Short answer: Greenhouse can run an AI voice pre-screen in two ways. Greenhouse now sells its own optional Voice AI, and any other provider can plug in through the Assessment Partner API, the Harvest API (v3) and webhooks. Whichever route you pick, the pre-screen should end as a score, a transcript link and a short rationale inside the Greenhouse application, and a human should make the call. This page shows where each piece fits, what Greenhouse documents, and what to settle with your data protection officer first (status 18 October 2026).

What do people mean by a "Greenhouse video interview"?

Most teams are not attached to video. They want an async first screen that every applicant takes in the same way, that costs no scheduling, and that leaves evidence in the application. Historically that step came from partners in Greenhouse's integration directory, such as Spark Hire or Hireflix. In May 2026 Greenhouse announced the acquisition of Ezra AI Labs, a voice AI interviewer, and now offers it as Voice AI. According to Greenhouse's data processing FAQ, it asks recruiter-configured questions, records audio and video, transcribes the answer and compares it with a rubric you define.

So the honest starting question is not "does Greenhouse support this?" but "native feature, partner tool or own build?" The table helps you decide.

How does a voice pre-screen fit into Greenhouse's structure?

Greenhouse organises hiring around the interview plan: stages per job, an interview kit per interview (instructions, questions, focus attributes, résumé, scorecard) and a scorecard that the interviewer fills in after the conversation, with an overall recommendation from Definitely Not to Strong Yes. Two consequences matter for a pre-screen:

• A scorecard belongs to an interviewer. It is a human's feedback. An AI score is better kept apart from it, for example as an assessment result or a clearly labelled note, so nobody mistakes a machine score for a colleague's rating.
• The stage is where the decision happens. The pre-screen should sit in its own stage, early in the plan, so that reporting shows how many applicants passed it and where they dropped out.

Which Greenhouse interfaces do you need, and for what?

This is where most "ATS × AI" pages stay vague. Greenhouse documents four separate interfaces, and they do different jobs:

Assessment stage: what actually happens

1. The provider gives the customer an API key, which is handed to the Greenhouse account manager. After that an assessment stage of that partner can be added to the hiring plan of eligible jobs.
2. Greenhouse calls list_tests to show the available tests, then send_test when a recruiter clicks "Send Test". The request carries candidate and application IDs and the Greenhouse profile URL. The résumé link in it expires after seven days.
3. The provider returns a partner_interview_id and invites the candidate.
4. When the candidate has finished, the provider calls the callback URL (PATCH). Greenhouse then reads test_status: partner_status "complete", partner_profile_url, an optional partner_score and flat metadata.

Polling still exists but is marked as deprecated for new integrations; the callback is the recommended way. Status updates only work for active, unhired candidates. All calls use HTTP Basic Auth over HTTPS (Greenhouse documentation).

Webhooks and Harvest: the custom route

Webhooks arrive as signed JSON (HMAC-SHA256 over the raw body). Greenhouse retries up to seven times over roughly 15 hours and disables a webhook whose test ping fails (webhook documentation). Write-backs go through Harvest. Greenhouse states that Harvest v1/v2 was to be removed on 31 August 2026, so any integration you buy or build must use v3. In v3 you can create scorecards, add attachments, create notes and move an application; a move requires the current from_stage_id, which protects against stale updates (Harvest v3).

What should be written back into Greenhouse?

Less is more. A pattern that works for recruiters who read dozens of applications a day:

• One overall score with the scale stated, shown as the assessment result.
• A link to the full report (transcript, quotes behind each criterion) via the partner profile URL.
• Two to three lines of rationale as metadata or a note.
• No automatic rejection. A move to "Rejected" stays a human click.

Mapping the AI result onto a scorecard is possible technically, but it blurs who rated what. Decide this deliberately with your hiring managers.

How does the Sprad voice interview work with this?

Sprad's Screening lets applicants chat or talk with Atlas instead of filling in a form. Interviews are audio first, voice or video can be set per role, and the finished report is meant to land in the ATS you already use. According to Sprad's AI ethics page, identity is removed before scoring, each criterion is scored three times with the median kept, every score quotes the sentence it came from, and there is no camera, face or emotion analysis and no automatic rejection.

What we cannot tell you from public pages: Sprad's product page names Personio, SAP SuccessFactors and Workday as ATS connections, but not Greenhouse. The connection is booked as an ATS integration (100 credits per connection, see pricing). Whether it runs as an assessment stage, via webhooks and Harvest, or in another way depends on your set-up, so clarify this before you plan a pilot.

What does it cost? A worked example with stated assumptions

Prices from sprad.io/pricing, status October 2026, net. Per application: CV review 2 credits, voice interview 30 credits. Flat option: 0.49 EUR per application including CV screening, plus 0.20 EUR for voice interviews for every fifth applicant.

The credit route is cheaper when you interview few and selectively; the flat rate wins when you want predictable cost or interview more. Atlas Base itself costs 450 EUR a year (6,000 credits included), which is not in the table. Credits cost 0.10 EUR at list price in the free plan.

Is a voice pre-screen allowed in the EU and DACH?

Yes, with conditions. This is general information, not legal advice. Our longer guide on what AI candidate interviews allow in the EU goes deeper; these are the anchors.

• GDPR Article 22. People have the right not to be subject to a decision based solely on automated processing that significantly affects them (GDPR). The Court of Justice held in the SCHUFA case (C-634/21, December 2023) that even a score can count as a decision if a third party draws strongly on it. A score that recruiters follow without looking is therefore risky; a score with real human review is much safer.
• EU AI Act. Systems used to recruit or select people, in particular to filter applications and evaluate candidates, are listed as high risk in Annex III, point 4. Employers then have deployer duties such as human oversight, informing affected people and keeping logs. The original application date was 2 August 2026; Regulation (EU) 2026/1744 (AI Omnibus, in force since 27 July 2026) moved the Annex III obligations to 2 December 2027. Preparing now still pays off, because GDPR Art. 22 and transparency duties already apply.
• Germany. In companies with a works council, technical systems that can monitor behaviour or performance trigger co-determination under section 87 BetrVG, and selection guidelines under section 95 BetrVG explicitly apply when AI is used. The AGG still requires job-related, non-discriminatory criteria.
• Your duties stay with you. Greenhouse says about its own Voice AI that notice, consent, an alternative process and legal compliance are the customer's responsibility. The same applies to any partner tool.

How do you roll it out in four weeks?

1. Week 1, scope and law. One role with high volume. Involve data protection and the works council. Decide the legal basis, retention period and the alternative for candidates who do not want an AI interview.
2. Week 2, questions and rubric. Three to five evidence questions, five criteria at most. Example: availability and shifts, one concrete example of similar work, one scenario.
3. Week 3, connection. Agree stage name, trigger (new application or stage change), what is written back, and who gets access. Test with internal applicants only.
4. Week 4, calibration. Listen to a sample from every score band. If good candidates score low, change the questions, not the threshold.

Typical mistakes: a generic "tell me about yourself" question, a threshold that rejects automatically, an AI score stored as a scorecard, and a pilot without a plan B for candidates who cannot or do not want to speak in the interview language. If accents or technical vocabulary are a concern, read our piece on AI interviews, German accents and technical vocabulary.

When is it not the right step?

• Roles where a portfolio, a work sample or a licence says more than a conversation.
• Very low volume, where a five-minute call is cheaper than the set-up.
• No agreement on what a good answer looks like. Then the score only automates disagreement.
• No time for the works council and data protection steps. Start there, not with the tool.

For a comparison of the formats themselves, see async video interview software compared and our overview of AI voice interview software.

FAQ

Does Greenhouse have a built-in AI voice interview?

Yes, as an optional feature called Voice AI, which Greenhouse took over from Ezra AI Labs in 2026. It is separate from partner integrations. Plans and regional availability are not listed in the data processing FAQ, so ask your Greenhouse contact.

Can I use any voice interview tool with Greenhouse?

Not automatically. The assessment stage works for partners that Greenhouse has set up. Anything else needs a custom integration through webhooks and the Harvest API, or a manual step with a link.

Where do the results appear?

With an assessment partner, score, status and a link to the partner profile are shown on the application. With a custom integration you choose: attachment, note or scorecard fields. A note or attachment keeps AI output apart from human feedback.

Does the Harvest API change affect me?

Yes, if an integration still uses v1 or v2. Greenhouse's documentation states those versions were scheduled for removal on 31 August 2026. Ask every vendor which version it uses.

Is a score alone allowed to reject someone?

That is exactly what you should avoid. GDPR Article 22 restricts solely automated decisions, and Greenhouse's own terms for Voice AI prohibit using the score as the sole basis for an employment decision.

Voice or video: which is better for a first screen?

Voice is lighter for candidates and collects less visual data. Video adds little signal for most first screens. Choose by role: customer-facing roles can justify video later in the process.

What does Sprad cost for this?

Credits: 2 per CV review and 30 per voice interview. Flat: 0.49 EUR per application, plus 0.20 EUR for voice interviews for every fifth applicant. An ATS connection costs 100 credits. Details on the pricing page, status October 2026.