How AI Video Interviews Work: The Two Formats Explained

AI video interviews work in one of two ways, and the label covers both. In the one-way format, a question appears on screen, the candidate records an answer alone with no interviewer present, and software transcribes and scores the recording afterwards. In the two-way format, a conversational agent asks the questions live, hears each answer and asks its own follow-ups before scoring. Both produce a transcript, per-competency scores and a ranked shortlist. Only the second can probe a thin answer, because the first has no way to react to what was said.

Almost every buyer guide treats "AI video interview" as one product. It is two, and they fail in different places. Teams that buy the wrong one usually discover it about three weeks in, when they realize the tool cannot do the thing they assumed was table stakes.

This is a plain walkthrough of what actually happens from the moment a candidate clicks the invitation to the moment a recruiter opens the shortlist, plus the parts vendors tend to skip: what gets scored, what does not, how long it takes, and what you have to disclose.

How do AI video interviews work?

There are five stages, and they are the same for both formats until stage three.

First, the question set is built from the job requirements and tied to competencies, before anyone applies. Second, candidates get an invitation with a link, and they complete the interview on their own schedule rather than on yours, which is the entire reason this format exists. Third, the interview runs, and this is where the two formats diverge. Fourth, the audio is transcribed in full and each answer is matched to the competency the question was written to test. Fifth, each answer is rated against anchored levels, and the ratings roll up into a ranked list a recruiter reads with the transcript attached.

Stage three is worth slowing down on. A one-way interview shows a question, gives the candidate a short window to think and a longer one to record, then advances automatically whether the answer was complete or not. A two-way interview is a conversation: the agent asks, listens, and if an answer is vague or a candidate claims something checkable, it asks a follow-up before moving on. The mechanics of the scoring step are identical in both cases and are covered in detail in our guide to how AI interview scoring works.

One-way and two-way AI video interviews compared

Both are legitimate tools. If you are screening 800 applicants for a role where the only real question is whether someone has a forklift certification and can work nights, the one-way format is efficient and honest about what it is doing. Our page on one-way video interview software covers that case. If the job turns on whether a claimed achievement holds up under one more question, a format that cannot ask that question is not screening, it is collecting.

What is an automated interview?

An automated interview is a first-round interview conducted by software rather than by a person, on the candidate's own schedule, with the answers scored against criteria written in advance. The automation covers the invitation, the interview itself, the transcription and the scoring. It does not cover the decision.

That last boundary is not a modesty claim, it is how these systems stay defensible. When an automated tool substantially assists or replaces a discretionary hiring decision it becomes a regulated instrument in several US jurisdictions, with audit and notice duties attached. Keeping a person on every advance and reject is both better hiring and a smaller legal surface. The definition that triggers those duties is unpacked in what is an AEDT.

What does the recruiter actually get back?

Four artifacts, and their usefulness depends almost entirely on whether they are linked to each other. You get the recording, a full transcript with the questions marked, a score per competency rather than one overall verdict, and a ranked list of candidates.

The link between the score and the transcript is the part to insist on. A rating that points at the exact sentence which produced it can be reviewed by a hiring manager, corrected when it is wrong, and defended when a candidate asks why they were rejected. A rating that floats free of the evidence cannot do any of those things. When you are evaluating vendors, ask to click a score and see what happens. How that evidence chain is built is on our candidate interview scoring page.

A per-competency breakdown matters for the same practical reason. One aggregate number hides which criterion drove the result, which means it also hides where a systematic problem would be if you had one. If every candidate over 45 is losing points, you will only ever see that in a breakdown.

Does the AI score how a candidate looks and sounds?

In a well-built product, no. The score comes from what the candidate said, which means the video is transcribed and the words are rated. The picture is there so a human can watch it later, not so a model can read a face.

This deserves a direct answer because the category earned its reputation honestly. Older video assessment products did score facial expression, vocal tone and speech cadence as proxies for traits like enthusiasm and confidence. Those signals track accent, culture, disability and webcam quality far more reliably than they track job performance. HireVue removed facial analysis from its assessments in 2021 following an independent audit, and VidCruiter states plainly that it uses no biometric or facial analysis. The practice is largely retired, and repeating the criticism as though it were current does not describe the market as it is in 2026.

What has not gone away is the need to ask. The question is narrow and any vendor can answer it: does any part of the score derive from face, tone or speech rate? Get the answer in writing. If a product offers you an enthusiasm or confidence metric, ask what it was validated against and who audited it, because those two questions usually end the conversation.

How long does an AI video interview take?

For the candidate, most first-round screens run 10 to 20 minutes, covering roughly five to eight questions. Shorter than that and you are not gathering enough to rank anyone; much longer and completion suffers, particularly for hourly roles where people are fitting this around a shift.

For the employer the arithmetic is the actual selling point. A 15 minute phone screen costs closer to 25 minutes of recruiter time once scheduling, no-shows and write-up are counted, so 100 applicants is about a working week for one person. Automating that round turns the week into a couple of hours of reading transcripts. The change is in where recruiter hours go, not in how many recruiters you need, and the honest version of the pitch is that candidates who would have been screened out for lack of time now get interviewed at all. That case is made in full on automated interview software.

We do not publish completion-rate benchmarks, and you should be wary of anyone who does. Every published figure in this category defines its denominator differently, and none of them survive contact with your own funnel.

Are AI video interviews legal, and what do you have to tell candidates?

Yes, with disclosure, and two separate bodies of law apply at once. Buyers routinely conflate them and then comply with only one.

The first is ordinary recording law, which cares that a conversation is being captured at all. Around a dozen US states require all-party consent, and because the obligation follows the candidate rather than your headquarters, a remote pipeline is far better off disclosing and capturing agreement from everyone than running a state-by-state matrix. We covered that in recording job interviews and consent.

The second is the AI hiring rules, which attach to the decision rather than the recording. New York City's Local Law 144 requires an independent bias audit from within the previous 12 months, a published summary, and candidate notice at least 10 business days ahead, and that duty follows the job location. Illinois requires notice, explanation and consent under AIVIA, and its amended Human Rights Act took effect on 1 January 2026. Maryland requires a signed waiver before any facial template is created. Colorado is the one most articles still get wrong: the June 2026 date was superseded, and SB 26-189, signed in May 2026, takes effect on 1 January 2027. The state-by-state position is in are AI job interviews legal, and who owes what is on AI hiring compliance.

Retention is the duty people forget. A recorded interview is biometric-adjacent personal data about someone you probably did not hire, sitting in a vendor's system indefinitely unless you set a policy. Decide how long you keep recordings, write it down, and make sure you can actually locate and delete everything you hold on one person when they ask, because that request will eventually arrive and "it is in the interview tool somewhere" is not an answer.

What AI video interviews are bad at

Three things, and none of them are fixable with a better model.

They cannot verify a claim. A candidate saying they managed a team of twelve or hold an active license is a sentence in a transcript, and scoring it well is not the same as confirming it. Verification is a separate step and it stays yours.

They cannot reliably tell whether an answer was written by an AI. Detection tools in this space are not accurate enough to act on, and a false accusation is worse than a missed one. The honest mitigation is structural rather than technical: unscripted follow-ups that depend on what the candidate just said are hard to feed to an assistant in real time, which is one of the clearer arguments for the two-way format. Identity fraud is a related and growing problem. Gartner forecasts that by 2028 one in four candidate profiles will be fake, and in June 2025 the DOJ and FBI acted against a scheme using the compromised identities of more than 80 US persons to obtain remote jobs at over 100 US companies, searching 29 laptop farms across 16 states. That is a forecast and an enforcement action, not a measurement of your pipeline, but it is the reason identity checks are becoming standard.

And they cannot judge fit. Whether someone will work well with a specific manager, on a specific team, in a role that will change in six months, is not observable in a 15 minute screen by any interviewer, human or otherwise. Tools that claim to score culture fit are usually scoring how much a candidate sounds like the people already there, which is the mechanism bias runs on.

Choosing between the two formats

Pick the one-way format when the questions are fixed, the answers are short and factual, and volume is the whole problem. Pick the two-way format when what you need to know sits one question deeper than the first answer, when candidates for the role are scarce enough that the experience matters, or when you want the same probing a decent recruiter would do applied to everyone rather than to the first thirty people who applied.

The buying question that separates vendors quickly is not on most feature grids: can the product ask a question that was not written in advance? Everything else in this category is broadly comparable. A full vendor comparison is on AI interview software, and the category page for this format is AI video interview software.

InterviewAgent.ai runs the two-way version. It conducts the first-round screening interview by voice or video, asks role-tailored questions, follows up when an answer is thin, scores each response against your rubric, and returns a ranked shortlist with the transcript evidence behind every score. Candidates get clear AI disclosure and consent before anything is captured, the screening is bias-audited against EEOC guidance and NYC Local Law 144, and a recruiter makes every decision about who advances. Pricing is public and starts at 149 dollars a month.