Nine hundred applications, four backend roles, two recruiters, thirty-minute phone screens. The arithmetic doesn't close, and that is the moment most teams start shopping for an AI video interview. An AI video interview is a recorded or live video conversation in which software, rather than a person alone, asks the questions, transcribes the answers and produces a score or summary that a recruiter uses to decide who advances. That's the whole idea. The implementation details are where teams get hurt, and this guide is mostly about those.
Adoption is no longer a frontier question. SHRM's State of AI in HR 2026 report, based on 1,908 HR professionals surveyed between 5 and 23 December 2025, found 39% of HR functions had AI in place, with recruiting the single most common practice area at 27% of use cases. On the candidate side, a Greenhouse survey of 2,950 job seekers published on 1 May 2026 found 63% had already sat an AI interview, up thirteen percentage points in six months. Whatever your position on this, your candidates have met it.
How do AI interviews work under the hood?
Strip away the marketing and the pipeline is short. A question set is defined, either fixed or generated from the job description. The candidate's audio and video are captured. Speech recognition produces a transcript. A model scores that transcript against a rubric, sometimes with prosodic features like pace and pause length layered in. The output is a numeric score, a set of competency ratings, or a written summary with quoted evidence.
The important shift over the past five years is that serious video interview technology now scores language, not faces. HireVue removed facial analysis from its assessments in March 2020 and announced it publicly in January 2021, with CEO Kevin Parker saying visual analysis no longer significantly added value once natural language processing improved, and that public concern about it wasn't worth carrying. If a vendor in 2026 still pitches you on micro-expressions or eye contact scoring, that tells you something about the vendor.
What are the three formats of AI video interview?
Teams meet three distinct things under one label, and conflating them causes most of the bad buying decisions.
Live AI-assisted keeps a human interviewer in the room. The AI transcribes, drafts the scorecard, flags where the interviewer skipped a competency, and sometimes suggests follow-ups. Nobody is being judged by a machine; a recruiter is being relieved of note-taking.
One-way recorded with AI scoring is the classic automated video interview. The candidate gets a link, records answers to fixed prompts alone, and software scores the result. No human is present at any point unless the candidate advances.
Fully automated conversational is the newest of the three. An AI agent conducts a real-time voice or video conversation, asks adaptive follow-ups based on what the candidate just said, and probes when an answer is thin. It feels like an interview. It is also the format with the most ways to go wrong.
| Live AI-assisted | One-way recorded + AI scoring | Fully automated conversational | |
|---|---|---|---|
| Human present | Yes | No | No |
| What AI does | Transcribes, drafts scorecards, flags gaps | Scores recorded answers against a rubric | Asks adaptive follow-ups, scores in real time |
| Recruiter time per candidate | 30–45 min | 3–6 min of review | 3–6 min of review |
| Best for | Final rounds, senior roles, anything with negotiation in it | High-volume fresher and campus screening with clear right answers | Mid-volume roles where follow-up probing genuinely changes the verdict |
| Fails at | Volume — it saves note-taking, not calendar hours | Anything requiring a follow-up question; candidates who freeze on camera | Ambiguous or highly specialised domains where the agent can't judge a partial answer |
| Candidate reaction | Broadly accepted | Worst of the three | Mixed; better than one-way when latency is low |
| Regulatory exposure | Low to moderate | High | High |
The pattern is consistent. The more the machine decides, the more it saves and the more it costs you elsewhere.
One thing recruiters discover late: the live AI-assisted format is the one that changes day-to-day work most and gets talked about least. It doesn't reduce headcount and it doesn't compress the funnel, so it never makes the slide. What it does is end the practice of writing up interview notes from memory two days later, which is where a startling amount of hiring bias enters the record in the first place.
Where does AI evaluation genuinely help?
Structured consistency is the real prize, and it's undersold. A human panel asks the fourth candidate of the day a different, lazier version of the question it asked the first. AI interview software doesn't get tired at 4pm and doesn't warm to a candidate who went to its college. Every applicant gets the same prompts in the same order, which is the precondition for comparing them at all.
Coverage at volume is the second. In Indian campus and fresher hiring, where a single IT services drive can generate application volumes no recruiting team can phone-screen, an automated video interview is often the difference between assessing everyone and assessing whoever applied in the first six hours. SHRM's 2026 recruiting benchmarking brief, drawn from 4,657 members, put the median time-to-fill for non-executive roles at 39 calendar days and found that organisations using advanced technologies including AI filled roles roughly five days faster. Five days is real but modest — worth knowing before you build a business case on speed alone.
Third, and least discussed: transcripts. Every AI interview produces a searchable record of what was actually said. When a hiring manager rejects someone for "communication," you can go read the answer. That single artefact improves hiring debates more than any score does.
See Xara AI interview a candidate live
Structured questions, adaptive follow-ups, a transcript and a scorecard your team can argue with. Book a 30-minute demo — no slides.
Book a demo →Where does AI interview scoring mislead?
The common belief that AI removes bias from hiring is backwards. It relocates it, and makes it harder to see.
A 2025 study by Mujtaba and Mahapatra tested a state-of-the-art multi-task neural network that predicts personality traits and interview scores from video. Using StyleGAN2, the researchers generated counterfactual versions of the same candidates with altered apparent gender, ethnicity and age, then re-scored them. In the baseline data, African-American candidates received a mean interview score of 0.472 against 0.520 for Caucasian candidates, and candidates aged 40 and over scored 0.494 against 0.521 for younger ones. Visual-only variants of the model performed at 91.1% against 91.3% for the full multimodal system, while text-only dropped to 88.0%. In other words, the model was reading faces far more than words, and its verdicts moved when nothing but appearance changed.
There's a subtler failure that matters enormously in India and anywhere else hiring across accents. Scoring language fluency as a proxy for competence penalises the candidate who thinks in Tamil or Marathi and answers in English, and rewards the confident generalist who says little at speed. A rubric built on "clarity of communication" will do this quietly and consistently unless you deliberately weight it down for roles where English fluency isn't the job.
Then there's gaming. Candidates now run AI assistance during one-way recordings, and the tell is a suspiciously well-structured answer with no specifics in it. Automated scoring rewards exactly that shape. Adaptive conversational formats resist this better, because a good follow-up question exposes a candidate who is reading rather than remembering.
Finally, false precision. A score of 72 out of 100 looks like a measurement. It is a model's guess, compressed. Treat it as a sorting aid, never as a threshold you fire people over, and never use AI scoring as the deciding input on senior or leadership hires. At that level you're assessing judgement under ambiguity, and there is no rubric for it.
What does an AI video interview cost you in candidate goodwill?
More than most teams budget for. In the same Greenhouse research, 38% of candidates said they had walked away from a hiring process that included an AI interview, and another 12% said they would. Seventy percent were never clearly told upfront that AI would evaluate them; 21% only worked it out when the interview began. After completing one, 51% received no feedback at all. Only 21% believed most employers use AI in hiring responsibly.
The most-cited reason for abandoning a process, at 33%, was a pre-recorded video interview scored by AI with no human present anywhere in the loop. That's the format most teams buy first because it's cheapest. Twenty-seven percent cited the absence of any disclosure, and 26% cited being monitored during the process — proctoring, tab-switch detection, that family of features. Fifty-seven percent thought disclosure should be a legal requirement rather than a courtesy.
The candidates you lose to this are not distributed randomly. The ones with three other live processes running walk away first, which means the format quietly filters for people with fewer options.
The same research found the effect cuts both ways: 38% said a well-executed AI interview improved their view of the employer, against 34% who came away with a worse one. The variable isn't AI. It's disclosure, feedback and whether a human ever appears. Candidates asked for specific things — 44% wanted upfront disclosure, 46% wanted the option to request a human interview, 38% wanted human review before an AI-influenced decision.
What does the law require in 2026?
Under the EU AI Act, recruitment and candidate-selection systems sit in Annex III as high risk. The obligations were set to apply from 2 August 2026, but the AI Act simplification package — endorsed by the European Parliament on 16 June 2026 and given final approval by the Council on 29 June 2026 — moved the Annex III compliance deadline to 2 December 2027. Do not read that as a reprieve. Article 50 transparency obligations still apply from 2 August 2026, meaning people must be told when they're interacting with an AI system, and the Article 4 AI literacy duty remains in force. The substantive high-risk work — risk management, logging, human oversight, technical documentation — has moved, not vanished.
In New York City, Local Law 144 has required an annual independent bias audit and candidate notice for automated employment decision tools since 2023. Enforcement has been thin: a New York State Comptroller audit published on 2 December 2025 found that the Department of Consumer and Worker Protection reviewed 32 companies and identified one compliance issue, while the auditors examining the same companies found at least 17 instances of potential non-compliance. Two complaints were received during the audit period. Weak enforcement is not the same as weak law, and that gap tends to close.
Illinois has the longest-standing rule specific to this technology. The Artificial Intelligence Video Interview Act, in force since 2020, requires employers to tell applicants when AI is used in a video interview, explain how it works, and obtain consent before proceeding. From 1 January 2026, amendments to the Illinois Human Rights Act under HB 3773 extend notice duties to any AI that influences or facilitates employment decisions, including the product name and vendor, with a four-year record retention requirement.
If you hire into any of these jurisdictions, the rules apply regardless of where your recruiting team sits. Most multinationals are settling on one policy built to the strictest standard rather than three.
Does your hiring volume actually justify it?
Run this before you buy anything.
- Count screening interviews per quarter per open role. Under about 15 applicants screened per hire, AI-powered hiring tools cost more in setup and goodwill than they return.
- Check whether your bottleneck is screening or scheduling. If offers stall at final round, an AI screen fixes nothing.
- Identify roles where the right answer is knowable. Fresher engineering, support, sales development, field roles — yes. Design leadership — no.
- Price the goodwill. If you're hiring in a market where your employer brand is still being built, a one-way scored interview is expensive in ways your ATS won't show you.
- Decide who reviews before rejection. If the answer is nobody, you've bought a liability, not a tool.
Two practical things that decide whether this works. Cap it at five questions with ninety-second answers; completion falls off a cliff beyond that, and nobody watches minute four. And allow at least one retake per question — candidates recording on a phone over mobile data, often during a lunch break while still employed and serving notice, will otherwise send you a first-take panic response that tells you nothing about them.
Xakal, a hiring platform launched in March 2024 with a built-in ATS, runs its Xara AI Interviews product on this model, with enterprise volume pricing starting at ₹100 per interview and a free trial that lets candidates practise one interview a day — worth noting mainly because the practice side is where candidate anxiety about the format actually gets addressed. You can see the approach at thexakal.com.
The teams getting this right treat an AI video interview as a first filter that a human signs off on, disclosed clearly, with feedback going back to everyone who completed one. The teams getting it wrong bought a scoring engine and stopped answering email.