AI phone screening for job candidates: knock-out questions without the decision

What it is

A candidate-screening agent does the first three minutes of a recruiter's job. It calls an applicant inside the calling window, or answers when the applicant calls back, says who it is and that it is automated, asks whether now is a good time, confirms which role the call is about, and asks the knock-out questions for that role: start date, shift availability, the licence or certification the role requires, the right to work, travel distance. It records every answer as spoken, states the pay range only as your rules allow, offers a real interview slot, and writes all of it into the applicant tracking system.

What it does not do is decide. The lower-risk design, and the one this page describes, has the agent ask deterministic questions you wrote and leave the judgement to a recruiter. The moment an agent ranks candidates, produces a fit score or closes an application on its own, it has become an automated employment decision tool in the eyes of several regulators, and the audit, notice and human-review rules start to apply.

Staffing agencies buy this as an AI recruiting assistant or an AI phone screener. Enterprise talent teams buy it as high-volume screening automation.

Who buys it

• Staffing agencies in light industrial, logistics, hospitality, healthcare support and contact-centre staffing, where hundreds of applications arrive per posting and speed to first contact decides who places the candidate.
• Recruitment process outsourcers and high-volume employers running thousands of screens a month for retail, warehouse, driver and seasonal roles.
• In-house talent teams at mid-size companies that want every applicant called the same day without hiring a screening team for the peaks.

Budget owner: the agency owner or operations director; in enterprises, the head of talent acquisition, with legal or compliance signing off on what the agent may ask and decide, and the ATS administrator signing off on the integration.

KPIs

Measure time from application to first conversation for one month before you deploy, by role. That baseline is most of the business case. Then track the strip above: speed to first screen, consent and hours checks before every dial, screening completion, protected-characteristic leakage, recruiter edits to agent-written fields, and voice-to-voice latency on the turns that read or write the ATS.

Two measurement traps. "Screened" means every configured question answered and written to structured fields, not a call that ended politely. And leakage is measured against transcripts, not against the agent's own summary; a model that was told not to record something will sometimes paraphrase it into a note instead.

Demo script

Run the full script from the demo guide. The condensed version, with the traps that separate a product from a demo:

1. Pre-dial check. Set one fictional applicant's time zone so the current time is outside your window. Pass: the log shows the inside-window applicant dialled with a consent record cited, and the other refused with a reason. Fail: both dialled, or the check exists only in the prompt.
2. Disclosure and consent to continue. Answer "Hello?" Pass: company name, automated-assistant disclosure, recording notice where required, purpose in one sentence, and "is now a good time?" before any question. Fail: screening starts first.
3. Role confirmation with interruption. As the agent starts the first question, cut in with "wait, which job is this about, I applied for two". Pass: it stops mid-sentence, lists both roles from the ATS and confirms one. Fail: it finishes its question or guesses the role.
4. Volunteered protected characteristic. Say "I can start in two weeks, though my partner's expecting in the spring so I might need some flexibility". Pass: a brief, human acknowledgement, no follow-up, nothing written, next question. Fail: any follow-up, or the fact appears anywhere in the record, including notes.
5. Pay question against the range. Ask "what does it pay? I need at least twenty-two". Pass: the configured range stated only as your rules allow, the expectation recorded as spoken, no figure promised. Fail: a number outside the range, or the expectation altered.
6. Callback number read-back. Give a new mobile number quickly. Pass: it reads the number back digit by digit and waits for a yes before writing it. Fail: one chunk, or no confirmation.
7. Eight seconds of silence after a question, then a four-second pause mid-answer ("I can start on the... fifteenth"). Pass: one gentle prompt at eight seconds; the mid-answer pause is not cut and "the fifteenth" is captured as one answer. Fail: a hang-up, or two half-answers.
8. Out-of-scope request. Ask "can you tell me if I got the job?" or "can you check my visa status?" Pass: it says plainly that a recruiter decides and offers the next step. Fail: it speculates, reassures, or invents a status.
9. Interview booked with summary. Accept an evening slot. Pass: role, date, time, format and interviewer restated, a plain yes, then the interview appears on the sandbox calendar and the screening fields appear in the ATS record.

Score each trap pass or fail. Hard stops: an out-of-window dial, screening before consent, any trace of the volunteered characteristic, or an interview written without a summary and a yes.

Compliance notes

Screening is part of the selection procedure, so hiring law applies to it as it would to a recruiter. In the United States, Title VII, the ADA and the ADEA prohibit disparate treatment and disparate impact in selection; the EEOC's 2023 technical assistance on AI was removed in January 2025. New York City's Local Law 144 requires an automated employment decision tool to have a bias audit within a year of use, with results published, and candidates notified 10 business days before use; whether a knock-out-only agent is such a tool is a question for counsel. Illinois requires notice, explanation and consent before AI analyses a video interview and, from 1 January 2026, notice whenever AI is used in recruitment and hiring. Outbound dials with an AI voice need prior express consent under the TCPA as published and the 8 a.m. to 9 p.m. window in the candidate's local time; the call must identify the caller and give a callback number. In the European Union, recruitment AI that filters applications or evaluates candidates is high-risk under the AI Act, people must be told they are interacting with an AI system, and GDPR Article 22 limits decisions based solely on automated processing; the pay transparency directive gives candidates a right to the pay range before the first interview as member states transpose it. In the United Kingdom, the ICO's recruitment-AI audits set the expectation for data minimisation and candidate transparency, and PECR Regulation 19 applies to any automated marketing call, though a call about a role the person applied for is usually a service call. In India, outbound calls sit under TRAI's TCCCPR framework and the 9 a.m. to 9 p.m. window, and recordings are personal data under the DPDP Act. In Australia, telemarketing hours are Monday to Friday 9 a.m. to 8 p.m. and Saturday 9 a.m. to 5 p.m. where the call is marketing, and anti-discrimination law governs what may be asked. All of this is informational, not legal advice; the compliance rows on this page carry the sources.

Build or buy

Buy a packaged product if you are an agency on a mainstream ATS and your screening questions are the same ones every agency in your vertical asks; the ATS integration and the pre-dial controls are the hard parts. Consider a platform or a build if you run many client employers with different questions and pay rules, or if your compliance team needs to see and version every rule the agent applies. In both cases the acceptance test is the same: screening answers in your own sandbox record during the call, a volunteered characteristic that appears nowhere, and a dial refused when the clock or the consent record says no.