Voice AI Interview Buyer Checklist: The VOICE Stack

By Brendten Eickstaedt — May 14, 2026

Voice AI interviewing is now the ATS default in 2026. Use the 12-point VOICE buyer checklist to separate operator-grade vendors from regulatory traps.

Voice AI interviewing went from pilot to platform feature in one quarter. Greenhouse acquired Ezra AI Labs. SeekOut shipped Sam. Every ATS roadmap now has "AI screening" on it. The pitch is the same everywhere: structured, scalable, fair, fast. The reality is that most buying teams cannot tell which products are defensible and which are a regulatory landmine in a slick UI.

In Brief:

• Voice AI interviewing and AI screening are now the contested layer in the AI hiring stack. Greenhouse's Ezra acquisition and SeekOut Sam reset buyer expectations in the same month.
• The vendors converge on the same talking points: explainability, audit trail, opt-out, bias monitoring. The execution gap between vendors is enormous and shows up in artifacts, not slide decks.
• Use the VOICE checklist to evaluate any voice AI or AI screening tool: Validation, Output evidence, Interview structure, Candidate transparency, Execution controls.
• Demand a worked example for every promise. If the vendor cannot show you a sample transcript, rubric, score breakdown, and audit log on a real role, the feature does not exist in production.
• The cheapest reduction in regulatory exposure is contractual. Require vendor cooperation on bias audits, NYC LL144 disclosure formats, and Colorado AI Act risk assessments in the MSA.

Why this checklist now

Greenhouse's Ezra AI Labs acquisition made voice AI a default expectation for an enterprise ATS. The pitch is structured, on-demand, AI-led voice interviews with transcript, score, and explainability built in. SeekOut Sam landed in the same window with a near-real-time screening flow that promises decision logs for every evaluation, alignment with EEOC and OFCCP guidance, SOC 2 Type II, and a posture that candidates are never auto-rejected.

The buyer problem is not whether to consider voice AI interviewing. It is which version of voice AI interviewing has the artifacts that survive an audit, an EEOC charge, or a NYC LL144 bias audit notice. Most demos look identical. The 12 questions below are how you separate them.

How to run this checklist

Bring the checklist to every demo. For each item, ask the vendor to show, not tell. Score each item as Pass, Conditional Pass, or Fail. A Conditional Pass needs a written follow-up commitment in the MSA. A Fail on any item under Validation, Output evidence, or Candidate transparency is disqualifying, full stop. Tie the result to a single owner on your side, usually the head of TA operations or the GC's HR lead, and re-run the checklist annually.

The VOICE Control Checklist

Validation

1. Job-specific rubric on file

• Ask: Show me the rubric your system used to score the last 10 candidates for a real role.
• Good answer: A role-specific rubric the customer authored, mapped to job-related competencies, with weightings.
• Red flag: A generic "personality" or "soft skills" rubric, or any rubric the vendor wrote with no customer input.
• The move: Require the rubric to be authored or signed off in writing by the hiring manager and the recruiter. No generic templates.

2. Bias audit, dated and scoped

• Ask: Show me the most recent bias audit report by independent auditor, scoped to the model and role family I am buying.
• Good answer: A report within the last 12 months, audit firm named, methodology cited, results disclosed for required protected categories under NYC LL144 and beyond.
• Red flag: "We do bias audits" with no documentation, or audits scoped only to internal applicant data with no demographic adjustment.
• The move: Require quarterly or at minimum annual bias audits, scoped to the configured rubric, with audit rights to review the raw report.

3. Bias monitoring in production

• Ask: How do you detect drift in scoring outcomes across demographic categories between formal audits?
• Good answer: A monthly statistical monitor with thresholds and an automatic flag to the customer.
• Red flag: "We retrain regularly" with no monitoring artifact.
• The move: Negotiate a 30-day notification window if drift exceeds threshold.

Output evidence

4. Full transcript on every interview

• Ask: Show me a complete transcript from a real recent interview.
• Good answer: Verbatim transcript with speaker labels, timestamps, and a downloadable export.
• Red flag: Summaries only, or transcripts that strip non-substantive responses.
• The move: Require transcript retention controls aligned to your records policy and a published export format.

5. Score breakdown tied to evidence

• Ask: Show me how a 7 out of 10 on competency X is calculated for a real candidate.
• Good answer: Per-question scoring, weight per competency, and evidence quotations from the transcript.
• Red flag: A single overall score with no decomposition, or an opaque "AI fit" rating.
• The move: Require the evidence-linked breakdown to be visible to the recruiter and exportable to the ATS.

6. Audit log of every decision

• Ask: Show me the audit log for any candidate where the system advanced, held, or returned the file to the recruiter.
• Good answer: A timestamped log with model version, rubric version, recruiter actions, and any overrides.
• Red flag: Logs scoped only to logins or admin events, not decisioning.
• The move: Confirm log retention aligns with the statute of limitations in your jurisdictions (often 4 to 6 years).

Interview structure

7. Same questions and rubric across candidates

• Ask: Confirm every candidate for the same role gets the same set of structured questions and is evaluated against the same rubric.
• Good answer: Yes, with a configuration screen showing the locked question set.
• Red flag: "We dynamically adapt questions per candidate" with no consistency commitment.
• The move: Lock the question set per role. Adaptive follow-ups are fine; the core question set is not.

8. No auto-rejection on AI score alone

• Ask: What is the system's behavior if a candidate scores below your suggested threshold?
• Good answer: The recruiter sees the score and decides. No automatic disposition.
• Red flag: An auto-reject toggle that some customers enable.
• The move: Disable any auto-reject pathway in your configuration and require it be off by default in the MSA.

Candidate transparency

9. Disclosure that AI is being used

• Ask: Show me the candidate-facing notice that AI is being used and what it is measuring.
• Good answer: Plain-language notice surfaced before the interview starts, archived for each session.
• Red flag: Buried in a 12-page privacy policy.
• The move: Customer controls the disclosure copy. Vendor logs the acknowledgment.

10. Opt-out path that does not penalize

• Ask: What happens to a candidate who declines the AI interview?
• Good answer: A human path of equal weight, no demerit in the rubric.
• Red flag: Opt-outs auto-route to a slower queue or get a lower default score.
• The move: Require equal pathways in writing and audit the opt-out funnel quarterly.

Execution controls

11. ATS sync and role-by-role gating

• Ask: Show me how I turn AI screening on for one job family and off for another.
• Good answer: Role-level toggles, ATS-native objects, clean state on each role.
• Red flag: All-or-nothing toggle, or screening fires for any role that hits the ATS.
• The move: Pilot on one high-volume role family before expanding.

12. Customer data isolation and model training

• Ask: Are my candidates' transcripts or scores used to train your models or any third-party model?
• Good answer: A clear no, with a written commitment in the MSA.
• Red flag: An opt-out toggle for training that defaults to opt in.
• The move: Require a no-training default and a no-third-party-sharing clause.

Quick Hits

Greenhouse + Ezra: Greenhouse's pitch positions structured AI interviewing as the first stage of the funnel, with transparency and a candidate opt-out built in. Why it matters: the bar for what an ATS includes by default just moved.

SeekOut Sam: Markets a near-real-time screen with decision logs, no auto-rejection, and alignment with EEOC and OFCCP guidance. Why it matters: pure-play screening vendors now have to compete with an ATS that ships voice AI in the box.

UKG voice agents: UKG is positioning voice agents for frontline background and skills validation. Why it matters: voice AI is moving beyond office knowledge work into hourly hiring, which is where most volume and most regulatory exposure live.

The Operator's Take

Inside any AI screening system worth buying, transcripts, rubric versions, decision logs, and drift monitors are the architecture, not features turned on for a demo. If a vendor cannot show a real transcript with a real evidence-linked score on a real role, the system was never instrumented to produce one. That is the test buyers should hold every vendor to in 2026. The risk this year is not that TA teams fail to adopt voice AI interviewing. It is that they adopt it without artifacts and absorb the regulatory exposure that follows. Run the VOICE checklist on every shortlist. Disqualify on any single Fail under Validation, Output evidence, or Candidate transparency. Pilot on one role family. Re-run the checklist annually. Write any conditional passes into the MSA before signature. TA leaders who treat artifact discipline as a quarterly habit will publish a defensible hiring record, satisfy NYC LL144 and the Colorado AI Act with documentation already in hand, and move faster than peers who treat voice AI as a procurement chore. The pattern is simple: evidence on demand earns the seat, the absence of it loses the seat.

Resource

Need a quick ROI model? Use a calculator to quantify time savings and cost tradeoffs. Get the AI Hiring ROI Calculator ($29, included with Pro subscription).

Planning rollout? Use a playbook that covers rollout sequencing and stakeholder enablement. Get the AI Adoption & Implementation Playbook ($39).

Related Reading

• The AI Hiring Buyer Has a Bigger Problem Than the Vendor
• 91% of CHROs say AI is their top concern. 47% can't measure if it's working.
• How to Measure AI Hiring ROI in 2026: 5-Input Model