How to use AI for phone screens: before, during and after the call

Use AI on a phone screen in three places: before the call, to turn the job description into knockout and must-have questions; during the call, to transcribe and keep track of which questions you have covered; and after the call, to draft the ATS note and submittal from the transcript. The hard part is the phone itself. An ordinary cellphone call has no meeting for a bot to join, so the audio has to run through your computer or through a phone system that transcribes. And because a phone call shows no recording badge, you say a consent line before the first question.

This page is about AI that helps a recruiter run their own screen. AI voice agents that conduct the screen instead of you are a different decision, covered near the end. For the screen itself, the call form is the recruiter phone screen template and the questions are in phone screen questions for recruiters.

Where does AI fit in a phone screen?

The biggest saving is usually the write-up. Recruiters who capture the call well spend minutes, not half an hour, turning it into a note, which is the point of the same-day screen to submittal workflow.

Before the call: build the screen from the job description

Give the model the job description and your intake notes, and ask for a short, ordered screen rather than a long question bank. A prompt you can adapt:

From the job description and intake notes below, draft a 20-minute
phone screen for [role].

1. KNOCKOUTS (max 4): only legal, job-related requirements stated in the
   intake (location or travel, work authorization question in a lawful
   form, required license, schedule). One question each.
2. MUST-HAVES (3 to 5, in the client's words): one open question each
   asking for a specific recent example, plus one follow-up that asks
   what the candidate personally did and what the result was.
3. RESUME GAPS: from the resume below, up to three questions about
   unexplained dates, short tenures or claims to clarify.
4. LOGISTICS: pay expectation, notice, other processes, consent to submit.

Do not include questions about age, family, health, religion, national
origin, current or past salary, or anything not in the intake.

Job description: [paste]
Intake notes: [paste]
Resume: [paste, with contact details removed]

Check the knockouts carefully. Salary history questions are banned in many states and cities, and work-authorization questions have lawful and unlawful forms; see work authorization questions in interviews. Never let a model add a knockout the client did not ask for.

How do you get a phone call into an AI tool?

This is where most phone-screen AI setups fail. Pick the route before the first call and test it with a colleague.

Interview Signal, for example, works on phone screens only when the call runs through the computer's audio: it labels your microphone as the interviewer and the call's sound as the candidate, with no bot involved. The setup options for every route, including what each one announces to the candidate, are in how to record a phone screen.

Consent comes first on a phone screen

Treat AI transcription as recording. Wiretap and eavesdropping laws are mostly written around capturing a conversation with a device, and several states require every party's consent. On a video call the candidate may at least see a badge; on the phone they see nothing. Say it before the first real question:

"Before we start, I use a transcription app on my calls so I can listen
properly instead of typing. It's only for my notes on the [role] position.
Is that OK with you, or would you prefer I take notes by hand?"

If they say no, switch the tool off, say that you have, and take notes by hand. More versions, including a strict one for all-party consent states, are in the interview recording consent script, and the statutes are summarized in is it legal to use AI note takers in interviews. This is practical guidance, not legal advice.

During the call: what live AI should and should not do

Live assistance helps when it reduces what you have to hold in your head. It hurts when it pulls your eyes to the screen.

• Useful: a list of must-have questions that ticks off as each is covered, so you do not finish the call and realize you never asked about the license.
• Useful: a suggested follow-up with a reason, such as "vague on their own role: ask what they personally did." You decide whether to ask it.
• Useful: a running transcript you can glance at to check a number the candidate just gave.
• Harmful: reading suggested questions out word for word. Candidates notice, and the call stops being a conversation.
• Harmful: scoring on screen while you are still listening. You start listening for the score instead of the answer.
• Harmful: anything that claims to read emotion, honesty or personality from the voice. In the EU, AI that infers emotions in the workplace is prohibited under Article 5(1)(f) of the AI Act, and the European Commission's guidelines on prohibited practices apply that to candidates in recruitment.

After the call: from transcript to ATS note

Decide first, write second. Before you open any draft, say out loud or write down: submit, hold or reject, and why in one line. Then ask the model for the note in your fixed format:

Using only this transcript, fill in the screen note below. Quote the
candidate word for word where you use quotation marks. Write "not
covered" for anything that was not discussed. Do not make a
recommendation; I will add the decision.

Must-haves: MH1 [ ] / MH2 [ ] / MH3 [ ]  (shown / partly / not shown + quote)
Gap and what they said about it:
Motivation: "[quote]"
Pay expectation as stated:
Notice / earliest start:
Other processes and stages:
Consent to submit to [client] for [role]: [yes/no, time]

Transcript: [paste]

Check the quotes, the pay figure and the consent line against the transcript, add your decision, and paste the note into the ATS. If you are submitting, the same transcript drives the client document; see how to turn interview notes into a submittal.

Should an AI voice agent run the phone screen instead?

Voice agents that call candidates, ask questions and report back are now sold for high-volume hiring. Before you replace a recruiter's screen with one, work through these questions:

1. Does the candidate know it is AI? In the EU, AI systems that interact directly with people must be designed so people are told, unless it is obvious (Article 50(1) of the AI Act). Elsewhere, telling candidates is still the honest default.
2. Does it score, rank or reject? If its output decides who moves forward, AI hiring rules may apply, such as New York City's bias audit and notice rules. Both are covered, with the statutes, in the legal guide linked above.
3. Is there a human route? A candidate with a speech or hearing disability, or who simply asks for a person, needs one, and it should not cost them their place in the process.
4. What does it do to your relationship with the candidate? For agency recruiters, the screen is where trust is built and where you learn what the candidate will not tell a client. That rarely survives automation.

For hourly, high-volume roles with simple knockouts, an agent may be reasonable with disclosure and a human fallback. For most professional searches, AI that supports your own call gets more of the benefit with fewer of the risks.

A phone-screen AI checklist

1. The capture route is tested with a colleague, and both voices come through clearly on separate channels.
2. The question set was drafted from the job description and edited against the intake.
3. No knockout or question goes beyond what the client asked for and the law allows.
4. The consent line is on screen or on a card, and you say it before the first question.
5. You can stop capture in one action if the candidate says no.
6. Live prompts help you track coverage; you are not reading them aloud.
7. The decision is yours and written before you read the AI draft.
8. Every quote, number and the consent to submit are checked against the transcript.
9. The transcript and note are stored where your retention policy covers them.