Interviewing

The problem it solves

Two problems, really.

The first is scheduling. A screening interview is the narrowest bottleneck in hiring: it needs half an hour of a specific person's time, and every candidate waits behind everybody else's calendar. An AI screen removes the queue for the first round.

The second is comparability. A human round produces whatever the interviewer wrote down afterwards, in a format nobody else uses, covering whatever they found interesting. Two rounds conducted by two people are not measurements of the same thing. Putting both kinds of round through one transcript pipeline and one scorer is what makes them comparable at all.

The interview modes

Each interview stage of a job says who runs its round.

• The AI screen: the AI interviewer runs the round in a browser-based room.
• A human round: one of your interviewers leads it, in the same room with the AI seat empty. If the round's meeting link is a Zoom, Google Meet or Microsoft Teams link, it happens there instead (see Integrations).

A round held anywhere else, in person for example, is filed afterwards by pasting its transcript.

The AI voice screen

The interviewer persona is built from the job's rubric rows — the same rows the assessment scores against — grouped into prompt sections by kind: deal-breakers to verify, required skills and competencies to evidence with specific examples, preferred skills to cover if the round has room, red flags to probe, success signals, and behavioural traits. Heaviest criteria come first. The interviewer probes exactly what the scorer measures.

It also receives the gap brief for this round — what earlier rounds and the résumé left uncovered. See Coaching the round.

Before the candidate joins, they consent to the recording and, separately, to automated assessment. Two consents rather than one, because a tick-box about a recording is not consent to an automated rejection. Both are required; the room refuses without them.

The screen is video. The candidate's camera has to be on: the AI interviewer waits for it and, if it never comes on, ends the session with a spoken explanation. The interview is recorded as video and audio and transcribed.

Only the transcript is scored. The video is never given to a model; it is there so a person can watch the screen back, and it is deleted after 12 months while the transcript and the assessment stay.

The questions it must never ask

One list of nine topics is written down in a single place and read by three different parts of the system: salary history, criminal history, age or graduation year, family or pregnancy or childcare, religion, national origin or citizenship beyond work authorisation, health or disability, union membership, and sexual orientation.

The AI interviewer's prompt states the list as absolute and outranking the job's own must-probe list, and carries a one-line redirect for a candidate who volunteers one of these unprompted. Two questions are deliberately not on the list, and the prompt says so: whether the candidate is authorised to work here, and whether they need an adjustment or accommodation. A rule that flagged the accommodation offer would train the interviewer out of making it.

Afterwards, the transcript is scanned for what the AI actually asked, because a prompt is a request and a transcript is a fact. See Compliance.

Human-led rounds

A human-led round uses the same room. The interviewer opens it from the app; the candidate joins from their own link, which opens that one interview and nothing else. Joining still requires both consents.

Because there is no AI in the interviewer's chair, a separate listener writes the transcript. It is deliberately not a captioner: only finished lines are shown, because the transcript exists so the co-pilot can say what has not been asked and so the round is scored on what was actually said — not so anyone reads captions mid-sentence.

Speaker attribution is two separate questions, answered two different ways. Which seat is ground truth from the room: the interviewer and the candidate hold distinguishable identities. Which voice within that seat comes from the audio itself, because one seat is not always one person — a panel shares a laptop, and so does a candidate with someone else in the room. A second voice on a seat is labelled as such rather than folded into the first. Merging them would turn "someone else in the room said X" into "the candidate contradicted themselves", which is the single thing the co-pilot must never get wrong.

A human round held on Zoom, Google Meet or Microsoft Teams instead of our room can have a note taker join it, record the audio and send the transcript to the same scorer. A round can also be filed from the interviewer's own Granola notes, and a connected Google or Microsoft 365 calendar keeps offered times clear of an interviewer's real meetings. How to set up each one is in Integrations.

Scoring the transcript

Every transcript is scored the same way, however it arrived: the AI screen, our own room, the note taker, a Granola import or a pasted transcript. So analysis cannot behave differently depending on how the round was conducted.

When a round ends, its assessment shows as queued until the score lands a moment later.

Then:

1. Score. The transcript is scored against the job's weighted rubric at the version the interview was conducted against. The model must quote the transcript verbatim for every criterion it returns; criteria it cannot quote for, and keys that are not in the rubric, are dropped. It returns an overall score, a recommendation, and one row per criterion the transcript actually evidenced.

   Uncovered criteria are omitted, never scored zero. Silence about mentoring is not evidence of bad mentoring, and a zero would be indistinguishable from a real failure and would poison the roll-up.

2. Roll up. Every interview assessment for the application is combined into a single cumulative assessment for that candidate on that job: a weighted mean over covered criteria, a coverage percentage, and a per-criterion ledger. A criterion scored in two rounds takes the mean of its scores, not the latest — two rounds probing system design are two samples of one thing, and picking the later one throws away evidence and makes the score jump on the ordering of rounds.

   Coverage is measured against the current rubric version. A criterion scored under an old version and renamed since has a different identity by design, so it stops counting — correct, because a renamed criterion measures something else.

   Nothing in the roll-up decides anything. It reports.

3. Decide. Handled separately, on the job's two numbers. See Deciding.

Scores link back to the rubric criteria they measure and, through the evidence table, to the transcript turns that support them. That chain is what makes an evaluation auditable rather than just a number.

Reading the results

One call's page shows the score on a bar marked with the job's reject and advance thresholds, the recommendation, how much of the rubric's weight the call reached, and the deal breakers met, failed and not reached. Under that come the scorer's verdict and every rubric criterion, heaviest first. A reached criterion shows its answer (yes, partly or no), its score, why, the quoted lines with the time each was said (a click opens the transcript there), when it was asked and what to probe next. A criterion the call never reached is drawn hatched with no number. Beside each criterion sit the résumé's score and earlier rounds' answers. They are shown for comparison and were never given to the scorer.

The band on the page is a label: advance, reject, or "a person decides". Between the thresholds the page also lists each criterion's follow-up question near the top. It never suggests a decision.

The candidate's review page has a picker: one interview shows that call's page, and "All interviews" shows the cumulative view. That view merges the counted rounds' strengths and concerns into one point per criterion (where rounds disagree, the latest counted round's point is kept). It also shows each criterion's answer round by round and the counted calls' metrics added together. All of it is read from stored rows, with no model call.

What the user does

• Pick each round's modality when configuring the job's stages.
• Invite the candidate. For an AI screen this is a link; the candidate picks their own time and consents on the way in.
• For a human round, open the room. The candidate joins from their own link.
• For an interview conducted elsewhere, paste the transcript.
• Read the scored assessment, criterion by criterion, with the transcript moment behind each score.

Limits and edge cases

• Coverage, not completeness. One round covers whatever it had time for. The honest statement of what is still unknown is the coverage number, shown beside every score.
• Only the words are scored. There is no video analysis, no facial assessment and no voice-based personality inference, and none is planned.
• A criterion nobody asked about stays uncovered forever unless a later round asks. That is what the gap brief is for.
• A transcript can be missing. A round completed with no transcript produces no scores, and the application's coverage is unchanged. With no new score there is nothing to decide, so the candidacy waits for a person.
• Speaker labels differ by source. The system recognises several conventions and treats anything unrecognised as neither speaker, which makes a talk ratio unavailable rather than wrong.
• Rescheduling and reminders exist, but a candidate who never books is not chased indefinitely. There is one reminder before the link expires.