Waterr AI Logo
GuideOctober 27, 2026

AI Interviewer with a Live Avatar: How It Works, What It Scores, How to Run One

An AI interviewer with an avatar is a live two-way interview where an AI asks, follows up and scores against a rubric you wrote. Whether the face matters, the legal basics, and a sample scorecard.

Harshit SharmaFounder & CEO, Waterr AI

An AI interviewer is software that runs a live, two-way interview: it asks your questions, follows up on the answers, and returns a transcript and a score against criteria you wrote. "With a live avatar" means the candidate sees a face instead of a voice orb. The face is optional. The largest field experiment I found on AI interviewing, 70,000 applicants, used a voice agent with no face, and the applicants it interviewed were 12% more likely to get a job offer (Jabarian and Henkel). The rubric, the follow-ups and the paper trail decide whether the interview is worth running. Below: how it differs from one-way video and phone screens, what to score, whether the face matters, the legal basics, and a sample scorecard.

AI interviewer vs one-way video vs phone screen

Three formats get lumped together under "automated video interview". They are different products.

One-way recorded videoPhone screen (human)AI interviewer (live, two-way)
Who asksPre-recorded or text promptsA recruiterAn AI persona, live
Follow-up questionsNoneYesYes, based on the answer
Candidate schedules itAnytimeNeeds a calendar slotAnytime, from a link
Reviewer effortWatch every videoOne call per candidateRead a score and a transcript

The one-way format is cheap to build and miserable to sit through: you talk to a webcam and can't ask what the question meant. The phone screen is good and does not scale. A live AI interviewer adapts like the phone screen and runs on demand like the recorded video.

The developer view is in AI interviewer API and build an AI interviewer. Product pages: screening and interviews.

How one interview runs

On Waterr the inputs are three objects: a persona (who the candidate talks to), a scenario (the script, the type, the time limit) and goals (what the candidate is scored on). The scenarios docs list the types as interview, roleplay, upskill and brainstorm, with a default session length of 30 minutes.

The candidate gets a link. They see a consent screen if you turned one on, check their devices, and the interview starts. Nobody schedules anything.

What comes out, per the post-meeting data docs: the full transcript, a recording (fetched with a separate call), a summary with strengths and growth areas, a score per goal plus an overall average, and speech analytics such as filler words and repetitions. Analysis lands 30 to 60 seconds after the call ends, and you can take it by webhook instead of polling. Scores come from a model reading the transcript against your instructions. That is a first pass for a human reviewer, not a verdict.

The same product can also join Google Meet, Zoom and Teams calls. This guide covers the link-based interview.

Tavus sells a version of this too. Its AI Interviewer page says each session returns structured evaluation data, full transcripts and perception analysis, webhook-ready for an ATS, and its calls to action are "Talk to a human" and "Answer the call". On capability we overlap. I would test how fast a developer on your team gets to a working, scored interview without a sales call.

What an AI interviewer should score, and what it should not

The score is only as good as the rubric. Waterr's goals docs give the rules, and they are good rules for any vendor:

  • Score observable behaviour. "Described a specific past project with their own role" is visible in a transcript. "Seems like a culture fit" is not.
  • Anchor the ends of the scale, or scores drift between calls.
  • Keep it to 3 to 5 criteria. Each is a separate grading pass.
  • Say what to deduct for.

Waterr scores 0 to 5 per goal from the transcript only. Tone, pauses and talk-time are not visible to the grader, which keeps the score on what the candidate said.

What it should not score:

  • Facial expression, eye contact, or "enthusiasm" from video. The EU AI Act prohibits systems that infer a person's emotions "in the areas of workplace and education institutions", except for medical or safety reasons (Article 5(1)(f)). Recital 44 cites "serious concerns about the scientific basis" of emotion inference (recital 44). The text does not say whether a job candidate counts as "in the workplace", so ask counsel before scoring expression on anyone in the EU.
  • Accent, grammar or fluency, unless the job is language. Interviewer.ai's avatar interview help page lists pronunciation, grammar and tone among the areas its avatar assesses.- Anything you cannot explain to the candidate in two sentences.

The AI ranks and flags, a human decides. Treat a low score as a reason to read the transcript, not a rejection.

Does the avatar need a face?

The best critique is a hands-on review in The Register, where the author interviewed with an AI avatar from Job Bolt. It "can't nod, raise an eyebrow, or make any other gesture", he wrote, and the result felt like "participating in an infomercial, not an interview". A face that cannot react makes the candidate perform into a void, the same problem as one-way video with a head on it.

The other end of the argument is Tavus's Griffin launch, which claimed 48% of people who talked to it live thought it was a human. That is a study Tavus designed and ran: 26 of 54 participants on one-minute calls (Tavus, Traictory), and Tavus says the doubters tended to suspect within 20 seconds. Protos reports a Community Note on X saying the results were not independently verified. Griffin-Lite is a research preview that Tavus says is not available to customers, with no API or price listed. The launch post reached about 21 million views by mid-October per the tracker doomers.ai, but the Hacker News thread got 15 points. None of it shows that a lifelike face makes hiring interviews better.

Our position: the behaviour is the product. An interviewer that waits for you to finish and asks a real follow-up beats a prettier face that doesn't.

What ships on Waterr, per the avatars docs: the default is a static face, picked from a catalog with categories such as recruiter and professional, or from your own image. An optional real-time animated avatar switch exists per scenario. I won't claim anything about lip-sync quality or render speed, because those need independent testing.

The legal basics

I am not a lawyer, and this is a map, not legal advice. Three regimes come up most. Status below is as of 27 October 2026.

Illinois: Artificial Intelligence Video Interview Act (820 ILCS 42). Read the statute on ilga.gov. If an employer asks applicants for Illinois positions to submit video interviews and uses AI to analyse them, then:

  • Before the interview, tell the applicant AI may be used, explain how it works and what characteristics it evaluates, and get consent. No consent, no AI evaluation (Section 5).
  • Share the videos only with people whose expertise or technology is needed to evaluate the applicant (Section 10).
  • On request, delete the applicant's videos within 30 days and tell anyone you shared them with to delete theirs (Section 15).
  • If you rely solely on AI analysis to decide who gets an in-person interview, collect race and ethnicity data and report it to the state each year (Section 20).

The Act was written for submitted video. Whether a live, recorded AI interview falls inside it is a question for counsel. I would follow it regardless.

New York City: Local Law 144 of 2021. The law text covers automated employment decision tools: computational processes that issue a score, classification or recommendation used to assist or replace discretionary hiring decisions. Before use, the tool needs a bias audit done no more than one year earlier, a published summary of the results, and notice to NYC-resident candidates at least 10 business days beforehand. The Department of Consumer and Worker Protection page says enforcement began on 5 July 2023.

EU AI Act (Regulation 2024/1689). Primary text on EUR-Lex. Status today:

  • In force: the Article 5 prohibitions, including emotion inference in the workplace and education, since 2 February 2025.
  • In force: Article 50(1) requires that people interacting with an AI system are told so, unless that is obvious. It applies from 2 August 2026.
  • Pending: Annex III, point 4(a) lists "AI systems intended to be used for the recruitment or selection of natural persons" as high-risk, with risk management, logging and human oversight duties. The original start was 2 August 2026, but the AI "digital omnibus" amendment, in force since 27 July 2026, moved Annex III to 2 December 2027 (European Commission, AI Act timeline). Those duties are not yet enforceable.

The practical overlap, stricter than any one regime requires:

  1. Tell the candidate it is an AI before it starts, and say what is evaluated and who sees the output.
  2. Get an explicit yes, and offer a human alternative.
  3. Delete on request.
  4. Test your rubric for adverse impact on a sample before you rely on it.

On Waterr, consent is a pre-join gate: a message you write, shown before the device check, with "I understand" or "Cancel". If the candidate cancels, the AI never receives them. The click is enforced but not logged as an auditable record of who agreed to which version and when. If you need that trail, capture it on your own backend before sending the candidate to the link.

Candidate experience and how long it should run

Candidates accept automated interviews less when the stakes are high, mainly because of lower perceived fairness and social presence (Langer, König and Papathanasiou, 2019). The fixes are cheap.

  • Say what it is. The first line of the invite: an AI will interview you, here's how long, here's what happens next.
  • Offer a human route. Some candidates will opt out. Let them.

Length: Interviewer.ai describes its avatar interviews as 10 to 15 minutes. For a first-round screen I would run 15 to 20 minutes, enough for four or five competency questions with follow-ups. Past 25 minutes it is a second-round interview and you want a human in it. That is my judgement, not a measured number. Set call_duration on the scenario so the cap is enforced.

Build vs buy

AI interviewer API covers when to build versus buy, and build an AI interviewer shows the four-call version on Waterr. Buy when interviews are not your product.

Ask any vendor, including us: can I see a real scored output today, can an engineer start without a sales call, what does consent log, what does a minute cost, and what is the face doing. Comparing against Tavus? Tavus alternatives covers six options, including ours.

An example: first-round screen for a customer success manager

Here is what the inputs look like on Waterr. The script is short and the rubric is where the work goes.

json
{
  "name": "Customer Success Manager, Round 1",
  "type": "interview",
  "persona_id": "PERSONA_UUID",
  "call_duration": 20,
  "visibility": "restricted",
  "welcome_message": "Hi, I'm an AI interviewer. This takes about 20 minutes and is recorded. Ready when you are.",
  "prompt": "You are interviewing for a customer success role.\n\n## Flow\n1. Intro (2 min)\n2. An account they saved or lost (8 min)\n3. A time they said no to a customer (6 min)\n4. Their questions (4 min)\n\n## Rules\n- Ask one question at a time\n- If an answer is vague, ask for a specific example before moving on\n- Never reveal scoring criteria"
}

That follows the scenario docs. The goals are separate objects, created with the goals endpoint. The agent cannot see them, and they never change how it behaves on the call. Copy the scorecard and swap in your own role:

GoalA 5 looks likeA 1 looks likeDeduct for
OwnershipNames the account, their role, the decision they made and whyTalks about "we" throughout, no individual decisionClaiming results without naming their part
EvidenceGives a figure or concrete outcome and how it was measured"It went well"Numbers with no source
Handling a hard customerDescribes a real conflict, what they said, and what changed afterNo conflict, or blames the customerBlame with no learning
ClarityAnswers the question asked, in orderWanders across topicsNot answering the follow-up

Each row becomes one goal, with instructions written in that style. Nothing here mentions "confidence", "energy" or the candidate's face. Every row can be checked by a person reading the transcript, which is what makes the score defensible. Goal feedback includes an example from the call, but per the goals docs it is written to the participant as coaching ("Open with a question before pitching"), not as a hiring verdict. A reviewer reads the score and transcript, and opens the recording only for candidates near the line.

Run the rubric on 10 interviews with known outcomes before it touches a real candidate. If scores disagree with your best interviewer, fix the instructions, then re-score the old calls with POST /analyses/trigger/{meetingId}.

Frequently asked questions

Do candidates prefer avatars to human interviewers?

I could not find a study showing they do, as of October 2026, and the ones I did find lean the other way. In a randomised study of 299 people, Kleinlogel and colleagues (2023) found more negative reactions to an asynchronous avatar-based video interview than to a face-to-face one. The 70,000-applicant field experiment by Jabarian and Henkel (2026) favours AI interviews on offers, but it used a voice agent with no avatar and measured outcomes, not whether candidates liked it. Test your own flow with a few candidates and ask them directly.

Does an AI interviewer need a lifelike face?

No. A static, well-chosen face with a good script and a sharp rubric will serve most screens. A face that animates but can't react, as The Register's reviewer found, can make things worse. Appearance also changes perceived fairness: in a 215-person study, racial mismatch with the avatar heightened perceptions of ethnic bias (Lau et al., 2026). If you want animation, test it on candidate comfort first.

Where this goes

Every candidate gets the same questions, every answer is on the record, and a reviewer reads a scored transcript instead of sitting through a calendar full of screens. Write the rubric first, disclose the AI up front, keep a human on the decision, and the avatar question gets a lot smaller.

AI InterviewerAI Avatar InterviewAutomated Video InterviewCandidate ScreeningHiring Compliance