How to Run a Fair First-Round Voice Screen With an AI Interviewer

When an AI first-round screen fits, and when it does not

Picture a 30-person startup in Bengaluru hiring its first customer support team. The founder posts four openings and gets a few hundred applications in a week. Resumes for support roles look alike, and the skill that matters most, explaining something clearly to an upset customer, does not show on paper. A 15-minute phone call with every plausible applicant would eat the founder's month.

This is where a structured voice screen earns its place. Every candidate gets the same setup, speaks in their own words, and your team reads the conversation later. It is not right for every role, though.

A useful test: if you would be happy for a junior recruiter to run this round from a fixed brief, a structured AI screen is probably reasonable. If the round needs judgment, persuasion or a relationship, keep it human.

Setting up the screen: step by step

The steps below use InterviewRoom as the example because that is the product we build, but the decisions apply to any AI interviewing tool.

1. Rewrite the job description first. The interviewer builds its questions from the JD you paste in. A JD that says "excellent communication skills, team player" produces generic questions. One that says "you will handle 40 chat tickets a day about failed payments and explain refunds to customers" produces questions about that work.
2. Choose the track, level and length. Pick the track that matches the role (student, software engineering, or a general role), the seniority, and a length between 10 and 25 minutes. For a first round, shorter is kinder to candidates and still gives you a usable transcript.
3. Add a coding round only if the job codes. It is optional. The interviewer sets one small problem, the candidate talks it through and submits code, and it is discussed rather than executed. Leave it off for support, sales or operations roles.
4. Add candidates. Enter one email, or paste a list or upload a CSV with up to 200 at once. Everyone in a batch gets the same interview setup and start time. Each person gets a private link that only opens for the email address you entered, and you can download all the links as a CSV.
5. Send the links with your own note. Do not send a bare link. Use the template below, through your own email, applicant tracking system or placement portal.

An invite note you can adapt

Subject: Next step for the Customer Support Associate role

Hi Priya,

Thanks for applying. The next step is a short spoken interview
with an AI interviewer, about 15 minutes, in English.

- Your private link: <link>
- It opens 15 minutes before <date and time>.
- You need a laptop with Chrome or Edge and a microphone. No camera.
- Our team will see your transcript and a score. A person on our team
  reads every transcript before any decision is made.
- You can practise in the same kind of room for free first.

If you would prefer to speak with a person instead, or need any
adjustment, reply to this email and we will arrange it.

Thanks,
<name>

Making it fair: notice, consent and a way out

A screen is only as fair as the process around it. These are the basics we recommend for any AI-assisted interview.

Tell people before, not after

Candidates should know they are talking to an AI, what it will produce, and who will see it. In InterviewRoom the invite page states that the company will see the score and the full transcript, and the candidate confirms that before the interview starts. Repeat it in plain words in your own email too, so nobody finds out at the last moment.

Offer another route

Some candidates will have a speech difference, a hearing impairment, unreliable internet or no laptop. Some will simply prefer a person. Offer a human call or another adjustment on request, and do not mark anyone down for asking.

Keep the setup identical

Fairness in a structured screen comes from consistency: same role, same level, same length, same JD for everyone being compared. Batch scheduling helps here. If you change the setup halfway through a drive, compare the two groups separately.

Be clear about data

Tell candidates what is kept. In InterviewRoom, speech is transcribed during the interview and the audio is discarded. The transcript, the score and any submitted code are kept, and the candidate can read their own feedback too.

What to review after each interview

After each interview you see its status, a score out of 5 on four dimensions (clarity, specifics, structure and bar fit), short notes on what landed and what to fix, and the full transcript. The score is written from the transcript, so the transcript is the evidence and the score is a summary of it.

A review order that avoids anchoring

1. Read the transcript first. If you see "2 out of 5" before you read, you will read looking for reasons it is a 2.
2. Write your own one-line verdict. "Explained the refund flow clearly, vague on the escalation example."
3. Then open the score. Where it disagrees with you, go back to the transcript and check who is right.

What to look for, by dimension

Calibrate before you scale

For the first ten or so interviews in a new role, have two people read each transcript separately and compare notes. You will quickly see whether your team agrees on what "good" sounds like, and whether the JD needs another pass.

Treat technical problems as no signal

A very short transcript, a session that ended after a long silence, or answers that read like garbled words usually mean a microphone or connection problem, not a weak candidate. Speech-to-text can also mishear names, accents and technical terms. Read for meaning, and offer a second attempt or a human call when the transcript cannot tell you anything.

What not to automate

An AI interviewer is good at running the same structured conversation many times and writing it down. It does not make hiring decisions, and InterviewRoom is built so that it does not: it gives your team evidence and leaves the decision with you. Keep these steps with people:

• Rejections. No candidate should be rejected because a number fell below a cut-off. A person reads the transcript behind every rejection.
• Automatic thresholds. "Everyone under 3 gets the decline email" turns a summary into a decision. Use scores to sort your reading order, not to skip reading.
• Later rounds. The final conversation is where candidates meet their future manager and judge you back. That needs a person.
• Adjustment requests. Anyone asking for a different format should get a quick human reply.
• Offers and feedback. If you share feedback, have a person write or at least approve it.

A simple rule of thumb: the AI can move candidates toward a human conversation, but only a human moves a candidate out of the process.

A one-week plan for your first AI screen

Here is how the startup from the start of this article could run its support hiring round:

Before you send the first invite, check

• The JD describes real tasks, not adjectives.
• Your invite note says it is an AI interviewer, how long it takes, and who sees the results.
• There is a named person who handles requests for another route.
• You have run the interview yourself at least once.
• Your team agrees that scores sort the reading order and never make a decision alone.

Hiring teams and placement cells are priced by volume. If you want to see the setup with your own role, book a walkthrough. The engineering story behind the product is in our InterviewRoom case study.