TL;DR: How AI interviews work: an AI interview is a structured interview in which software asks the questions, records and transcribes the answers, and scores them against criteria the employer sets. People review the evidence and make the decision. This page walks through seven stages, with what the AI does, what the human does, and what is recorded at each.
Updated 28 October 2026.
Key takeaways
- An AI interview is structured: the same questions, scored against the same rubric, with a transcript behind every score. People make the decision.
- There are seven stages: define criteria, build questions, invite and disclose, run the interview, transcribe and score, integrity checks, human review and decision.
- 63% of job seekers have faced an AI interview, and 70% were not clearly told AI would evaluate them (Greenhouse, n=2,950, 1 May 2026; vendor survey). The invitation and notice step is where candidate trust is won or lost.
- What the AI should not do matters as much as what it does: no emotion inference (the EU has banned it in workplaces since 2 Feb 2025) and no facial scoring (HireVue withdrew its facial analysis in a January 2021 announcement).
- In the one large field experiment, AI-led voice interviews with human decision-making raised job offers 12% and 30-day retention 16% (as reported by Chicago Booth), at one recruitment firm (RPO), in one country, for one role family.
AI interview software is the tool that runs this process; this page covers the process. It is written for hiring teams. If you are a candidate looking for preparation tips, this is the wrong page, and nothing here is designed to help anyone pass an interview.
What is an AI interview?
Definition (ours): An AI interview is a structured interview in which software asks the questions, records and transcribes the answers, and scores them against criteria the employer sets. It supports early-stage evaluation. People review the evidence and remain responsible for the hiring decision.
We found no neutral definition source in the top results, so this one is ours, and it is written to be quoted. Our definition separates three types:
- Automated one-way video: the candidate records answers to set prompts, and software scores them.
- AI-assisted interviews: a person conducts or attends the interview, and AI supports with questions, notes or scoring.
- Fully AI-led interviews: software conducts the conversation itself, and a person reviews the output afterwards.
Check which type a vendor is selling you before comparing prices or features.
How does an AI interview work, from setup to decision?
Follow one interview from start to finish. Each stage below has three lines: what the AI does, what the human does, and what is recorded. After the seven stages, an illustrative evidence record shows what a reviewer would read.
1. Define the role, criteria and rubric
What the AI does: Nothing yet. What the human does: Decides what the job requires and writes it down as criteria and a scoring rubric. What is recorded: The rubric and its version.
The rubric is the contract. Criteria come from the job, not from the tool's defaults, and everything downstream is only as good as this step.
2. Build structured questions and scoring anchors
What the AI does: Drafts questions and describes what a strong, middling and weak answer looks like. What the human does: Edits and approves them. What is recorded: The final question set and anchors.
Every candidate gets the same questions. That is the point: a structured format lets you compare candidates on the same footing. SIOP's summary of the 2022 Sackett, Zhang, Berry and Lievens paper in the Journal of Applied Psychology says structured interviews may in fact be the strongest predictor of job performance, not cognitive ability.
3. Invite and disclose
What the AI does: Sends or hosts the invitation. What the human does: Writes the notice, offers the alternative process and collects consent. What is recorded: The notice text shown and each candidate's consent.
The notice, the alternative route and the consent record all start here. The next sections cover what candidates ask for and what the law requires. The candidate's side of the flow is not described on this page.
4. Run the interview
What the AI does: Conducts the interview as an asynchronous conversational voice or video session, or assists a live panel. What the human does: Either nothing during the session, or runs the panel. What is recorded: The recording and its transcript.
This page does not state a duration, because duration depends on the format and the question set.
5. Transcribe and score each answer against the rubric
What the AI does: Turns speech into a timestamped transcript and scores each answer against the rubric criterion. What the human does: Nothing yet. What is recorded: Each score tied to a transcript line and a criterion.
A score is only useful if a reviewer can follow it back to the words. For what a good explanation contains, see how scores are explained.
6. Integrity checks
What the AI does: Surfaces integrity signals, such as multiple faces in frame, a face absent from the frame, liveness issues or suspected reading from a screen. What the human does: Looks at the signal and decides whether a follow-up is warranted. What is recorded: The signals and when they occurred.
The line that matters is between detected and inferred. A tool can detect that a second face appeared in the frame. It can only infer why. A signal is a prompt for a person, not a verdict. Proxy interviewing is a documented problem, and the proxy interview fraud hub covers what is known. This page makes no claim that any tool verifies identity or detects impersonation.
7. Human review and decision, logged
What the AI does: Presents the evidence and, optionally, a recommendation label. What the human does: Reads the evidence per answer and decides. What is recorded: The decision, who made it, and when.
Each action, whether advance, hold, reject or submit to client, should be logged with who took it and when. The reviewer's name matters as much as the candidate's score.
What a reviewer sees: an illustrative evidence record
Illustrative example, not a customer record. The values are invented to show the fields.
| Field | Example |
|---|---|
| Question | "Tell me about a time a delivery slipped. What did you do first?" |
| Transcript excerpt | [00:04:12] "...I called the client the same morning, before they heard it from anyone else, and gave them two options with dates." |
| Rubric criterion | Communication under pressure: informs stakeholders early and offers options |
| Alignment with reference answer | 70% |
| Score and written reason | 3 of 5: contacted client early and gave options; did not describe preventing a repeat |
| Recommendation label | Advance (advisory) |
| Reviewer action and time | Advanced to panel by [reviewer name], 09:42 |
Every row traces back to something the candidate said or something a named person did.
What does the AI evaluate, and what should it never infer?
What the AI does NOT do: It does not decide who gets hired. It should not infer emotion or personality. It should not score facial expression or voice tone.
The legal baseline in the EU is clear. Article 5(1)(f) of the AI Act bans emotion recognition in the workplace, and has done since 2 February 2025. The commercial precedent runs the same way: HireVue removed the facial-analysis component of its assessments after an announcement in January 2021, which is the only decision this page relies on.
Beyond that, the buyer's question is scope. Proctoring-style signals such as look-aways and head turns are a category to ask about. Are they in the score, or only beside it? Can they be excluded? A vendor that cannot answer has told you something.
Bias evidence has a scope problem too. A University of Washington audit, published 31 Oct 2024, tested LLM resume ranking, not interviews. It does not tell you how interview scoring behaves, so treat it as context, not as evidence about the stages on this page.
Which formats of AI interview exist?
The formats differ in who asks, who scores, and what ends up on record. The descriptions below are by design, not market data.
| Format | Who asks | Who scores | What is recorded | Candidate effort |
|---|---|---|---|---|
| One-way video with AI scoring | Pre-set prompts, text or recorded | Software against a rubric | Video answers and transcript | Records answers alone, within the invitation window |
| Conversational voice or video AI interviewer | Software, adapting to answers | Software against a rubric | Recording, transcript | A live exchange with the software, at a time the candidate picks |
| AI-assisted live | A person, with AI suggesting follow-ups | A person, with AI support | Recording, transcript, notes | Joins a scheduled call with a human |
| Phone | Software or person, by voice | Software or person | Audio, transcript | Takes or places a call |
The full formats table lives in the pillar page.
What must you tell candidates, and what does the law require?
This section summarizes obligations by jurisdiction. It is not legal advice.
NYC Local Law 144 applies to automated employment decision tools used for NYC roles. According to the NYC Department of Consumer and Worker Protection, the employer must have an independent bias audit within one year before use and give candidates notice at least 10 business days beforehand.
Illinois has its own Artificial Intelligence Video Interview Act (820 ILCS 42) covering AI video interviews. In our reading, it asks employers to give notice, obtain consent and explain how the AI works, and it covers deletion on request; read the statute itself before relying on that summary. For the consent question, see interview recording consent.
In the EU, recruitment AI is classed as Annex III high-risk under the EU AI Act for recruitment, listed under Annex III point 4, "Employment, workers' management and access to self-employment". Regulation (EU) 2026/1744 of 8 July 2026, the Digital Omnibus, moved the application of the high-risk rules for Annex III systems to 2 December 2027; Gibson Dunn tracks the changes. The emotion-recognition ban has applied since 2 February 2025.
Candidates have their own list of requests. The full set of candidate-trust figures, with samples and verification dates, is on the sourced AI recruiting stats page. The figures most relevant to this step:
| Figure | Source |
|---|---|
| 70% were never clearly told AI would evaluate them | Greenhouse, n=2,950, 1 May 2026 (vendor survey) |
| 21% learned only once the interview started | Greenhouse, same survey |
| 44% want upfront disclosure; 39% want clarity on what is measured; 46% want a human-interview option | Greenhouse, same survey |
| 38% withdrew because of an AI interview | Greenhouse, same survey |
| 51% of completers never heard back; 28% advanced; 13% were formally rejected | Greenhouse, same survey |
| 26% trust AI to evaluate them fairly | Gartner, 31 Jul 2025; the release cites a 1Q25 survey of 2,918 candidates |
| 25% trust employers less if they use AI to evaluate them | Gartner, same release |
The implication: invite-and-disclose is a funnel metric as well as a legal one. A candidate who is not told, or who never hears what happened, has every reason to withdraw.
Do AI interviews improve outcomes?
One large field experiment says they can, under specific conditions. In "Voice AI in Firms" by Jabarian and Henkel (SSRN 5395709, August 2025), the authors studied more than 70,000 applicants to entry-level customer-service postings at one recruitment firm (RPO), PSG Global Solutions, hiring in the Philippines. According to Chicago Booth's summary, AI-led interviews were associated with 12% more job offers, 18% more job starts and 16% better 30-day retention (PSG's press release gives 17% for retention), and 78% of applicants chose the AI recruiter when offered the choice. Recruiters made the final decisions.
Read that in context: one recruitment firm, one country, one role family. The study says nothing about senior or low-volume roles, and it does not show that AI beats a trained human using a structured format.
To check claims like these in a demo, use the vendor scorecard for AI hiring tools.
FAQ
Does AI interview software decide who gets hired?
No. People review the evidence and decide. Some tools let the employer configure what happens automatically, so ask the vendor what is and is not possible.
What does the AI evaluate in an interview?
Answer content against the employer's criteria. Ask whether voice, face or timing is in the score. In the EU, emotion recognition at work has been banned since 2 February 2025.
Is the interview recorded?
It depends on the tool and the jurisdiction. In our reading, Illinois requires notice and consent for AI video interviews (820 ILCS 42). For the consent question in detail, see interview recording consent.
How long does an AI interview take?
Duration depends on the format and the question set, and this page does not quote a figure because it has no sourced one. Ask your vendor for the range across their formats.
Can candidates opt out?
That depends on the employer's process and on the law; in our reading, Illinois requires consent for AI video interviews. In the Greenhouse survey, 46% of candidates said they wanted a human-interview option, so offering one is also a trust measure.
How are AI interview scores calculated?
Per answer, against a rubric criterion, with written reasoning and the transcript as evidence. The question to put to any vendor is whether you can see the evidence behind every score.
How do AI interviews detect cheating?
They surface integrity signals for a human to look at. The distinction that matters is between what a tool detects and what it infers. No identity or deepfake claim is made here. For tools built around this question, see interview cheating detection software.
Is an AI note-taker the same as an AI interviewer?
No. Note-takers listen to an interview a person runs; an AI interviewer asks the questions. For the first category, see AI interview note-takers.
Next step
Book a demo to see what the candidate sees, then read the evidence it produces.
Comments
No comments yet — start the conversation.