Job applicant screening
Early candidate calls are repetitive but consequential. A recruiter needs the same job-related facts from every applicant: which role they mean, when they can work, whether they meet stated requirements, and how to reach them. When those calls happen between interviews or after office hours, answers can end up incomplete, inconsistent, or trapped in voicemail.
An AI phone agent can run a recruiter-authored question set and record what the candidate says. Its role should be deliberately narrow. It does not decide who is qualified, score answers, rank applicants, infer traits, or reject anyone. Recruiters and hiring managers remain responsible for the questions, their use, reasonable exceptions, and every employment decision. This page focuses on the screening-call function; the industry guides cover the broader staffing and recruiting context.
A concrete applicant screening call
Consider a candidate calling about an operations coordinator opening.
Agent: Thank you for calling Northstar Operations. I am the company's AI recruiting assistant. I can collect your answers for a recruiter to review. I do not make hiring decisions. Which opening are you calling about?
Caller: The operations coordinator role.
Agent: Thank you. May I get your full name, phone number, and email address? I will confirm each detail back to you.
Caller: My name is Sam Ortiz, and you can use this number.
Agent: I have a few questions approved for this role. Are you available for the schedule described in the posting? What date could you start? Which of the listed scheduling tools have you used in prior work?
The agent asks the same configured questions in the same sequence, permits the candidate to clarify an answer, and records the response without commentary. If Sam asks whether an answer is “good enough,” the agent explains that a recruiter reviews the information. It can then offer an interview only when the recruiting team's rules explicitly allow that branch. Otherwise, it confirms that the completed record will be reviewed and explains the approved follow-up process.
The boundary matters. Consistent collection can reduce clerical variation, but consistent questions do not make automated hiring judgment appropriate. Treat the phone call as a structured phone intake form with conversation and a human reviewer.
How to configure the workflow on ThunderPhone
Start with a question set approved for the specific role. Avoid one universal script that asks for information the recruiter does not need. The agent prompt should define:
- an opening disclosure, the employer or agency name, and the agent's limited role;
- how to identify the correct opening before asking role-specific questions;
- required contact fields and confirmation rules for names, email addresses, phone numbers, dates, licenses, or certifications;
- the exact approved questions, including allowed follow-ups and neutral handling for “not sure” or “I need an accommodation”;
- a prohibition on evaluation, ranking, prediction, rejection, or improvising new screening criteria;
- when to offer scheduling, record a callback request, or transfer to recruiting staff; and
- what to say when a candidate asks about compensation, status, policy, or an exception outside the approved material.
Attach only approved job descriptions, application instructions, location information, and recruiting policies to the agent's knowledge base. ThunderPhone supports text, Markdown, CSV, PDF, and DOCX files, but each document must be selected for the agent. The built-in knowledge search can retrieve approved information during a call. It should not be used to infer a decision from a résumé or compare one applicant with another.
Use an inbound demo U.S. number while testing. Demo numbers are inbound-only and not intended for production. For production, bring a number through a supported direct VoIP connection or manual SIP configuration. Verified imported numbers support inbound and outbound calling. If a candidate needs a person, cold transfer is available; attended warm transfer requires an eligible production number that can place an outbound SIP leg.
Choose integrations based on where the recruiting team works. Google Sheets can support a structured review queue. Slack can notify the recruiting team that a record needs attention. Cal.com or Google Calendar can support an approved interview-booking branch. REST or webhooks can deliver the completed call record to an applicant system without implying a native connection. A remote Zapier MCP server or another streamable-HTTP MCP server can be added by URL, synchronized for tool discovery, and attached when the business controls the available operations.
Test the boundary before launch. Simulate candidates who ask for a hiring verdict, volunteer unrelated personal information, need a question repeated, decline to answer, refer to a different job, or request a person. ThunderPhone supports browser mic tests, AI-caller simulations, reusable scenarios, graded call logs, and regression suites with pass-rate gates. Simulations are billable real calls, and the interface shows the charge before the run.
What to measure
Measure quality and consistency without turning the call into a hidden ranking system. Review:
- completion and correction rates for required fields;
- records linked to the wrong opening;
- candidate requests for clarification or human help;
- recruiter rework caused by unclear or missing answers;
- scheduling, transfer, and integration failures; and
- prompt-boundary failures, especially evaluative or unapproved language.
Use graded calls to check the disclosure, exact question set, neutral follow-ups, confirmation, and closing summary. Keep outcome analysis with human reviewers. A human-in-the-loop design is not an exception here; it is the intended operating model.
Where a human remains necessary
People must decide which questions are appropriate, interpret answers, consider context, handle accommodations and exceptions, and make every advancement or hiring decision. A person should also handle disputes, sensitive disclosures, unclear identity, or any request the approved script cannot address. The agent can produce a consistent record, but it cannot make the record fair, lawful, or sufficient by itself.
When a candidate advances, use a separate interview scheduling flow with its own permissions and fallback rules.