An AI Interview Without a Human Interviewer, Explained

A question that keeps coming up on demo calls is about the AI interview without a human interviewer: if nobody from my team is on the call, who is actually interviewing the candidate?

The short answer is that a voice agent runs the call, but the job description wrote the questions and the scoring before any candidate was contacted. Since May 12, 2026, when per-job screening went live on our production stack, that has been the default path at HireQwik. A candidate books a slot, joins from a browser, talks for fifteen to twenty minutes, and a scored record with a verdict is waiting in the recruiter’s queue. No recruiter dialled in. No panel. Nobody took notes.

That sounds like screening without judgment. It is not. The judgment is all still there. It just moved, and if you do not know where it moved to, you will trust the output for the wrong reasons or distrust it for the wrong reasons. This post is about where it went.

How Long Is the AI Screening Call? The 20 Minutes Explained

The call itself runs fifteen to twenty minutes. That is the length of the structured conversation, not the time between uploading a job description and holding a scorecard. Before the call there is a resume score, an invite, and a candidate who picks a slot on their own schedule. The setup happens once per role.

Around that call, a candidate may pick a slot that evening or on Sunday night. We have written the full elapsed-time story in the application to result timeline, and the honest summary is that the human-free part is fast while the human-dependent part, a candidate deciding to show up, is exactly as fast as that candidate feels like being.

So when I say twenty minutes, I mean the interview. The setup is once per role, and everything after it repeats without you.

How Does the Job Description Run the AI Screen?

Here is the part most buyers do not expect. The job description is not a reference document that the model glances at. It is compiled into a structured screener document before a single candidate is contacted, and that document is what runs the call.

It carries the competencies for that specific role, the questions that test each one, the disqualifiers, and how an answer maps to a score. A field sales role and a platform engineering role do not share a rubric, because they did not share a brief.

We shipped this as per-job screening rather than a single template applied everywhere, and the difference shows up immediately in the score spread: one generic rubric bunches everybody into the same stretch of the range, and no cut-off drawn through that bunch separates a fit from a non-fit.

This is why a thin brief produces a thin screen. If the job description says good communication and relevant experience, the screener built from it will ask soft questions and score them softly, and you will end up with a queue of candidates who all look like a six out of ten. The lever is upstream. We covered what that looks like in practice in what a vague job description does to your rubric.

The practical read: the recruiter who writes the brief is the interviewer. They are just doing the interview once, in writing, instead of again and again on the phone.

What Do Phase-0 Knockout Questions Do First?

The first questions on the call are not warm-up. They are the disqualifiers, and they run first on purpose. If a night-shift role meets a candidate who cannot work nights, there is no value in eighteen more minutes of assessment. The call ends politely, and the record is tagged with the exact requirement that was missed.

We first verified this behaviour on production on May 10, 2026, in a flow where a candidate passed the first disqualifier and failed the second, and the call closed itself correctly.

Two things follow from that design that are worth saying out loud.

The first is that a large share of rejections never consume a full slot. That is where the throughput comes from, more than from the model being clever. The second is that a badly written disqualifier is the most expensive mistake available to you, because it rejects people fast and at scale. A knockout that asks about a certification the role does not genuinely require will quietly delete a good part of your funnel by Tuesday. Knockout questions deserve their own design pass for exactly that reason.

Why Score Every Call With Two Evaluators?

When a recruiter sits on a screening call, they are doing two jobs they never separate. They are judging what the candidate said, and they are judging how the candidate said it. Hesitation, pace, whether an answer arrived or was assembled.

Strip the human out and a transcript only preserves the first half. So since April 2026 there have been two evaluators on every completed call. A language model scores the content of the answers against the role’s rubric. An audio layer scores delivery from the recording itself: pace measured in words per minute, how often the candidate hesitates or reaches for a filler, pronunciation, and a CEFR fluency read.

The rule that binds them matters more than either score. The blend is deliberately asymmetric. Speech evidence can lift a borderline candidate from a reject into a hold, and a rejection is only ever recommended when the two of them agree. It never works the other way around.

A strong candidate with a heavy regional accent does not get marked down by the audio layer, because the layer is not allowed to demote. That constraint is the only reason I am comfortable running this unattended at scale, and we explain the mechanics in how the two-evaluator check works.

What Is in the Scorecard When the Call Ends?

The scorecard is written when the candidate leaves the call, not after a batch job overnight. It carries a verdict tier, Strong Go, Go, On Hold or No Go, the per-question evaluation with quotes from the candidate’s own answers, the speech read, and the recording with its transcript.

The quotes are the load-bearing part. A number with nothing under it is an opinion with a decimal point. A number with the sentence the candidate actually said under it is something a hiring manager can argue with, and that is what you want.

That record is also shareable without an account. A hiring manager who will never log into a screening tool can be sent a read-only page carrying the verdict, the scored dimensions, the notes and the transcript, and that link can be revoked later, which kills every copy of it at once. We wrote up that mechanism separately in sending a scorecard to a hiring manager without a login. For an agency desk, the same page replaces call notes, which is the subject of AI screening for staffing agencies.

Who Owns the Verdict? Auto-Decide Bands Explained

Auto-decide bands are per role and opt-in. A recruiter sets a floor below which candidates are auto-rejected and a ceiling above which they are forwarded, and only the uncertain middle band reaches the review queue at the HR dashboard. Nothing switches the bands on by itself. If a role has none configured, every candidate stays in the queue for a human decision.

That is the difference between unattended and unsupervised. I would go further: leave the auto-reject side off on a new role until you have read thirty scorecards yourself and agree with them. The bands are a way to spend your attention on the uncertain middle once you trust the edges. They are not a way to skip the part where you learn whether the rubric is any good.

The accountability question has a regulatory edge now, not just an operational one. Hiring sits in the high-risk category under the EU AI Act, and some regulators expect a documented human review point rather than a claim that one exists.

SHRM’s 2026 State of AI in HR research found that 49% of organisations have an AI-use policy and only about a quarter of those believe it is future-proof, which tells you how thin the governance layer still is at most companies. An unattended screen with evidence attached to every verdict is defensible. An unattended screen with a score and no transcript is not.

How Does the Outcome Loop Catch a Wrong Screen?

The part I care most about is the one that runs after the hire. A screening record can carry what actually happened to the candidate: hired, offered, not selected, or withdrew. That turns into a per-role view comparing the verdict the screen gave against the outcome the team reached.

The cell worth staring at is the candidate the screen said No Go to and the team hired anyway. That is the miss that matters, and it is surfaced on its own rather than averaged into a satisfying accuracy number. Below ten records the card shows raw counts instead of rates, deliberately, because a percentage over six candidates is theatre. None of it auto-tunes the thresholds either. Changing a band stays a human decision, on purpose. We covered that discipline in recording what actually happened to your screened candidates.

What Goes Wrong When Nobody Is on the Call?

Unattended means nobody is there to rescue a call that goes wrong, so it is worth knowing what wrong looks like.

A candidate whose connection drops eight minutes in has not given you a twenty-minute interview, and you should not judge them on the twelve minutes that never happened. Check how your queue marks a call that ended early, and re-invite that candidate rather than rejecting them.

A candidate who answers three questions in monosyllables is a harder case: both evaluators have very little to work with, so a confident verdict on a short transcript deserves suspicion, and the right move is to read the transcript before you trust it. If you see a queue full of On Hold verdicts with short transcripts, look at the invitation before you look at the candidates. It may not have told people what the call would be, so they joined expecting a two-minute automated form.

The one that costs the most is a disqualifier written for a requirement the business has since relaxed. Nothing fails loudly. The call simply ends early for a steady stream of people who would have been fine, and the queue looks thin in a way that reads like poor sourcing. Re-read your own knockouts monthly against what the hiring manager is currently willing to accept, because that is the one input that drifts without anyone updating it.

What the Unattended Screen Does Not Do

It does not sell the role. A voice agent asking a candidate about their notice period is not going to talk a reluctant senior engineer into your process. It can answer a few company questions from documents you have uploaded, and it passes the rest to a person, which is a different thing from persuading someone. What the agent will say when a candidate asks is the subject of candidate questions during an AI interview.

It does not replace a hiring conversation. It replaces one call, the repetitive twenty-minute screen that a recruiter runs again and again and mostly forgets. Some roles should skip it entirely, and roles that need a human first round sets out which.

And it is not free of your errors. The rubric, the disqualifiers and the bands are yours. The screen executes your judgment faithfully at volume, which is excellent when the judgment is good and unforgiving when it is not.

Where the Interviewer’s Judgment Goes

Unattended screening does not remove the interviewer. It front-loads the interviewer into the brief and back-loads the evidence into the scorecard, and it leaves a twenty-minute conversation in the middle that runs whether you are awake or not.

So what you should put to a vendor, us included, is not whether the AI is accurate. It is this: show me where my judgment enters, and show me the evidence under one verdict. If a vendor cannot point at both, what you have is not an unattended screen. It is a guess with a user interface.

With no recruiter on staff, the setup is laid out in a first round interview without a recruiter. The fastest way to see what that output looks like is to try the screening interview as a candidate. It runs about five minutes, needs no signup, and the scorecard that lands in your inbox is built by the same engine a real candidate’s is.