AI Voice Agents for Recruiting: Screening Calls, Scheduling and the Rules

Why recruiting teams are adding voice AI

In high-volume hiring, such as retail, logistics, warehouses, contact centres, hospitality, healthcare support roles and staffing agencies, speed decides who you hire. Good candidates apply to several employers at once and take the first reasonable offer. A recruiter who calls back in four days often finds the candidate already working elsewhere.

Most of the first contact is repetitive: confirm the candidate is still interested, check they can work the shifts and location, confirm a required licence or certificate, answer questions about pay and the process, and book an interview. Recruiters spend hours on the phone doing this, chasing voicemails and rescheduling no-shows.

An AI voice agent can call every applicant within minutes of applying, at a time that suits them, in their language. It asks the same job-related questions every time, answers common questions from approved information, books interviews straight into recruiters' calendars and sends reminders. Recruiters then spend their time talking to candidates who are interested and available, and making the decisions.

Many candidates in these roles apply from their phones and prefer a quick call or text to long online forms. Done well, an AI phone screen feels faster and more responsive than a silent application portal.

Recruiting use cases

The safest and most valuable starting point is usually scheduling and logistics: getting interested candidates in front of a recruiter quickly. Screening questions come next, kept strictly objective and job-related.

How an AI screening call works

• Before calling: check the candidate gave a phone number and agreed to be contacted, and that any required notice (such as NYC's 10 business days) has been given.
• Open by saying it is an AI assistant calling on behalf of the employer about their application, and ask if now is a good time.
• Explain briefly what the call covers, how answers are used and how to ask for a person or an accommodation instead.
• Ask a short set of objective, job-related questions approved by HR and legal, the same for every candidate.
• Answer the candidate's questions from approved role information; pass anything else to the recruiter.
• Offer interview slots from the recruiter's calendar and book the one the candidate picks.
• Confirm by text or email, write answers and transcript to the ATS, and flag anything for recruiter review.

Keep the call to a few minutes. Long AI interviews feel like an exam and push candidates away, and they drift into assessment, which carries more legal weight than fact-gathering.

Keep hiring decisions with people

The most important design choice is what the agent does not do. It gathers facts and books interviews; it does not judge, rank or reject candidates on its own.

This design is also simpler to justify under the rules. A tool that only collects facts and books interviews may fall outside some definitions of automated decision-making, while a tool that scores or filters candidates usually falls within them. Your counsel decides which applies; the narrower the tool, the easier that conversation is.

The rules that apply

This is a planning summary, not legal advice, and AI employment rules are changing quickly in several US states and in the EU. Treat the rules as configuration, such as notices, required questions, retention periods and which features are switched on per country or city, so the agent can follow them as they change.

Your recruiting agency or software vendor is not the only one responsible. Under NYC Local Law 144 and the EU AI Act, employers using these tools carry duties of their own, so keep your own records of audits, notices and human reviews.

Example: objective screening flags, not rejections

This simplified example shows how screening logic can live in code that HR controls. The AI gathers structured answers; the code compares them with requirements and produces flags for a recruiter, never a rejection.

# screening_rules.py: simplified screening flags reviewed by a recruiter
from dataclasses import dataclass

@dataclass
class Requirement:
    key: str            # e.g. "forklift_licence"
    question: str       # read by the agent, approved by HR and legal
    required: bool      # a genuine, job-related requirement
    expected: object    # e.g. True, or a set of acceptable values

REQUIREMENTS = [
    Requirement("work_authorised", "Are you legally allowed to work in this country?", True, True),
    Requirement("forklift_licence", "Do you hold a current forklift licence?", True, True),
    Requirement("shift", "Which shifts can you work: early, late or night?", False, {"early", "late", "night"}),
]

def screening_summary(answers: dict) -> dict:
    """Never rejects. Produces facts and flags for the recruiter to review."""
    flags, missing = [], []
    for r in REQUIREMENTS:
        if r.key not in answers:
            missing.append(r.key)                      # ask again or let the recruiter follow up
            continue
        value = answers[r.key]
        ok = value in r.expected if isinstance(r.expected, set) else value == r.expected
        if r.required and not ok:
            flags.append({"key": r.key, "answer": value, "note": "Does not meet stated requirement"})
    return {
        "status": "needs_recruiter_review" if flags or missing else "ready_to_schedule",
        "flags": flags,
        "missing": missing,
        "answers": answers,          # stored with the transcript in the ATS
        "decided_by": "recruiter",   # the AI never sets a final outcome
    }

Requirements are kept short, objective and job-related, and HR can change them without touching the AI. Candidates who meet everything go straight to scheduling; everyone else is reviewed by a person, who may well schedule them anyway.

Fairness, accessibility and bias testing

Voice brings fairness issues that text forms don't. Speech recognition is less accurate for some accents, dialects and speech differences, background noise affects candidates calling from busy homes or workplaces, and people with hearing or speech disabilities may not be able to use a phone screen at all.

• Test speech recognition accuracy across the accents, languages and phone conditions of your real applicant pool before launch.
• Confirm key answers back to the candidate ("so that's yes to the forklift licence, is that right?") to catch recognition errors.
• Offer alternatives: text, web form or a human call, and accommodations on request, without penalty.
• Monitor outcomes by group where lawful: who is flagged, who is scheduled, who drops out, and investigate differences.
• Commission an independent bias audit where required, such as for NYC, and repeat it at least yearly.
• Review questions regularly with HR and legal to remove anything not genuinely job-related.

If error rates differ by group, fix it before launch: better models, confirmation steps, shorter answers or alternatives. Don't let recognition errors turn into screening flags.

ATS, calendar and messaging integration

Write to the ATS through narrow, logged actions. The agent should be able to add notes and propose a stage, while only recruiters move candidates to rejection. Check each ATS vendor's API terms early, as partner programmes and rate limits can shape the design.

Designing a good candidate experience

Candidates are also customers, neighbours and future applicants. A clumsy AI call can damage your employer brand quickly, especially on social media.

Measuring results

Review a sample of calls every week, especially calls where candidates were flagged or dropped out. See our guide to voice agent testing and monitoring for how to set up reviews and alerts.

Cost and timeline

Budget separately for any independent bias audit you need, and for legal review of questions and notices. Running costs are per call minute and per message; see our AI voice agent cost guide.

Off-the-shelf recruiting assistants exist and may suit standard hiring on a supported ATS. A custom build makes sense for staffing agencies, unusual workflows, several countries and languages, or when you need full control of questions, data and audit records.

Common mistakes

Frequently asked questions