AI for recruiting: Building video-based screening into your platform

Hiring decisions begin with a human judgment: how someone thinks, communicates, and recovers when a conversation takes an unexpected turn. Application forms can't show that, so the first wave of AI for recruiting parsed resumes and ranked keywords, scaling the paperwork while leaving the conversation untouched. 

The screening interval between application and recruiter conversation stayed text-based or asynchronous, meaning most of what candidates communicate never reached the decision-makers.

Live video brings that missing conversation into a recruiting platform. The product can observe how candidates reason, hesitate, and recover before a human recruiter joins, which requires precise timing, relevant follow-up questions, and the kind of presence that makes someone willing to speak honestly.

What AI-driven candidate screening actually does

For an applicant, the screening stage sits between submitting an application and speaking with a recruiter. AI-driven candidate screening handles work in that interval, including ranking resumes, booking interviews, matching candidates by semantic search, and answering applicant questions through a large language model (LLM).

Fewer than half of organizations are expected to use AI in HR in 2026, and recruiting is where that use is most concentrated, at 27%. Automating resume and scheduling tasks does not recover the signals that disappear when screening remains text-based.

Why text-based screening loses most of what candidates actually communicate

Text shows what a candidate chose to write down. A live conversation shows how they reason when a question isn't the one they prepared for and whether their tone matches their words.

Keyword matching misses all of this. A transcript drops the hesitation, energy, communication style, and tone. Asynchronous video records those signals but often goes unwatched by any human reviewer. Live video lets you react while the candidate is still talking, making perception and response part of the screening experience.

How generative AI changes what screening has to measure

Candidates can polish resumes and practice answers with the same generative models platforms use to screen them, so written materials converge on one baseline. A live conversation resists scripting because the next question depends on the last answer. That shift changes what screening has to capture:

• Reasoning in the moment: Written materials show prepared answers; a live exchange reveals how a candidate thinks when the next question is unexpected and how they recover when their rehearsed path falls apart.
• Delivery under adaptive questioning: Pre-recorded answers can be retaken until they sound smooth, but a live follow-up probe tests whether tone, pacing, and confidence hold when the interviewer picks up on a detail the candidate didn't anticipate.
• Consistency across turns: A single transcript can be edited into coherence, while a multi-turn exchange exposes contradictions between examples and lets the interviewer surface them with one clarifying question.

Preparation still helps; the rubric simply scores how a candidate reasons when answers can't be pre-written. Measuring that fairly depends on what the platform allows its models to infer and score.

How PALs capture richer candidate signal in a live screen

Capturing richer signal requires perception, conversational timing, and structured scoring to work as parts of the same live exchange. Tavus is the human computing company building Personified Application Layers (PALs), real-time applications candidates talk to that see, hear, remember, and respond. 

A PAL runs four components as one closed loop: Raven-1 handles perception, Sparrow-2 provides real-time conversational understanding, the LLM layer reasons about what to say next, and Phoenix-4.5 handles real-time human rendering.

Multimodal perception during a live screening session

An applicant for a field-service role describes working alone in well-worded sentences, but her tone is flat, and she glances away from the camera. Raven-1, the multimodal perception system, fuses those signals with her words and catches the mismatch, describing it in plain language as "confident phrasing, hesitant delivery." 

That description shapes the follow-up question but stays out of the score. Raven-1 runs on sub-100ms audio perception, with rolling context no more than 300ms stale.

Conversational flow and knowing when to respond

Many spoken dialogue systems wait for a fixed stretch of silence before responding, while human responses in a landmark study of ten languages peaked within 200 milliseconds of a question's end. A silence timer can cut off a candidate before they recall a start date.

Sparrow-2, the real-time conversational understanding model, replaces the timer with continuous floor-ownership prediction on raw audio at a 10-millisecond frame rate. On TurnBench, it scores 92.4% end-of-turn recall and 97.4% interruption recall. Those predictions let the LLM layer prepare a response early, then commit or discard it as the prediction updates.

Visible attention mid-sentence

The response has to look attentive while the candidate is still speaking. Phoenix-4.5, the real-time human rendering model, renders the face, head, and upper body, so a nod appears mid-sentence rather than after the candidate finishes. Visible body language tells a candidate the interviewer is following along, not just waiting for a turn to speak.

Structured scoring against a rubric

Candidates need to be compared on consistent evidence. The PAL scores answers against a rubric built from job analysis, the foundation of a structured interview.

Dana, applying for a warehouse lead role, meets measurable Objectives: two examples of resolving a shift conflict and confirmation of her forklift certification. Knowledge Base returns the matching rubric anchor in about 30 milliseconds through its retrieval-augmented generation (RAG) model, and when her second example contradicts her first, the LLM layer asks one clarifying probe before scoring anything.

How to build video-based screening into your platform

Building a defensible live screening layer requires decisions about bias, candidate trust, access, and compliance before the first interview goes live. The following tips cover what belongs on the backlog alongside latency and turn-taking.

1. Separate perception from scoring

Bias enters a screening layer through what it's trained on and what it scores. Federal guidance under the Americans with Disabilities Act (ADA) warns that facial and voice analysis may screen out candidates with autism or speech impairments.

Perception should shape conversation flow, such as timing a follow-up or handing off to a person. Only rubric-scored job content should shape the ranking.

2. Design for candidate trust

Gartner's candidate survey found that only 26% of applicants trust AI to evaluate them fairly, and 25% trust an employer less when AI evaluates their information.

Measure drop-off at each stage of the screen rather than a single completion rate, so you can see where trust breaks. Presence belongs on the backlog alongside latency and turn-taking.

3. Audit training data on a schedule

Historical hiring patterns can re-encode themselves into any model trained on them. Audit training data before launch and at set intervals afterward, and send borderline scores to a human reviewer rather than letting the model close the loop on its own. 

Those safeguards must carry through from model design to practical implementation, which means documenting what the system scores and what it ignores so a candidate who asks why they were screened out gets a defensible answer.

4. Meet access and compliance requirements

Candidates should join from a browser. Web Real-Time Communication (WebRTC) transport handles real-time video delivery, and relay servers support networks that block direct connections. 

Product flows must cover notice, consent, data deletion, and bias audit requirements in the jurisdictions where you hire, such as Illinois and New York City. SOC 2 Type II provides an independent audit of security controls, and eligible Enterprise plans include Zero Data Retention.

5. Build disclosure into the candidate experience

Disclosure has to be built into the candidate experience, not treated as a separate consent form. 

PAL Maker includes a no-code Interviewer starter template, and each PAL's configuration includes verbal and visual disclosure settings, so candidates know they're talking to a PAL from the first moment of the conversation. Teams can also bring their own LLM through an endpoint that returns compatible chat completions, keeping their scoring prompts and rubric logic in place.

6. Plan for scale and implementation speed

Once access, compliance, and disclosure are covered, scale becomes an economic question. If your platform runs 10,000 first-round screens a month, every minute a recruiter doesn't spend on them is capacity you can price.

As a named example of implementation speed, Final Round AI, a candidate-side mock-interview product, runs on CVI. The team went from concept to live product in 30 days and reached 100,000-plus active users in Q2 2025.

Candidates can tell when something is actually listening

Candidates can tell whether a system is waiting for a silence timer or following what they mean, and that difference determines whether the screening exchange feels like another automated hurdle or a conversation someone is willing to finish. Presence is what turns the first interval between application and recruiter into usable signal rather than a form candidates fill out alone.

At Tavus, the focus is human computing: giving recruiting platforms a more attentive, human-like way to meet every applicant, including the many who would otherwise encounter a form or an unwatched one-way recording before any person.

See it for yourself. Book a demo.