What is AI hiring? Using video agents to screen every candidate the same way

Every applicant hopes the first screen judges only their qualifications. Often it doesn't. 

Two candidates with matching qualifications apply for the same customer support role. One gets a Monday morning recruiter who asks four questions and listens to every answer. The other gets a Friday afternoon call from someone on their eleventh screen of the day, and only the first advances.

AI hiring, the use of software to source, filter, schedule, and assess candidates, can apply the same questions and standards to everyone. Live video screening keeps that first conversation consistent, while other uses of AI hiring happen without any conversation at all.

What AI hiring means

AI hiring uses software to find, filter, schedule, and assess job candidates, with different systems handling each stage. Machine learning (ML) ranks candidates against a role's requirements, natural language processing (NLP) reads resumes and answers, and real-time conversation systems hold live screening calls.

A candidate may never notice some of these systems and meet others face to face. A resume parser works silently, while a live, two-way screening conversation sees and hears the candidate, asks questions, and responds in real time. How much a candidate notices depends on where the software sits in the funnel.

Where AI enters the hiring funnel

Software touches an application long before a person does, and each stage changes what candidates experience and what recruiters have to review.

Sourcing and candidate matching

Sourcing systems scan job boards, professional networks, and applicant databases for people whose skills fit a role, including passive candidates who aren't actively applying. 

A retailer staffing 200 seasonal store roles receives a ranked list of past applicants and nearby candidates, each with the reasons they matched. The list is only as good as the requirements behind it, so vague job criteria produce vague matches. Those same requirements then become the criteria for screening incoming applications.

Resume and application screening

For the candidate, application screening is a silent step that often ends in an automatic rejection with no explanation. 

Software parses each application against structured criteria and ranks or filters the pool before a human opens a file. It's one of the most widely used stages: the World Economic Forum reports that more than 90% of employers use automated systems to filter or rank applications. Candidates who clear the filter next face the problem of finding a time to talk.

Interview scheduling and coordination

Scheduling systems handle the availability back-and-forth, confirmations, and rescheduling without a coordinator. 

The candidate sees a link, a few open slots, and a confirmed time, and the recruiter's calendar fills without an email thread. For high-volume roles, that removes days of waiting between application and first conversation. What happens in the booked slot decides how much the recruiter has to review afterward.

Live candidate assessment

In a live structured screen, a video screener asks predetermined questions, listens to each answer, and hands the recruiter a transcript, a summary, and a rubric score. Candidates get room to answer, and recruiters get one standard across every conversation. 

One-way video interviewing works differently: candidates record answers to on-screen prompts, and a reviewer or an algorithm scores them later. Only live interviews lets the screener respond to what a candidate says.

The consistency problem that structured video screening closes

Phone screens vary from recruiter to recruiter. One opens with availability, another with a behavioral question, and a third skips the rubric and goes on feel.

Energy varies too. The same recruiter is sharper at 9 a.m. than at 4 p.m. after ten screens, and candidates who are comfortable with small talk can score higher whether or not the role rewards it.

Order changes outcomes as well. Research published in the Review of Economic Studies found that the candidate evaluated immediately before has roughly three times the influence on an evaluator's judgment as the average other candidate, and that minor changes in ordering can change who is selected.

Most recruiters also carry their standard in their heads. When a rejected candidate asks why, the team has nothing written down to point to.

Software can apply one standard to every applicant, but a script that talks over people creates a new unfairness. A fair screen needs the same questions for everyone and a screener that waits for the answer. That's the job a Personified Application Layer (PAL) is built for.

How a PAL conducts a live screening conversation

Tavus, the human computing company, builds real-time applications that candidates talk to and build a relationship with, ones that see, hear, understand, remember, and respond face-to-face. 

Teams use a PAL for interview practice, first-round AI screening, and candidate questions. In hiring, a PAL recruiter runs an adaptive first conversation that adjusts follow-ups to each candidate's answers.

Four components run as a closed loop during every call, and each one protects something a candidate cares about:

• Time to think: Sparrow-2, a real-time conversational understanding model, predicts who has the floor from raw audio, so a candidate who pauses isn't cut off. It reaches 92.4% end-of-turn recall and 97.4% interruption recall.
• Attention to the whole answer: Raven-1, a multimodal perception system, fuses tone, expression and gaze, so PAL notices when someone is still forming a thought.
• Follow-ups that fit: The large language model (LLM) layer turns Raven-1's plain-language notes and the team's rubric into the next question, so every screen covers the same ground yet responds to the person.
• A face that listens: Phoenix-4.5 renders the PAL's face, head, and upper body in real time and reacts while the candidate talks, so they feel attended to.

Take Maya, who's applying for a customer support role at a regional insurer. Her 15-minute screen has five questions in the same order as every other candidate's. When she goes quiet for five seconds to find a detail, the PAL waits instead of cutting to question three, and when she slows down mid-answer, it holds the follow-up until she finishes. For Maya, that's presence: the sense of being attended to.

Every session ends as structured evaluation data in the applicant tracking system (ATS), with the transcript, so a recruiter reviewing twenty screens reads twenty answers to the same five questions. 

Product teams embed a PAL recruiter through the Conversational Video Interface (CVI) API, and the screen produces evidence while people still decide what to do with it.

What AI hiring does not replace

People decide whether to interview and, later, extend an offer. A PAL recruiter screens: it asks the agreed questions, records the answers, and scores them against the rubric the team wrote.

Questions about the team, pay and start-date negotiation, and judgments about organizational fit depend on context a screening rubric doesn't hold, and the more senior the role, the more of the decision lives there. 

For many high-volume roles, the real alternative to a PAL is a keyword filter or no screen at all. Whichever tool runs the screen, employment law still governs how a team sets it up.

What to put in place before running AI screening at scale

Agree on the questions, scoring rubric, and candidate disclosure before the first PAL screen so recruiters can interpret and compare the results.

• A structured question set tied to the role: Generic screening questions produce generic answers. Questions built from the role's core competencies, such as de-escalation for a support job, produce comparable answers.
• A scoring rubric recruiters agree on before the first screen: If the team disagrees on what a strong answer looks like, the PAL surfaces rankings nobody trusts. Writing the rubric as Objectives forces that disagreement into the open early.
• A candidate disclosure policy: Plain wording about what the PAL will ask and who reviews the result reads as respect. Buried legal boilerplate reads as a warning.

Disclosure is already legally required in some places. Under New York City Local Law 144, employers and employment agencies can't use an automated employment decision tool unless it has had a bias audit within one year of use, information about the audit is publicly available, and candidates or employees have received certain notices.

The Illinois AI Video Interview Act covers AI analysis of recorded video interviews. It requires notice, an explanation, and consent beforehand, plus deletion within 30 days of a request.

Federal anti-discrimination law, including Title VII and the Americans with Disabilities Act (ADA), also applies to screening tools, so those obligations should shape the questions, rubric, and disclosure before screening begins. Sequence the three in that order, because the rubric depends on the questions and the disclosure describes both.

Every candidate deserves the same first conversation

A fair first conversation gives every candidate the same questions, the same standard, and time to be heard. Maya, the insurer's customer support applicant from the screen described earlier, finished each answer before the PAL moved on, so her screen reflected what she said. The next applicant gets the same five questions and the same room to work through a long pause.

Tavus builds human-like PALs that see and hear candidates, ask questions, and respond in real time. Hiring teams that build on them give every applicant a first conversation with presence, the feeling that someone is paying attention.

See it for yourself. Book a demo.