Live interview copilot

A live interview copilot is a real-time transcription aid that proposes question prompts and drafts structured notes while an interviewer leads the session. The quality outcome is narrow and operational: every note that ships into the hiring record must cite a transcript span. If the audio is unusable, those note fields stay empty. The copilot does not auto-score. It does not invent a competency rating. The interviewer still owns the scorecard.

Treat this as a writing and recall aid sitting next to a human-led conversation, not as a substitute for structured interviewing. It is a different job from ai screening interviews, which usually sit earlier in the funnel and often run at higher volume, and from async video interview analysis, which scores or summarizes a recording after the candidate has already left. In a live session the interviewer is present, accountable, and the person who will defend the rating in debrief.

The copilot can surface a follow-up when a rubric cell is still empty. It can quote what was said. It cannot decide that the candidate exceeds on systems thinking.

Before anyone joins, load the competencies, behavioral anchors, and planned probes for this requisition. The prompt bank should be the scorecard, not a generic interview script. If the copilot is working from a different list than the panel will use in Greenhouse, Lever, or Ashby, it will pull the conversation toward questions that never land in the rating form.

Preparation is part of the how-to, not a nice-to-have. Confirm which competencies must have evidence before a hire recommendation is allowed. Confirm which topics are out of bounds. Confirm what enough signal means for this level, so the copilot does not keep pushing filler questions after the interviewer already has a defensible sample. ATS records in tools such as Greenhouse, Lever, and Ashby are typically where that scorecard already lives. Point the copilot at that definition. Do not maintain a second, unofficial rubric in a sidebar.

A loaded rubric also constrains prompts. When the copilot suggests a question, the interviewer should be able to map it to a cell they still need. A clever prompt with no home on the scorecard is noise. Decline it and stay on the planned structure.

Start transcription when the interview starts. The useful artifact is a time-aligned transcript the interviewer can point at later, not a polished essay generated while they are still talking.

Audio fails in ordinary ways: speakerphone, two people talking over each other, a bad VPN, a second interviewer who never comes off mute, a candidate thinking out loud in a language the model was not prepared for. When you cannot point at words, you do not invent a paraphrase and file it as evidence. Empty stays empty. That is the quality rule, not a fallback you apologize for in debrief.

The failure mode to train against is a note with no transcript cite. A line such as strong ownership of the incident, with no span, is the interviewer's memory wearing a structured-note costume. Memory is allowed in the scorecard, because the interviewer owns the score. It is not allowed to masquerade as cited evidence. If the transcript is garbage, write nothing in the evidence fields and say so: audio unusable, score from live impression only, no quoted support.

Interview platforms in the same hiring stack, including HireVue and peer tools in that class, may already hold recordings or structured artifacts from other stages. A live copilot should not silently overwrite those records or pretend it watched a different session. Its job is notes tied to this conversation's transcript, for this interviewer, on this scorecard.

After the call, or in a brief pause during it, draft notes that attach each observation to a transcript span: speaker, timestamp or turn range, short quote or close paraphrase, and which competency the quote is offered as evidence for. Then stop. The interviewer scores.

Here is a single illustration, not a case study. A hiring manager is running a live product-sense loop for a PM role. The loaded rubric asks for problem framing, tradeoff reasoning, and stakeholder communication. The candidate walks through a pricing change. The transcript captures, around minute eighteen, that they would ship the cheaper plan first because sales already promised it, then measure churn. The copilot drafts a note under tradeoff reasoning, citing that span, and suggests a probe: what you would have done if sales had not promised it. The interviewer asks that probe, hears a thin answer, and after the call marks tradeoff reasoning as below the bar. The note did not contain a rating. The rating came from the interviewer, who used the cite plus the follow-up they themselves ran.

Two other failure modes show up at this step. Treating notes as the score looks like pasting the copilot bullets into the ATS comment box and skipping the competency cells, or copying a model-written overall hire into the recommendation. Inventing a competency rating looks like filling Collaboration because the form had a blank and the transcript never discussed collaboration. Both break the outcome. Cited notes can be empty. Scores cannot be hallucinated to complete a template.

If a panelist missed a stretch of the call, they still do not get to adopt the copilot's implied verdict. They can read cited spans they did not hear and decide whether those spans are enough for them to score, or they recuse from that competency. Completeness of the form is not a reason to fabricate a number.

A live copilot does not replace structured interviewing, and it does not replace the rest of select-stage evidence. structured resume scoring belongs earlier, against a job-specific rubric, not as a live-call substitute. After the loop, a hiring decision bias audit is the place to check whether panels used the scorecard they loaded or drifted into gut feel, proxy variables, or borrowed notes with no cites.

Operationally, interview-ops should treat the copilot as a controlled input to the ATS packet: transcript when usable, cited notes or an explicit empty, and interviewer-owned scores. Train panels on the three failure modes above until they can catch them in calibration. If a debrief cannot point from a rating to a span or to a live impression the interviewer will own out loud, the rating is not ready to move a candidate.

Enablement owners should also decide who is allowed to see the raw transcript. A copilot that drafts notes for the assigned interviewer is a different risk posture from a copilot that publishes a full recording to every stakeholder. Default to the narrower share: the people who were in the room, plus whatever your existing interview-packet policy already allows.

That is the whole loop: load the rubric, transcribe, draft notes with cites, interviewer scores. Anything that auto-completes a competency the conversation never touched is out of scope for a quality outcome.