All sixteen AI interview agents, compared

Five of sixteen AI interview tools can act on a candidate with nobody checking. Eight publish a current outside bias audit. HireVue’s published audit is from 2023. Use this grid to check a demo against what each vendor publishes.

By Ashish Manchanda Published 23 Sep 2026 Last updated 25 Sep 2026 Sixteen products, side by side

Every product but Tengai can follow up on a vague answer

All sixteen tools interview by video, voice, phone or text. Every answer here opens with a short verdict, such as Yes, No, Not stated or a count. Not stated means the vendor has not published an answer. The sentence that follows gives the detail.

How each interview runs · sixteen products · read 19 Sep 2026
Product How it runs, and how long Languages Does it follow up? What it reads Who writes the questions and the scoring
HireVue By voice. You pick the time. HireVue gives no length. 4 named languages, and more. French, English, Spanish and Portuguese are named. HireVue adds “and more.” Yes. HireVue says it “asks, probes, and listens.” Transcript only. HireVue says it does not read “facial expressions… or tone of voice.” AI drafts both, from the job description. You approve before it goes live.
Eightfold By video, audio or text, in your browser. Eightfold gives no length for a standard run. 22 languages or more. Yes, on an incomplete answer. Eightfold says “it asks a follow-up.” Words only. Eightfold says it does not read “facial expressions, tone of voice, or accent.” AI drafts both, from the job you post. You approve before it goes live.
Talview Ivy By voice and text. You pick the time. Talview gives no length. Not stated. Talview says “multiple languages.” Yes. Talview calls these “intelligent follow-ups” on a vague answer. Words, plus “non-verbal cues.” Get Talview to say in writing whether that includes emotion. Ivy drafts both, inside rules you set.
Apriora By video, phone, SMS or WhatsApp, any time of day. Apriora gives no length. 26 languages or more. Yes. Apriora says it follows up on “what the candidate actually says.” Eye movement, tab switching and background voices, plus the transcript. Apriora does not say it reads emotion. AI drafts both, from the job description. You run it.
HeyMilo By phone, video, SMS or forms. HeyMilo sets this up differently for each role, and gives no length. 14 languages or more. Yes. HeyMilo says it “listens, follows up, and keeps full context.” Voice and video, to check for cheating. HeyMilo does not claim to read how a candidate feels. AI drafts the questions, through what HeyMilo calls an “AI Agent Builder.” Recruiters set the scoring.
Humanly By video avatar, or audio only. Humanly gives no length. Not stated. You can set a language for each question. Optional. You turn this on or off per interview. Transcript only. Humanly makes no claim about reading tone or facial expressions. You or your integration partner write both, based on the job description. Humanly has no question bank of its own.
Ribbon By phone or in your browser, voice or video. Five questions by default, and no length given. 10 languages or more. Yes. Ribbon says it “follows up when an answer needs more detail.” The sound around the call, timing, speech patterns and attention. Ribbon makes no claim about reading a candidate’s mood. You pick or write the questions. Scoring is set by default, by you, or by what Ribbon calls “AI magic.”
Greenhouse Voice AI By voice. The candidate picks the time. One case study interview ran 25 minutes. 11 named languages, including English, Hindi, Spanish, German, French and Japanese. Yes. Candidates can also ask it questions back. Transcript only. The contract rules out reading emotion. It also rules out biometrics, like using your face or voice to identify you. Recruiters write both, before any candidate is seen.
Metaview Screening In your browser, before a deadline. Metaview calls it “short” but gives no minutes. 18 languages. Yes, on a vague answer. Candidates can also ask it to repeat a question. Transcript only. If you turn the camera on, it also “watches for” a second screen. AI drafts both, from the job description. You can edit both.
Employment Hero By video avatar, spoken aloud. 15 to 30 minutes, built from four to six core questions. Not stated. Yes, on a short answer, to draw out a real example. Audio and video, not just the transcript. AI drafts both, from the job description, automatically. The recruiter can edit both.
CodeSignal By phone, browser voice, or text. About 15 minutes, and you can change that by role. 20 languages or more, for live phone interviews, as of Dec 2025. Yes, on a vague answer. CodeSignal says “it probes and asks follow up questions.” Voice or text content only. CodeSignal makes no claim about reading facial expressions or tone. You set the criteria and tone, or start from a named template agent.
HackerRank Chakra By voice, from one link. HackerRank calls it “voice-based” on one page and “voice/video” on another. Not stated. Yes, in real time. HackerRank says it follows up “based on candidate responses.” Transcript only, for scoring. HackerRank checks the webcam separately, not for scoring. An “Interview Creator Agent” drafts both. You refine it.
Glider AI By video. An admin sets the length: 30, 45 or 60 minutes. 35 languages or more, with live translation offered for others. Yes. Glider calls these “intelligent follow-up questions tailored to each candidate’s answers.” Answer content only, scored against a set guide. Glider checks the face and voice only to catch fraud. AI drafts both, from the job description. An admin adjusts the skills and scoring.
Maki Mochi By phone or browser. About 10 minutes. 5 named languages, with more on request. English, French, Spanish, German and Italian are named. Yes, on a vague answer. Maki calls this unscripted. Words only, for scoring. A separate feature can add voice and gaze checks. Mostly Maki itself. Recruiters pick from a ready-made library and cannot write custom questions.
Spotted Zebra A live, spoken conversation, on any device, any time. Spotted Zebra gives no length. Not stated. Yes. Spotted Zebra calls these “dynamic follow-up questions that probe deeper.” Not stated. No page says it reads tone, face or emotion. Its AI, called Spot, drafts both from your skills profile.
Tengai By video avatar, in your browser. About 20 minutes. English and Swedish, spoken. Since September 2026 it translates questions into Norwegian and Danish too. No. It only rephrases a question if it misunderstands the candidate. It keeps to one script. Words only. Tengai uses no facial recognition, by design. Recruiters write the questions, since its 4.0 update. The core scoring stays fixed.

Tengai keeps every candidate on the same script, by design, and never adapts. Humanly lets you switch follow-up on or off for each interview.

HeyMilo, Humanly, Maki, Apriora and Metaview can act on their own

Every vendor says a recruiter makes the final call. These five also have a setting that acts without one. Tengai’s own pages disagree on this point.

What each one checks and tells the candidate · sixteen products · read 19 Sep 2026, reconciled 25 Sep 2026
Product Checks who is answering Checks for live AI help Can it act without a recruiter? Published bias audit? What the candidate is told What it connects to What comes out
HireVue Not stated. HireVue folds this into what it calls an “integrity layer.” Not stated. No. HireVue says the recruiter should “always make the final call.” Published, but old. HireVue says its models get yearly independent audits. Its published audit dates from 2023. Yes. Candidates agree first. HireVue names an opt-out and a way to ask for accommodation. 45 or more systems, including Workday, SAP SuccessFactors, Oracle and Greenhouse. It sends scores back to them. A recording, transcript and score.
Eightfold Yes. Eightfold checks the applicant through the identity services ID.me and CLEAR. Yes. It tracks “12+ cheating signals,” including tab switching and duplicate voices. No. “The hiring team makes the hiring decisions,” Eightfold says. Yes. Audited under New York City’s Local Law 144, and passed. Yes. Candidates agree to recording and an ID check. Get Eightfold to state whether a candidate can ask for a human. Oracle Recruiting Cloud by name. For other systems, Eightfold says only “your existing ATS,” meaning applicant tracking system. A summary, video and transcript.
Talview Ivy Not stated for Ivy. Other Talview products check identity by face, voice and device, and flag deepfakes. Not stated for Ivy. Deepfake checks live in a related Talview product. No. Talview says its services “do not qualify or disqualify anyone.” Not stated. No outside audit is published for Talview. Yes. Candidates agree each session. Talview describes no human option. Workday and Greenhouse, confirmed both ways. Other systems are listed for the wider Talview platform, not confirmed for Ivy. A recording, transcript and scores.
Apriora Yes. It matches a live selfie to a government ID during the interview. Yes. Its Proctor tool rates the chance that a second person is helping, or that an answer is scripted. Yes, to advance a candidate. Its Coordinator books the next interview once a candidate clears your bar. Apriora says it does not auto-reject. Yes. Warden AI audits it every month. Yes. Candidates are told what Alex, Apriora’s AI interviewer, is. They get no feedback and no human option. Workday, Greenhouse, Bullhorn and iCIMS. It sends results to them. A recording, transcript and scores.
HeyMilo Yes. It matches the candidate’s voice as ID, what HeyMilo calls “voice authentication.” Yes. HeyMilo runs “AI-generated-answer detection” to catch coached responses. Yes. A score can decide whether a candidate goes further. It can move them on or archive them in Workday. Its rule builder does the same in Bullhorn and Avionté. In iCIMS, SAP SuccessFactors and SmartRecruiters, a score only moves a candidate forward. That is six systems. Someone on your team has to switch this on. Yes. Warden AI audits it every month. Yes. Candidates are always told it is AI. HeyMilo logs their consent, and they can take it back. Nine named tracking systems get results back, including Workday and Greenhouse. HeyMilo lists more than 20 partners in all. UKG Pro only sends data in. A recording, transcript and score.
Humanly Not stated. Not stated. Yes. A score you set can decide whether a candidate goes further, with no one reviewing it. Not stated. No outside audit is published for Humanly. Yes. Candidates are told they may be recorded. Not taking part is the only way to opt out. Mainly Greenhouse, through one linked field. Links to other tracking systems are still being built. Audio, a transcript and a score.
Ribbon No. Ribbon describes no identity check like this. Yes. It flags text-to-speech tools, a second voice, and patterns that look “AI-generated.” No. The recruiter casts what Ribbon calls the “hire/no-hire vote.” Yes. Holistic AI audits it every year, as its data contract states. Yes. Candidates agree before recording starts. In California and New York, they must also be offered an opt-out. 60 or more systems, through a tool called Kombo, including Workday, Greenhouse and SAP. It sends scores back to them. A recording, transcript and score.
Greenhouse Voice AI Not built in. A separate Greenhouse product handles ID checks. Yes. Built-in cheat detection flags and timestamps signs of “outside help.” No. The contract says a score can never decide alone. Yes. Warden AI audits it every month. Yes. Candidates are told an AI is interviewing them. They are told it is recorded and scored, and offered another option. 60 or more systems, including Workday, SAP and iCIMS. Jobs and scores sync both ways. A recording, transcript and score.
Metaview Screening Not stated. Screening has no ID check or deepfake detection. Yes. Metaview reads the transcript for signs of a scripted or “externally assisted” answer, and shows this as a risk level. Only for a missed deadline. Metaview never auto-rejects a candidate on the rating. One setting, off by default, rejects a candidate who misses the deadline. Yes. Warden AI audited it, most recently on 17 Aug 2026. Yes. Candidates are told directly it is AI. An opt-out link exists but is off by default. 34 named systems. Decisions sync back to all of them except Crelate and Workday. A recording, transcript and rating.
Employment Hero No. Its pages and help centre describe no identity check. No. Its pages describe no cheat or fraud detection. No. Employment Hero says it “does not make hiring decisions autonomously.” Self-reported only. Employment Hero tests its own model. No outside auditor is named. Yes. Candidates are told upfront it is AI. They can ask for a human to review the result. Its own tracking and HR system only. No outside tracking system is named for this feature. A recording, transcript and a score for each question.
CodeSignal Yes. It matches a government photo ID to the person on camera. Not confirmed. A “Suspicion Score” flags patterns that look AI-assisted. CodeSignal built it for tests, not confirmed for interviews. No. Its contract bars using the product to automate hiring decisions. Not published. CodeSignal offers an adverse-impact study during a pilot instead. Not stated. Get CodeSignal’s candidate notice and consent screen before you sign. Workday, iCIMS and Oracle, only on its top plan, which has no public price. Its Build and Grow plans get basic tracking-system links. A transcript, a skills report and a recommendation.
HackerRank Chakra Partly. The webcam flags a “secondary face” during the interview. Chakra has no ID-document check of its own. Yes. Screenshots flag “external AI assistants” such as Cluely, an AI cheating tool. Its Desktop App Mode blocks two named AI cheating tools, only when switched on. No. “Hiring decisions stay with your team,” HackerRank says. No audit of Chakra. Two BABL AI audits from July 2024 cover its plagiarism check and its webcam check. Yes. Candidates must acknowledge AI is being used. One mode adds a list of actions they cannot do. Workday, Ashby and Greenhouse, confirmed both ways. HackerRank also names Eightfold. A report, score and transcript.
Glider AI Yes. A photo is matched to an ID and to earlier attempts. Its “Audio Genuineness” check flags a second voice. Not stated. Nothing is named for catching live AI help during the interview. No. Nothing in Glider’s pages describes an automatic reject. Not stated. No outside audit is published for Glider AI. Yes. Candidates must “agree to the disclosure,” or the interview ends automatically. Named for SAP, Workday and Greenhouse. Get written confirmation that results sync back. A transcript and a scored report. A summary and a hire recommendation are off by default.
Maki Mochi Not confirmed for Mochi. A face and device check exists on the wider Maki platform. Not confirmed for Mochi. A speech-pattern check exists on the wider Maki platform. Yes, through a setting. Recruiters can set filters that reject a candidate automatically. Maki’s product page also says no candidate is ever auto-rejected. Both statements are on Maki’s site. The filter is real, and it is a setting you choose. Yes. Holistic AI audited it in March 2026. Yes. Candidates agree first. Accommodation is available on request. Whether a human option exists depends on local law. 30 or more systems advertised, including SAP and Workday. Results sync back automatically. A score, transcript and status.
Spotted Zebra Not stated. Spotted Zebra claims “cheat detection… built in.” Get it to name its identity check. Not stated. Get Spotted Zebra to name its check for live AI help. No. “Spot surfaces evidence. Humans make decisions,” Spotted Zebra says. Yes. Dated 10 Sep 2026. Not stated for this interview. Its general AI principles promise disclosure, but it has no notice page specific to this product. Unnamed major systems, through Kombo. Status and results sync both ways. A transcript, summary and a ranked score.
Tengai No, by design. Tengai keeps “no access to candidates’ personal data” at all. No. Depends which page. One Tengai page says it can screen and recommend candidates with no human interference. Other pages say a person decides. Not stated. No outside audit is published for Tengai. Yes. Candidates are told the avatar is just an interface, and they get written feedback right after. Its named systems are Teamtailor, Zerolime, Jobylon, Talenttech and Varbi. Its recommendations sync back to the tracking system. A shortlist, scores and an index.

Before your first call, get each of these five and Tengai to show that setting in their own terms.

“Not stated” means unpublished, not missing

The feature may still exist. For a few questions, the check sits on a vendor’s wider platform. Those answers say it is not confirmed for this interview tool.

Two questions move a shortlist more than the rest. “Can it act without a recruiter?” tells you if an automatic switch exists, and who owns it. “Published bias audit?” tells you which vendors have already had their tool checked by someone outside the company. In New York City, the bias-audit and candidate-notice duty falls on the employer, not the vendor, for every tool here.

Two guides split these sixteen by company size

Click any name here to open its full write-up. It carries the contract language behind every answer.

For a shortlist sized to your company, read the guide for mid-market companies or the guide for enterprise companies. For the argument behind this grid, read the guide to AI interview agents.

Sources and method Every answer here comes from the feature sheet built for this market, which draws on each product’s own pages, help centre, developer documentation and legal terms, read directly on 19 September 2026. The answers on whether a tool can act on a candidate without a recruiter, and whether it publishes an outside bias audit, were checked again against this market’s reconciled vendor files on 25 September 2026. We did not summarise, rank or guess beyond what those sources state. No answer is left blank. Where a vendor has not published an answer, it says “Not stated.” The market guide carries the fuller note on how this research was done. Two products you may meet in this market, Ashby’s AI Interviewer and Indeed Smart Screening, are not included here. Neither one can be licensed yet.