What a Greenhouse Voice AI Screen Records

A Greenhouse Voice AI screen is a structured, rubric-scored voice interview run by the employer's applicant tracking system rather than by a separate assessment vendor. It records audio and video of your answers, converts them to a transcript, and scores the transcript against criteria the employer defined in advance. According to Greenhouse's own Voice AI data processing FAQ, last updated August 28, 2026, the audio and video are "used only to generate that transcript" and are not themselves used to produce the score.

That distinction matters more than it sounds, and it is the opposite of what candidates usually assume about an AI video interview. This article covers what the round captures, which third parties touch it, what it flags beyond the score, and which of those things you can actually refuse.

Why this round appeared in the Greenhouse pipeline

Greenhouse agreed to acquire Ezra AI Labs, a voice-based AI interviewing startup, in May 2026 (HR Brew, May 5, 2026). The product now ships as Greenhouse Voice AI. In an October 2026 release announcing what it calls the Orchestrated Hiring Ecosystem, Greenhouse positioned Voice AI as a replacement for résumé-based triage: applicants get a first interview "with the same questions and criteria" instead of a résumé deciding who gets one (Greenhouse press release, October 2026).

The stated motive is volume. Greenhouse reported that applications per recruiter on its platform rose 412% since 2023, and that fewer than 7% of applicants receive an interview. Read that as the reason the round exists: it is a phone screen the employer could not otherwise staff, not a final-round evaluation. Greenhouse's own glossary entry, dated October 9, 2026, describes an AI interviewing agent as something that "typically stands in for an early-stage recruiter phone screen, not a final-round interview."

Because Voice AI lives in the ATS rather than in a separate assessment tool, the invitation may not name a vendor you recognize. If the scheduling link and the careers page are Greenhouse-hosted and the interview is a voice conversation with follow-up questions, this is the round you are in.

What a Greenhouse Voice AI interview records

Greenhouse's data processing FAQ lists the categories it processes. The ones that concern a candidate:

• Interview media: "audio and video recordings of interviews conducted through Voice AI."
• Transcripts generated from that audio.
• Derived artifacts: a Voice AI Score, question-level scoring rationale, stated strengths and growth areas, and separate cheat and fraud signals.
• Supporting data: identifiers, résumé content, scheduling data, and account and device data.

Two exclusions are stated explicitly in the same FAQ, and both are worth knowing because competing vendors do not all disclaim them. Greenhouse says Voice AI does not perform emotion recognition and does not extract biometric identifiers. It also says the score is derived from the transcript, not from the recording.

If that holds in practice, the practical consequence is that tone, pacing, pauses, and facial expression are not what the rubric reads. The words are. That is a narrower evaluation surface than the affect-scoring approach that older video interview platforms were criticized for, and it is a meaningfully different setup from the voice-first screens run by standalone vendors — see what a Ribbon AI voice interview records and scores for a comparison point outside the ATS.

How long the recording survives

Greenhouse's FAQ says recordings are retained "for the duration of the contract unless deleted earlier by you," where "you" is the employer. There is no candidate-facing expiry. The company adds its own editorial position: that transcripts and scores satisfy hiring-records obligations, and that keeping raw media longer "should be justified by a separate, documented purpose." That is a recommendation to its customers, not a limit on them.

So the realistic retention answer for a candidate is: as long as that employer keeps paying for Greenhouse, unless someone there deletes it or you ask. The mechanics of asking are covered in how to get your interview recording deleted.

Which companies process your audio

This is the part no marketing page leads with, and the FAQ names the chain. As of its August 28, 2026 revision, Greenhouse lists these subprocessor roles by example:

Text-to-speech, cloud hosting, and fraud and verification providers are referenced but not named.

Note what the LLM tier receives: transcript text together with the employer's rubric and ideal-candidate profile. The model is doing a comparison against a target, not an open-ended assessment. Greenhouse states that it has data processing agreements with each subprocessor prohibiting use of personal data for their own purposes, including model training, and that "Personal data isn't used to train Greenhouse's or any third party's models." Those are contractual commitments disclosed by the vendor; treat them as the company's representations rather than as independently verified facts.

What it flags that is not part of the score

The FAQ lists cheat and fraud signals as a separate output from the Voice AI Score, and says those signals are decision-support only and are not part of the score. Greenhouse's glossary adds that Voice AI "flags signs of scripted or AI-generated answers."

Both halves of that matter. A scripted-answer flag does not mechanically reduce a score, but it does reach a human reviewer as a note attached to your interview. In practice this is the same category of signal as the integrity flags that assessment platforms attach to coding tests, and it behaves the same way: a flag is an invitation for someone to look harder, and what happens next depends entirely on that employer's process. What triggers a fraud flag on your job application covers how those signals usually get handled upstream of an interview.

The direct implication for how you approach this round: reading prepared answers aloud is the specific behavior the round is instrumented to notice. Rehearsed structure is fine. Recited text is the thing with a detector pointed at it.

Who sees the result, and who can override it

Scores, transcripts, and rationale are surfaced, in the FAQ's wording, "to authorized customer personnel in the customer's Greenhouse account." That is the hiring team, not the public, and not other employers using Greenhouse.

Two process details are unusual enough to be worth knowing:

• Overrides are logged. Reviewers can override a Voice AI score, and the FAQ says each override requires a written rationale. A human can disagree with the machine, and has to say why.
• You can request human review. The FAQ says candidates can request human review through the post-interview email flow. That is a built-in path, not a favor you have to negotiate.

Greenhouse support staff cannot see candidate data unless the employer grants what the FAQ calls a Temporary Access Grant.

What you can decline

The FAQ states that applicants can decline Voice AI, and that an applicant who declines "isn't penalized, disadvantaged, or subject to any adverse inference." It also makes clear where the obligation sits: the employer, not Greenhouse, must define and run the alternative process.

That is the catch worth planning around. The no-penalty language is the platform's commitment; the alternative round is the employer's to build, and the quality and speed of it will vary by company. Greenhouse also requires customers to give notice before the session covering AI interviewing, recording, transcript analysis, human final decisions, and candidate rights, with a transparency notice on the pre-interview landing page and a banner during the interview. If you did not see those, the employer skipped a step its vendor told it to take.

For how to phrase the request and what a reasonable alternative looks like, see what you can actually request when you are sent an AI interview.

Where the rules are still moving

Two legal markers sit underneath this round, and neither is settled enough to rely on blindly.

Illinois. The Artificial Intelligence Video Interview Act, effective January 1, 2020, requires employers to tell candidates that AI will assess the interview, explain how it works and what characteristics it evaluates, obtain consent, limit sharing, and destroy the interview within 30 days of an applicant's deletion request, including copies held by others. Greenhouse's FAQ says it deletes audio, video, transcripts, and scores "within 30 days of a customer-initiated deletion request" and instructs its model providers to do the same. Read the seam carefully: the statutory right runs against the employer, and Greenhouse's clock starts when the employer asks. Illinois also amended its Human Rights Act effective January 1, 2026 to address AI in employment decisions.

European Union. Greenhouse expects Voice AI to be high-risk under Annex III of the EU AI Act. The application date for Annex III systems, which include recruitment tools, was moved from August 2, 2026 to December 2, 2027 under a provisional political agreement reached in May 2026 (Gibson Dunn). Greenhouse's FAQ flags the same date as subject to Official Journal publication. The practical reading for a candidate in 2026: the disclosure you get rests on the employer's policy and on state law, not yet on EU high-risk obligations.

Greenhouse says bias audits run on every Voice AI algorithm change, with monthly third-party audits by WardenAI published at trust.warden-ai.com/greenhouse/ai-interviewer. That is a vendor-reported arrangement; the published results are the thing to read if the audit outcome matters to your decision, and this article did not verify them.

How to prepare for a transcript-scored voice screen

The scoring surface tells you what to do. If the rubric reads the transcript against a stated ideal-candidate profile, then specificity in words is what converts, and polish in delivery is not where the marginal point comes from.

• Say the concrete noun. Name the system, the metric, the team size, the timeframe. A transcript keeps "cut p99 latency from 800ms to 120ms" and loses "significantly improved performance."
• Answer the question asked, then stop. The agent asks follow-ups. Leaving the obvious follow-up available is better than pre-empting it in a monologue.
• Expect to be able to interrupt. Ezra's founder described candidates being able to interview the agent back about the role, compensation, and benefits. Treat the time as two-way.
• Do not read. The scripted-answer detector is the one documented behavioral flag in the round.

That last point also sets the honest boundary on live assistance here. Control is a desktop AI interview assistant for live technical interviews, coding assessments, and screen-share workflows on Windows and macOS; a transcript-scored behavioral voice screen with a scripted-answer flag is not the round it is built for. Reading generated answers into a detector pointed at exactly that is a bad trade, and the preparation above is the better use of the hour.

Where live help does fit is the round after this one. If the Greenhouse screen passes you to a technical interview or a shared-screen coding exercise, that is a different setup with different mechanics, and Control's free allowance of five messages and two minutes of voice transcription is enough to see how it behaves before you commit to a pass. Prices as of October 19, 2026 are $12 for a day pass, $19 for a 72-hour sprint pass, and $39 monthly.

The short version

A Greenhouse Voice AI screen records your audio and video, scores the resulting transcript against the employer's rubric, attaches separate fraud and scripted-answer signals for a human to read, and keeps the recording in the employer's account for the life of their contract. You can decline it, you can ask for human review through the post-interview email, and in Illinois you can demand destruction within 30 days. What you cannot do is treat it as a formality: fewer than 7% of applicants reach an interview on this platform by Greenhouse's own count, and this is now the interview.