An AI Phone Screen Is Audio Only. Ask Who Reads It

What Actually Changes on an AI Voice Screen?

Three things: the channel, the pace and the record. Alex, one live AI interviewer, describes itself on its own About page as an always-on first-round interviewer that runs a personalised interview by phone call, video or text at a time the candidate picks ¹. That is a marketing page, and the scale claims on it are the vendor's own and unaudited. The format it describes is real enough to plan for.

The pace changes first. The employer-facing settings documentation for the same product carries an example welcome message telling candidates that when Alex finishes speaking it is their turn, and that after two seconds of silence their response automatically ends ². Employers can edit that text, so the exact cutoff will not be the same everywhere, and the page does not document the underlying voice detection separately. The habit it rewards is easy to describe: finish the thought, then stop.

A pause to gather your thoughts can read as the end of your turn.

There is nothing visual to fall back on either. No nod, no screen to share, no expression telling you a point landed.

The third change is the record. Alex's candidate FAQ says the recording, the transcript and the notes are shared with the employer's hiring team, and that humans make all hiring decisions ³. It also says the product does not automatically reject candidates. That is a vendor describing its own product rather than an audit, it says nothing about retention or deletion, and another caller's product may work differently. What travels onward is a file, and the file is made of your words.

Where you take the call goes into that record too. The same settings page describes an optional Cheat Detection add-on that watches for signals such as background voices, excessive tab switching and unusual delays before answering, with the employer setting a flag threshold from 0 to 100 ². It is an add-on, so plenty of interviews run without it, and the page publishes no accuracy or false-positive figures and no candidate notice. A housemate on a call in the next room is a poor backdrop regardless.

Ask Who Is Calling and Whether a Person Reviews It

Ask on the call, in the first minute: who the company is, whether the conversation is recorded, and whether a person reviews it. Most candidates say they were not told. In Greenhouse's 2026 candidate report, a survey of 1,200 US respondents, 70% said they were never clearly told upfront that AI would be evaluating them and 21% found out only once the interview started ⁴.

Treat that as what it is.

It is self-report from a company that sells hiring software, and it measures what candidates noticed rather than what a consent screen contained. It still says that not knowing is the common experience rather than a personal lapse.

In the EU, whether you are talking to an AI at all now has a statutory answer. Article 50(1) of the AI Act requires providers to design systems that interact directly with people so the person is informed they are talking to an AI, unless that is obvious in the circumstances. That duty applies from 2 August 2026 ⁵. The duty falls on the company that built the system rather than the employer using it.

It stops at disclosure. Nothing in Article 50(1) says what happens to your answers, and nothing in it gives you a right to a person instead. The UK has no equivalent rule, and elsewhere what you are owed depends on local law, so treat the question as a courtesy the caller can decline.

A recording of your words is not automatically a voiceprint. Illinois' Biometric Information Privacy Act lists a voiceprint as a biometric identifier and expressly excludes photographs and writing samples ⁷. Whether a particular product extracts one is a question of fact about that product, and courts have split on where the line falls. BIPA reaches private entities operating in Illinois and nowhere else.

You lose nothing by asking, and the answer itself tells you something.

Four questions worth putting to the caller, or into a reply to the invitation:

• Which company is this, and which role is the call for
• Is the call recorded, and how long is the recording kept
• Who reads the transcript, and does a person make the decision
• Is another format available if this one does not work for me

Should You Take the Call Right Then?

Only if the room is right. Alex's About page describes the interview happening at a time the candidate chooses ¹, and its candidate FAQ says someone who misses a slot because of a technical issue can take the interview later on the same link ³. Those are one vendor's published terms, not a rule that covers every caller. Asking for a time still costs less than a screen taken beside a road.

Say it plainly: "I want to give this a proper answer. Could I do it in an hour, from somewhere quiet?" A recruiter's voicemail would get the same sentence.

Some people decline the format outright. In the same Greenhouse survey, 38% said they had walked away from a hiring process because it included an AI interview, and another 12% said they would ⁴. That is a reported past action with no verification, and the release does not say whether the AI round was the only reason it happened. It is common enough that you are not unusual either way.

Walking away is a real option with a real cost, and only you can price it.

If you do take the call where you are, the audio decides how you come across. A supermarket car park hands the transcript your words plus a tannoy and an engine. Headphones with the microphone near your mouth beat a phone on speaker on a dashboard. None of that tests your answers, though background voices are one signal the optional Cheat Detection add-on watches for ². All of it shapes what ends up written down and passed on.

Treat It as the Recruiter Screen It Replaces

Prepare the way you would for a twenty-minute call with a recruiter, because that is the slot this fills. In a natural field experiment at one Philippines customer-service employer, 70,000 applicants were randomly assigned to a human recruiter or an AI voice agent, and human recruiters read the transcripts and made every hiring decision in both arms ⁶.

Applicants interviewed by the AI were 12% more likely to receive an offer, and transcript analysis found the agent following the firm's own question guideline more closely than the human recruiters did ⁶. That is one firm, one country and one job family: high-volume customer-service hiring on short screening calls with very high turnover. It is a 2026 working paper rather than a peer-reviewed result, and it reports no bias or adverse-impact analysis. Read it as evidence about consistent question-asking, not about a machine deciding.

The costs are in the same paper. 5% of applicants ended the interview rather than speak to an AI, and the agent hit technical difficulties in 7% of cases ⁶.

Guidance written for interviewers is public, and reading it is cheap preparation. What follow-up questions expose whether someone understands an answer is written for interviewers, and it describes a move to expect from any interviewer, human or not: take one claim, then press on the specifics behind it. What to ask an employer about the AI in their hiring covers how to put process questions to a recruiter as plain logistics.

Four kinds of AI interview sorts the formats, and a live voice call is only one of them.

Have one example ready that you can tell in order, out loud, in ninety seconds.

A transcript carries a specific answer well and a vague one badly.

See how it works >