Skip to main content
An interview is a real-time voice conversation between a calibrated AI interviewer and one participant. It runs in the browser or over a phone call, lasts 10–45 minutes depending on the template, and ships a scored transcript within 2 minutes of ending.

​
What makes it an interview, not a chatbot

Adaptive follow-up

Probes vague answers, asks for real examples, and adjusts depth based on what the participant just said. No fixed question list.

Voice-first, sub-second

Real voice latency (<800ms) means the AI feels like a person on a phone call, not a chatbot streaming tokens.

Calibrated scoring

Every dimension you defined on the template gets graded against your rubric — same bar every session, every region.

Authenticity-aware

Tab-switch, paste, voice-spoofing signals captured continuously and surfaced for human judgment.

​
Channels

ChannelWhen to useLatency
Web (browser)Default. Candidate clicks a join link, AI joins immediately. Camera optional.<800ms
Outbound phone (SIP)High-volume call-center hiring, or when participants can’t install/permission a browser.<1s
Embedded widgetInside your own product or job page — see Embed guide.<800ms
The session record is identical across channels — same scorecard format, same transcript shape, same webhook payload. Pick the channel that matches where your participants are.

​
Languages

intervyo.ai conducts interviews in English, Japanese, Hindi, Spanish, and several more, with rubric-aligned scoring per language. Set the language on the evaluation template (language field) or per-session by passing it to the create-session endpoint.
Scores are directly comparable across languages — the AI uses the same rubric dimensions and weighting regardless of which language the interview ran in.

​
Lifecycle in one diagram

  ┌──────────────┐   created    ┌──────────────┐   joined     ┌──────────────┐
  │   scheduled  │ ───────────► │  in_progress │ ───────────► │   completed  │
  └──────────────┘              └──────────────┘              └──────────────┘
         │                              │
         │ candidate doesn't show       │ candidate hangs up early /
         ▼                              ▼ tech failure
  ┌──────────────┐              ┌──────────────┐
  │   cancelled  │              │    failed    │
  └──────────────┘              └──────────────┘
See Sessions for what’s emitted at each transition and what fields are populated.

​
What gets captured

Every interview produces a structured session record with these fields:
​
transcript
string[]
Time-stamped speaker turns. AI and participant labeled separately.
​
recording_url
string
Presigned URL to the audio recording. Short-lived (1 hour by default); request a fresh URL each time you need access.
​
score
number
Overall rubric score 0–10. Weighted average of per-dimension scores.
​
evaluation_breakdown
object[]
Per-dimension scores with reasoning paragraphs and transcript citations. See Rubrics for the shape.
​
authenticity_signals
object
Tab-switch count, paste events, voice-spoofing flags, screen-share detections. Surfaced verbatim — no auto-pass-fail interpretation.
​
ai_feedback
string
Two-paragraph plain-English summary: what the participant did well, what the next-round interviewer should probe. Generated against your rubric.

​
Re-running an interview

Interviews are immutable once completed or failed. To run another one for the same participant — say after a technical glitch — create a new session. The participant’s prior sessions stay attached for history.
Multi-stage templates auto-create a fresh session for each stage. You don’t need to manually re-create sessions when a candidate progresses through stages — see Templates.
Last modified on June 3, 2026