AI phone screen questions: what it actually grades

An AI phone screen grades the words in your transcript and how you structure them, far more than how your voice sounds. The scoring runs on what you say, and a human usually reviews the candidates who pass, so you are answering for two readers at once. The industry line that AI cuts time-to-hire by 90 percent while holding the same accuracy is half true, because the speed is plausible but the best published accuracy is weak and comes from the vendors' own researchers. So the answer that survives is a tight account, spoken from memory, of one situation, what you did, and the number it moved, delivered in 60 to 90 seconds and defended when the follow-up lands.

By UnchartedCareer

Last updated: July 2026

What does an AI phone screen actually grade?#

Your words and your structure, not your voice. The generic prep lists get this backwards. They coach your tone and pace as if the machine listens to your delivery, when it mostly reads a transcript of what you said.

HireVue is the clearest case. It removed visual analysis from its scoring, announced in January 2021, once its own research found that non-language inputs barely moved the model. HireVue's figures put the nonverbal contribution at about 0.25 percent of predictive power in most roles, and about 4 percent even for heavy customer-contact jobs (HireVue, reported by SHRM and Fortune, January 2021). What was left is the transcript. The model scores the substance of your answer, then attaches the score to the quote it came from.

The newer voice-first tools describe the same mechanism. HeyMilo sets weighted competencies for the role and returns a per-question score with the reasoning attached. Ribbon runs a voice screen the candidate completes on their own time, then hands the recruiter a transcript and an AI summary. Apriora's "Alex" holds a real-time, two-way call and writes a structured report after. Different products, one pattern: transcribe the call, match your words against a rubric, score the content.

Read that plainly and it changes what you train. A warm voice wrapped around a vague answer scores low. A flat voice around a specific, well-ordered answer scores well. HireVue's recorded video rounds run on this same transcript scoring, and clearing a HireVue round covers that format directly.

Does a human ever hear it?#

Often, yes. Automated scoring is one mode of review, not the universal one, and a person commonly reviews the people who pass. The research on asynchronous interviews treats machine scoring as one option among several, not the default on every call (International Journal of Selection and Assessment, 2025).

The format is common enough to prepare for deliberately. Recruiting is the HR function with the highest AI adoption, at 27 percent of US organizations in an SHRM survey fielded December 2025 (SHRM, 2026). But 27 percent is a long way from every screen being automated, and you rarely learn on the day which kind you got. So the answer has to clear both bars at once: clean enough for a model to parse, substantive enough for a recruiter to trust when they skim the transcript.

Is the "90 percent faster, same accuracy" claim true?#

The speed half is believable. The accuracy half is marketing. The line circulates across vendor pages and recruiting blogs, sometimes pinned to HireVue's research, that AI cuts time-to-hire by 90 percent while keeping accuracy comparable to traditional hiring. Two things break it.

First, "same accuracy" as what, measured how. The strongest published number belongs to HireVue's own team: automated competency scoring correlated with later job performance at an uncorrected r-bar = .24 across five US samples totaling 1,124 people (Liff et al., Journal of Applied Psychology, January 2024), ranging from .20 for maintenance roles to .27 for call-center roles. The same authors place that below the r-bar = .32 that Sackett and colleagues (2022) report for human structured interviews. On the personality side, the independent evidence on whether these assessments predict performance is thinner still. Stevenor and colleagues found AI-scored interview personality had weak, statistically nonsignificant links to supervisor ratings of performance, in a Study 2 group of just 25 people (Stevenor et al., 2024). A weak correlation from the vendor's own lab is not "the same accuracy."

Second, the 90 percent figure does not hold still. Across the same body of vendor and blog copy the time-to-hire reduction shows up as anything from a third to 90 percent, depending on who is selling. A number that swings that far is a marketing range, not a finding. Faster is real. Same accuracy is unproven, and the accuracy the tools do have is modest. That is exactly why trying to perform for the scorer is the wrong goal. Answer for the human who reviews the score.

What questions will an AI phone screen actually ask?#

Predictable ones, by design. These tools exist to gather comparable data from every applicant, not to surprise you, so the opening set repeats across roles. Expect four buckets.

The self-pitch. Some version of "tell me about yourself" or "walk me through your background." They want 60 seconds tying your experience to this specific role, not your life story.

Logistics and eligibility. Notice period, work authorization, salary expectation, earliest start. Answer these flatly and move on. They are gates, not scored competencies.

A behavioral prompt or two. For customer-facing and high-volume roles, a conflict or pressure question is close to guaranteed, something like "tell me about a time a customer was angry and what you did."

Role-specific probes. A real problem pulled from the job, like "how would you handle model drift in production" or "walk me through triaging a P1 outage."

The twist the generic lists miss is the second question. The two-way voice tools, Apriora's Alex and Ribbon and HeyMilo among them, can generate an unscripted follow-up off whatever you just said. The opener is predictable. The probe that comes after it is not.

What is the answer shape that survives audio-only?#

On a phone screen there is no camera to read you, so the whole signal is your words and your audio. Structure and specifics carry everything. The shape that survives is one line of situation, the bulk of your airtime on what you did, and one line of result with a real number, inside 60 to 90 seconds. The common advice to cap answers at 60 to 90 seconds is right for a mechanical reason: the model has to compress you, and past 90 seconds it starts clipping your point.

You can drill this tonight, free, and you should.

Pick three questions: your self-pitch, one conflict story, one role-specific problem. Open your phone's voice recorder. No notes, no camera. Answer each out loud, all the way through, the way you would to a stranger who can hire you.

Then play it back with your eyes closed. If you cannot hear where the situation ends and your actions begin, a transcript cannot show it either, so add a plain signpost ("so what I did was"). If your result landed as "it went well" instead of "cut average handle time by 18 percent," say the number out loud. Cut anything that drags you past 90 seconds.

Last, throw yourself the follow-up you would dread, the "why that way and not the cheaper option," and answer it cold with no restart. That is the manual version of what the tool does to you: transcribe, check the structure, probe the gap. Hearing your own filler on playback is what fixes it, and one honest rep beats a week of silent rehearsal. The case for mock interviews lays out why.

Where the solo drill hits its wall#

You can rehearse every question you predict. You cannot generate the follow-up you did not see coming. Alone, you already know why you chose the expensive option, so the follow-up you throw yourself is soft by definition. A two-way voice AI, or a live recruiter, hears the exact gap in your answer and pushes precisely there. That unscripted push is where rehearsed answers come apart, and it is the one thing a voice recorder cannot reproduce.

When you want that follow-up thrown at you on demand, scored, and repeatable, AI interview practice runs a live two-way voice mock and asks the question you did not plan for, after you have done the solo audio drill yourself.