How AI interview scoring works
What happens to a video interview answer after it is recorded, how the three-grader panel works, and what is never automated.
This page explains how InsightHire turns a recorded interview answer into a score, and where people stay in charge. For the candidate notice and the human review option, see AI interview fairness.
From answer to score
- Transcript. The answer is transcribed to text.
- Contact details removed. Email addresses, phone numbers, web links and ID numbers are removed before any AI reads the transcript.
- Scored against your rubric. The AI grader scores the answer from 0 to 100 against the criteria you set for that question, and writes a short summary, strengths and concerns. Your organization's scoring calibration (Harsh, Balanced, or Generous, plus any score adjustment) is applied here.
- Interview score. The answer scores are combined, weighted by each question's points, into the candidate's interview score. Video behaviour analysis is shown for context only and never counts toward the score.
The three-grader panel
Next to the normal grader, InsightHire runs a panel: three independent AI graders that each read only the transcript, and a supervisor that combines their grades.
- Three lenses. One grader looks for concrete evidence (specific actions and results), one for relevance to the question, and one for clarity of reasoning.
- Told what to ignore. Every grader is told to ignore accent, grammar slips, filler words, and anything about who the candidate is.
- Independent. The graders do not see each other's grades.
- Supervisor. The supervisor reads the three grades and their reasons and gives one final grade from 1 to 5, noting why the graders disagreed when they did.
- Disagreement flag. If the graders are more than 1.5 points apart (out of 5), the answer is flagged Needs human look.
Today: running in the background
Right now the panel runs alongside the normal grader. Its result is saved for comparison only. Your scores and score gates still come from the normal grader, and you will not see panel results in reports.
When the panel becomes the score
Our team compares the panel with the normal grader on real answers. We switch to panel scoring only after the panel has scored at least 200 answers and the two agree closely (on average within 0.3 points out of 5, with disagreement flags on no more than 1 in 10 answers). InsightHire staff make the switch, per organization or for everyone. You'll be told before it changes for your account.
When the panel is the score:
- The answer score is the supervisor's grade, shown on the usual 0 to 100 scale.
- The report notes that a panel scored the answer.
- A flagged answer shows Needs human look. A score gate never rejects a candidate because of a flagged answer: the candidate moves on and your team reviews.
- If the panel cannot finish (for example, an AI service is down), the normal grader scores the answer instead, so no candidate waits.
- Every score records which method produced it, for audits.
The human review option
If a candidate chooses Prefer a human review, no AI scores their answers, neither the normal grader nor the panel. Their answers are transcribed for your team to read, marked Needs review, and no score gate fails them automatically. See AI interview fairness.
What is never automated
- Hiring decisions. Offers and final hiring decisions are made by people. AI scores support your team's judgement.
- Rejections on disagreement. When the panel graders disagree, the answer goes to a person, not to an automatic rejection.
- Candidates who asked for a human. They are never AI-scored.
- Protected characteristics. Graders are told to ignore anything about who the candidate is, and contact details are removed before scoring. Use adverse impact testing to check outcomes across groups.
Score gates you add to a journey can move a candidate on, or stop them, based on the score. You choose whether to use them and where to set the threshold.
