes, after review
Assessment	No	No	Yes
AI interviewer	Yes, two-way	Yes	Yes, for a human to decide on

The quick test: if the tool can’t ask a second question that depends on the first answer, it is not an interviewer.

How a voice AI interview works, end to end

Strip away the branding and every serious AI interviewer runs the same six steps.

1. A rubric is authored from the job. The role’s competencies come out of the job posting, each with observable levels such as “named the tradeoff and what it cost” instead of vague ones like “seems strong.” The rubric is then frozen as a version, so every candidate for the role is measured against the same standard.
2. The candidate gets a link. An invitation goes out, reminders follow, and the candidate starts now or picks a time. No recruiter calendar involved.
3. The conversation runs. The interviewer opens, covers each competency, follows up where an answer is thin, watches the clock, and closes on time. In a live, two-way voice interview the candidate can interrupt, ask about the role and switch language, and the interviewer builds each follow-up from the previous answer. In a voice interview, that is about ten minutes.
4. Speech becomes a transcript. The transcript is the record: it is what gets scored and what you can audit later.
5. Each competency is scored against the rubric. The score maps the answer onto one of the rubric’s levels and cites the moment in the transcript that earned it.
6. A human decides. A reviewer reads the scorecard, checks the evidence, and decides who advances. Automation can route candidates, but the judgment stays with a person.

The quality of an AI interviewer lives in the joints between those steps: whether the rubric holds still, whether the follow-ups stay inside it, and whether every score points back to words the candidate actually said.

Why the category exists: structure, at volume

AI interviewers exist because an old research finding has run into a capacity problem.

Structured interviews (the same competencies for everyone, a consistent way into each, answers scored against defined levels) have long been among the best predictors of job performance. Schmidt and Hunter’s 1998 meta-analysis put their validity at .51, against .38 for unstructured interviews. When Sackett and colleagues revisited those estimates in 2022 and corrected for a systematic overcorrection, most numbers fell. Structured interviews still came out ranked first of all selection procedures, at .42, while unstructured interviews dropped to .19.

The capacity problem is one those same authors name. They note that in their traditional in-person form, structured interviews are generally not viable for high-volume jobs. They add that as interviewing moves to technology and automation, its validity has to be evaluated under those new conditions. That second point is a fair warning to everyone selling in this category, us included. We worked through the arithmetic of the first in the math of high-volume screening.

The early field evidence is encouraging, with caveats attached. In a natural field experiment with 70,000 applicants, Brian Jabarian and Luca Henkel randomly assigned candidates to human recruiters or AI voice agents, with human recruiters making every hiring decision. Applicants interviewed by AI were 12% more likely to receive a job offer, the gains carried into job starts and retention, and hired workers were no less productive. The authors found the AI interviews more structured and consistent while still responsive to each applicant. The study covers one setting, entry-level customer service roles in the Philippines, and it is a working paper, so the verdict is still open.

The AI interviewer buyer’s checklist: 11 questions for any vendor

Each question has a good answer and a tell. Ask all eleven.

1. Where does the rubric come from, and does it hold still? Good: authored per role from the job, written as observable levels, versioned and frozen so every candidate is measured against the same version. Tell: a rubric generated per candidate, or one you are not allowed to see.

2. Can every score be traced to the candidate’s own words? Good: each rating cites a verbatim transcript moment you can open. Tell: a “match score” with a paragraph of adjectives. We wrote a full guide to what to cite and what to ignore.

3. Does it analyze faces, vocal tone, or emotion at all? The right answer is no. The EU AI Act’s Article 5(1)(f) prohibits AI that infers emotions in the workplace, and has applied since 2 February 2025. The Commission’s guidelines treat hiring as part of the workplace context and tie the ban to inferences from biometric data. Tell: “confidence,” “enthusiasm,” or “energy” anywhere in the demo.

4. Which rubric version and which model produced this score? Good: both recorded on every scorecard, and changing the bar creates a new version instead of rewriting the old one, so last quarter’s scorecards stay comparable. Tell: “the model keeps improving,” with no version on the result.

5. Who makes the decision, and can the tool reject anyone on its own? The EU AI Act requires high-risk systems to be built so people can effectively oversee them, including the ability to disregard or override the output. Good: a human decides, and automation only routes. Tell: an “auto-disqualify” toggle.

6. What does the candidate actually go through? Candidates do not automatically warm to automation. In one experiment, people shown a highly automated interview accepted it less, through lower perceived social presence and fairness (Langer et al., 2019). Yet in the field experiment above, 78% of applicants offered the choice picked the AI interviewer (ERE). The design of the experience makes the difference, so ask: is an app or account required? Does it work on a phone? (For 16% of US adults, a smartphone is how they get online; they have no home broadband.) Is the AI disclosed up front? How long does it take? What is the route for a candidate who cannot do a spoken interview?

7. Which languages, and how good is the transcript? A voice interviewer scores what speech recognition heard. A 2020 study of five commercial speech recognition systems found an average word error rate of 0.35 for Black speakers against 0.19 for white speakers (Koenecke et al.). The systems have improved since, and the lesson still applies: ask how accuracy holds across accents and languages, whether a candidate can switch languages mid-interview, and whether you can read the exact transcript a score was built on.

8. Where do candidates drop off? Completion decides whether any of this works. Ask for the funnel by stage (link opened, microphone allowed, first answer, halfway, finished) and what happens when a connection drops. Ask to see those numbers from a pilot with your own candidates.

9. Does it fit the systems you already run? Good: candidates import from your ATS, interviews trigger when someone reaches a pipeline stage, results flow back, and events arrive as signed webhooks. Tell: “integration” means a CSV export.

10. How long is data kept, and who controls it? The rules pull in different directions. The GDPR says personal data should be kept no longer than necessary. California’s rules on automated-decision systems require keeping those records for four years. Illinois requires deleting AI-analyzed video interviews within 30 days of a candidate’s request, backups included. You need retention you can set per jurisdiction, and deletion on request that reaches every copy.

11. What does it give you for the law? New York City’s Local Law 144 requires a bias audit within one year of use, a public summary, and notice to candidates, and has been enforced since July 5, 2023. The EU AI Act classifies AI that evaluates candidates as high-risk; after the Digital Omnibus, those obligations apply from 2 December 2027. Ask what the vendor hands you: the data an audit needs, instructions for use, logs. Our guide to the rules is AI interviews and the law.

Red flags, in one list

• Emotion, confidence, or “culture fit” read from a voice or a face.
• Scores with no citations behind them.
• A rubric you cannot see, or one that shifts between candidates.
• Automatic rejections.
• No exportable transcript.
• Validity or fairness claims with no method you can inspect.

Where Hure lands

Hure is a voice AI interviewer, so here are our answers to the same list.

Rubric: authored per role from the job posting, then sealed and versioned, so every candidate for a role is scored against the same version and scorecards are comparable. Evidence: every score cites a verbatim transcript moment mapped to an observable level of the rubric. Biometrics: none, and no emotion inference. Hure scores only what candidates say, with no input from how they look or sound. Follow-ups: adaptive, but only with the techniques your team approved for each competency, and the interview closes gracefully inside about ten minutes. Decision: always a human’s. Workflows automates routing and has no reject action. Candidates: no app and no account. They get one signed, single-use browser link on a phone or a laptop, in 100+ languages with live switching, and start now or at a time they pick. Integrations: Greenhouse, Lever, and Ashby to import candidates and trigger interviews by stage, plus HubSpot, Gmail, Notion, Slack, and signed webhooks.

We have two gaps to name. There is no public validity study or independent bias audit of Hure to point you to today. And a spoken interview is not right for every candidate, so keep another route open for anyone who needs one. Weigh both the way this checklist says to weigh any vendor.

See how the interviewer works on the product page.