TL;DR: AI interview software conducts, transcribes and scores structured candidate interviews against criteria the employer sets, then returns evidence for a person to review. What separates good tools from weak ones is whether each score traces to the candidate's actual answer, and who holds the reject decision. Before anything else, ask a vendor to show the evidence behind one score and how its decision gate is configured.
General information, not legal advice. Have counsel confirm how this applies to you.
Updated 21 October 2026 — EU high-risk dates, state-law status and Greenhouse/SHRM figures updated.
Disclosure: ZeaHire sells AI interview software. This page is built to be useful if you buy elsewhere.
Key takeaways
- AI interview software conducts and scores structured interviews and hands a person the evidence. The formats differ in what they decide and what they leave behind.
- 63% of job seekers have faced an AI interview, and 70% of surveyed job seekers said they were never clearly told upfront that AI would evaluate them (Greenhouse, n=2,950, 1 May 2026; the vendor's own survey). Disclosure is now a buying criterion, not a footnote.
- Who holds the reject decision is a configuration question. Ask every vendor whether the system can reject or advance a candidate with no human action, and whether that is enforced or a setting.
- Structured interviews may be the strongest single predictor of job performance, according to SIOP's summary of the Sackett et al. (2022) re-analysis (SIOP TIP). (Validates structure, not any AI product.)
- Recruitment AI is a high-risk use under the EU AI Act. Regulation (EU) 2026/1744, the Digital Omnibus on AI, moves the Annex III high-risk obligations to 2 December 2027.
A demo shows you a scoring screen. A scoring screen does not tell you the two things that decide whether a tool is safe to run: what evidence sits behind each score, and who can reject a candidate without a person ever looking.
Candidates are already on the other side of these tools, and most were never clearly told. This page defines the category, sorts the formats, and gives you six questions that expose weak tools in a demo. If you want the mechanics first, start with how AI interviews work step by step; the version here is a summary.
What is AI interview software?
AI interview software is software that conducts, transcribes and scores structured candidate interviews against criteria the employer sets, and returns evidence for a person to review: the transcript, the reasoning behind each score and the criteria applied. The software runs the interview; a person reviews the result.
That definition separates it from three neighbours. Video interview software records answers for a human to watch. An applicant tracking system tracks candidates and does not interview them. AI note-takers listen to an interview a human runs and write it up afterwards, where AI interview software runs the interview itself.
We sort the category into three working labels: automated one-way video, AI-assisted live, and fully AI-led. The labels matter because they answer different questions. In the first, a candidate talks to a camera and nobody answers back. In the second, a person interviews and the software supports. In the third, the software asks, listens and follows up. Which one fits depends on volume, role type and who must sign off, so there is no single best tool. That is also why this page carries no ranked list.
This page covers the software employers and staffing firms buy. It does not cover candidate preparation, "how to pass an AI interview" advice or coding assistants, which serve a different reader.
What formats of AI interview software exist?
Add skills simulations to those three labels and the category contains four formats. They differ less in the technology than in three practical things: what the AI decides, what evidence it leaves behind, and who makes the final call.
| Format | What the AI decides | What evidence it leaves | Who makes the final call | Typical disclosure need |
|---|---|---|---|---|
| Conversational voice or video AI interviewer | Which follow-up to ask; scores each answer against the employer's criteria | Transcript, per-answer score, and reasoning if the vendor exposes it | Set by configuration: a person, or a rule | Notice that AI conducts and evaluates the interview |
| One-way recorded video with AI scoring | Scores recorded answers; no live follow-up | Recording and score; ask what reasoning sits underneath | Set by configuration | Notice that recordings are analysed, and what is scored |
| AI-assisted live interview | Nothing on its own: a human interviews and the AI supports with notes, prompts or summaries | Notes, summary, transcript | The human interviewer | Notice and consent for recording and AI processing |
| Skills simulation or assessment | Scores performance on a task | Task output, score, scoring criteria | Set by configuration | Notice of what is assessed and how it is scored |
Simple one-way video is genuinely simpler for teams that want recorded answers and a human watching. That is a fair reason to choose it, and it is not a ranking.
The format question has a precedent worth knowing. In January 2021, HireVue announced it was removing the facial-analysis component of its assessments. (The 2021 decision only; no claim about HireVue's current product.) Our position is that facial scoring should be treated as disqualifying.
Why facial and voice-tone scoring is the line to hold
Answer content can be quoted, checked and challenged. A score derived from a face or a tone of voice cannot be quoted back to the candidate in any meaningful way, which makes it hard to review and harder to defend. The EU has drawn a legal line nearby: Article 5(1)(f) of the AI Act prohibits "the use of AI systems to infer emotions of a natural person in the areas of workplace and education institutions," except for medical or safety reasons (EUR-Lex). The text says workplace and does not name recruitment, so how far it reaches interview analysis is a reading that counsel should confirm. Whether a particular scoring feature counts as emotion recognition is a question for counsel. Whether it is in the score is a question for the vendor, and it is the third row of the demo table below.
How does AI interview software work?
In summary: the employer defines the criteria for the role. The software builds structured questions from those criteria. Each candidate receives an invitation that discloses AI use. The interview runs, whether by voice, video or text. A transcript is produced and each answer is scored against the criteria. Integrity signals are recorded, such as whether more than one face appears or the candidate looks away from the screen. Finally, a person reviews the evidence and decides, with the degree of automation depending on how the decision gate is configured. The stage-by-stage version of AI-led screening interviews, with what gets recorded at each step, is on its own page.
What should it leave behind? Evidence, not verdicts
The rule is simple: every score should trace to the candidate's own words and to a named criterion. A reviewer should be able to open the transcript, find the answer, read the reasoning and see the rubric that was applied. That is what separates evidence, not verdicts from a number on a dashboard. We use the word "consistent" for what a scoring system can offer. "Unbiased" and "reduces bias" are claims that need audit evidence.
Structure is the reason the evidence is worth collecting. Structured interviews put the same questions to every candidate, which gives a reviewer a consistent basis for comparison. SIOP's summary of the Sackett, Zhang, Berry and Lievens re-analysis (Journal of Applied Psychology, 2022) says "structured interviews may in fact be the strongest predictor of job performance—not cognitive ability" (SIOP TIP). The paper itself is on PsycNET, where it reports a mean operational validity of .42 for structured interviews. (Supports structure, not software over a trained interviewer.)
There is field evidence on AI-led interviews with people keeping the decision. Jabarian and Henkel's working paper "Voice AI in Firms" (SSRN 5395709, August 2025) reports a field experiment in entry-level customer-service hiring at one recruitment firm, where humans made the final hiring decisions. Read the paper for its outcome figures. (One firm, one country, one role family.)
The scoring layer has its own bias evidence, and it cuts the other way. In an audit by the University of Washington (31 October 2024), three large language models ranked more than 550 resumes against more than 500 job listings. White-associated names were favoured 85% of the time, against 9% for Black-associated names. (LLM resume ranking, not interviews.) It does show why "the model scored it" is never an answer on its own.
Who makes the decision?
Ask the buyer question plainly: can the system reject or advance a candidate with no human action, and is that enforced or a setting? In a 2024 survey commissioned by Resume Builder and fielded by Pollfish (n=948 business leaders, 22 October 2024; vendor-commissioned), of the companies using AI in hiring, 21% automatically reject candidates at all stages with no human review, 50% use AI-only rejection at the initial resume screen, and 29% keep human oversight for all rejections. (2024 context, not a current benchmark.)
Gates are architecture decisions. A gate can be a setting an administrator changes, or a rule the software enforces, and the two carry different risk. The stage-by-stage map of where the human approval gate sits is in the buyer's guide.
ZeaHire's answer to that question: candidates wait in Pending Review until a person approves or rejects them.
Is AI interview software legal, and can it be biased?
The answer depends on where the candidate and the employer sit. The state-by-state detail is in the tracker of AI hiring laws by state; the table below is the short version.
| Jurisdiction | What it requires | Status |
|---|---|---|
| NYC Local Law 144 | Annual independent bias audit of automated employment decision tools; notice to candidates 10 business days before use (NYC DCWP) | In force; enforcement began 5 July 2023 per NYC DCWP |
| Illinois | AI Video Interview Act: notice that AI will be used, an explanation of how it works, consent, and deletion of the video on request (ILGA) | In force since 1 Jan 2020 |
| EU AI Act obligations for recruitment | Recruitment and selection AI is an Annex III high-risk use; obligations apply from 2 December 2027 under Regulation (EU) 2026/1744, in force 27 July 2026 (EUR-Lex; Gibson Dunn summary); emotion inference in the workplace is prohibited under Article 5(1)(f) (EUR-Lex) | Obligations deferred to 2 Dec 2027; Article 5 prohibition applies since 2 Feb 2025 |
| Colorado | SB 26-189 (ADMT), signed 14 May 2026; detail on the Colorado page | Duties from 1 Jan 2027 |
On litigation, start with Mobley v. Workday. In the Northern District of California, the court granted preliminary certification of a collective action under the Age Discrimination in Employment Act on 16 May 2025. That is a court-record fact about procedure. The plaintiffs' claims are allegations, not findings, and no court has ruled on liability. As of 28 October 2026, the CourtListener docket shows a Fourth Amended Complaint filed on 10 July 2026 and a motion to certify a class filed on 14 September 2026. Other matters, with court, status and legal theory for each, are in the tracker of AI hiring lawsuits, which also owns the current schedule.
Bias in this context is a measurable property. A bias audit typically compares selection rates by group, often against the four-fifths benchmark (29 CFR 1607.4(D)); what a bias audit measures is selection outcomes, not intent. The buyer question is the practical one: can the vendor export selection rates by stage and group, and when was its last independent audit? This page makes no claim about what any particular product, ours included, reports.
The full two-sided evidence is on its own page: the advantages and disadvantages of AI in recruitment.
How common is it, and what do candidates think?
Start with how many companies use AI in hiring, which has its own page, and with the problem of definition, because the sources count different things. The headline figure from SHRM's State of AI in HR 2026 (1,900+ HR professionals, 3 April 2026) is that 39% of organizations have implemented AI in HR, with recruiting the top function at 27%. The same report found that 56% of HR functions do not formally measure AI success. (Counts organizations, not users or candidates touched.)
Candidates see it from the other side. The Greenhouse 2026 Candidate AI Interview Report (n=2,950, 1 May 2026; the vendor's own survey) found that 63% of job seekers have experienced an AI interview, up 13 points in six months. 70% of surveyed job seekers said they were never clearly told upfront, 38% withdrew from a process because of one, and 57% say disclosure should be legally required. Among those who completed an AI interview, 51% never heard back.
On trust, Gartner's press release (n=2,918, 31 July 2025) found that 26% of job candidates trust AI will fairly evaluate them (the release cites the 1Q25 survey of 2,918 candidates).
Why this data belongs on a buying page: candidate walk-away is the loudest counter-argument to the whole category. A candidate who withdraws is a lost applicant and a lost data point, so disclosure stops being legal housekeeping and becomes a funnel metric.
What about interview fraud?
Interview fraud, which includes proxy interviewing and impersonation, has its own page: interview fraud. For a buyer the useful question is narrow: what does a vendor detect, and what does it only infer? Gartner predicts that by 2028, one in four candidate profiles worldwide will be fake (Gartner, 31 July 2025). (A forecast, not a measurement.)
Ask what the tool reports and how it is labelled. Integrity signals generally mean observations from the session itself, such as multiple faces on screen, a face absent from view, liveness issues or suspected reading from a script. Ask whether each signal feeds the score or is reported separately, and what a reviewer is meant to do with it. A candidate using an AI assistant to prepare or answer is a different problem from impersonation, and candidates using AI in interviews has its own page.
How do you choose AI interview software?
For the short version, which AI recruiting platform to buy walks through the decision step by step. The detail follows here.
The six questions below resolve most of what a demo can show you. They are ordered by how much a weak answer matters. The last column maps each to the fuller scorecard.
The six questions to ask in the demo
| # | Question | Why it matters | Good answer (green) | Red flag (red) | Scorecard row |
|---|---|---|---|---|---|
| 1 | Show the evidence behind one score, from the candidate's actual answer. | A score with no answer behind it cannot be reviewed or challenged. | Opens the transcript and shows the quoted answer, the criterion and the reasoning. | Shows a number, ranking or "fit" label with nothing underneath. | Q1 |
| 2 | Who can reject or advance a candidate, and is that enforced or a setting? | It decides whether a person sees a candidate before a rejection. | Names each mode, who can change it, and that changes are logged. | "A human is always in the loop," with no answer on configuration. | Q2–Q3 |
| 3 | What is measured: answer content only, or voice, face, gaze, timing? Is any of it in the score? | Any signal in the score is part of the decision. | Lists every signal and says which feed the score and which are reported separately. | Calls it proprietary or will not list the signals. | Q4 |
| 4 | What does the candidate see before and after: notice, criteria, alternative, human-review route? | Most candidates were never clearly told (the 70% Greenhouse figure above). | Plain-language notice before, a route to a human or an alternative, and a way to request review. | Disclosure is a footer link. | Q7 |
| 5 | Can you export selection rates by stage and group, and when was the last independent bias audit? | A bias audit measures selection rates; without an export you cannot check. | Shows the export and names the auditor and date. | Makes bias claims with no audit and no export. | Q5–Q6 |
| 6 | How are impersonation and cheating handled: what is detected, what is only inferred? | Detection and inference are different claims and need different handling. | Separates what is observed from what is inferred, and says what a reviewer does next. | Claims to catch all fraud, or does not distinguish the two. | Q9 |
The full scored version has twelve questions: the 12-question vendor scorecard.
If you want a ranked comparison, see top AI interview tools compared. On cost, this page prints no price range: vendor prices change and need a dated source, and the HireVue and Spark Hire pricing pages carry them.
Staffing agencies and high-volume desks
Volume and client submission change the criteria. A staffing desk needs evidence it can put in front of a client, not just a score it can read itself, and it needs the economics to work per requisition rather than per seat. Everything else is on the page for AI for staffing agencies.
FAQ
What is AI interview software?
AI interview software conducts, transcribes and scores structured candidate interviews against criteria the employer sets, and returns evidence for a person to review. It differs from a video recorder, an ATS and an AI note-taker because the software itself runs the interview.
How does AI interview software work?
The employer defines criteria, the software asks structured questions, transcribes the answers and scores each against the criteria, and a person reviews the evidence. Integrity signals are recorded along the way. The stage-by-stage detail is in how AI interview software works.
Is AI interview software legal?
It depends on the jurisdiction. New York City, Illinois and the EU each set requirements, covering bias audits and notice, notice and consent, and high-risk obligations respectively. For the detail, see the state-by-state AI hiring rules.
Can AI interview software be biased?
Yes. Any model can be inconsistent, and bias evidence exists for AI scoring layers. A University of Washington audit found white-associated names favoured 85% of the time against 9% for Black-associated names, in LLM resume ranking rather than interviews (UW, 31 October 2024). Structured interviews give a consistent baseline to compare against. The pros and cons of AI in hiring are set out on their own page.
How is AI interview software different from video interview software?
Video interview software records answers for a human to watch. AI interview software evaluates the interview itself, which changes who scores, what gets recorded and what the candidate must be told.
How much does AI interview software cost?
Prices vary by vendor and need a dated source, so this page does not print a range. The current published figures, with dates, are covered on HireVue pricing and Spark Hire pricing.
Does AI interview software replace recruiters?
The software is built to cover the first-round interview while people keep the decision; in the Jabarian and Henkel field experiment, humans made the final hiring decisions (SSRN 5395709). In ZeaHire, candidates wait in Pending Review until a person approves or rejects them.
What is an agentic AI hiring platform, and is it the same thing?
"Agentic" describes software that plans and carries out multi-step tasks with limited prompting, so an agentic hiring platform may run several stages without being asked each time. An AI interview tool may or may not be agentic. The buyer questions are the same either way, and more is on agentic AI in recruitment.
See the evidence behind one score
See what the evidence behind one score looks like: see one in a demo. In ZeaHire, candidates wait in Pending Review until a person approves or rejects them.
Comments
No comments yet — start the conversation.