Fairness, explainability and data
How Panayam's AI interviews and scores candidates
Statement 1.2 · 11 November 2026. Applies to every interview run on Panayam. Each clause describes what the software does today, not what we intend.
1. What the AI does, and does not do
- It conducts the first-round interview. A conversational AI interviewer asks the role's approved questions by voice, follows up on the candidate's answers and keeps to the time limit.
- It produces a scorecard. After the call, a separate scoring step rates the transcript against the role's rubric and writes an assessment with evidence.
- It never decides. No candidate is advanced, held or rejected by the software. A recruiter reviews every scorecard and the hiring team makes the decision. The interviewer never tells a candidate their score or outcome, never coaches, and never goes off the record.
- Nothing runs before two people sign off. The questions and rubric for a role are approved by the recruiter and the hiring manager. Any edit cancels both approvals, and each approval is logged with a timestamp.
2. How a score is made and explained
- One rubric per role. Every candidate for a role is scored against the same weighted dimensions and the same written anchors for low, middle and high.
- Evidence for every dimension. Each score carries verbatim quotes from the candidate, tied to the moment in the recording, so a reviewer can check it.
- A written recommendation, not a verdict. The scorecard gives an executive summary, a recommendation from Strong yes to Strong no, and a separate list of red flags. The reviewer may disagree with any of it.
- Call health is shown. Connection problems, long silences and technical faults during a call are reported on the scorecard, so a weak interview is not mistaken for a weak candidate.
- On-camera presence. A still frame is taken from the recording about once a minute. The frames are shown to the reviewer on the scorecard and rated for engagement, eye contact, framing and composure. For professional roles the rating carries 8–12% of the total. For high-volume roles it carries no weight: the total comes from the transcript only. At walk-in events, where the venue is the employer's, only what the candidate controls is rated.
3. What is never scored
The scoring instructions prohibit scoring on age, ethnicity, gender presentation, perceived disability, religious dress or skin tone. The rubric builder refuses any dimension that would score a protected characteristic. Background noise and other people's voices are not penalised; a second voice coaching the candidate is noted, not scored.
4. Consent and the right to a person
- Before joining, every candidate ticks an explicit consent to an AI-conducted, recorded interview, written for the Data Privacy Act of 2012 (RA 10173). The interview cannot start without it.
- The consent text says that a human recruiter reviews the scorecard and that no hiring decision is made solely by AI.
- A candidate may decline and ask for an interview with a person by replying to the invite email. The recruiting team, not the software, answers.
5. Data: what, where, who
- Conduct during the interview. The employer chooses, position by position, whether this applies; it is off unless chosen, and it never applies at a walk-in, where the candidate is with the employer's staff. Where it applies, the candidate is told before joining. While such an interview runs, a still frame from the camera is checked by software about every 45 seconds for two things only: a candidate who is undressed, and an interview being made a joke of. The check is told to ignore dress that is religious or cultural, simple or worn clothing, disability, appearance and the room. Longer answers are checked for text read aloud from an AI assistant, and the page reports if it was left during the call. These frames are not stored. Depending on the employer's choice for the position, the software either only notes what it found for the recruiter, or can also remind the candidate once and stop the interview and pass it to a recruiter. It cannot reject anyone: a recruiter watches the recording and decides.
- What is collected. Name and contact details, the resume if provided, the interview audio and video recording, a transcript, periodic video frames for professional roles, and the scorecard.
- Where it is held. Google Cloud, asia-southeast1 (Singapore): a managed database for records and scorecards, and object storage for recordings and frames.
- Processing during the call. Live audio is processed by speech-recognition, language and voice providers to run the conversation. We name the providers to customers under agreement; they are not named publicly because they can change.
- Who can see it. Named recruiters at the employer under role-based access; recruiters see only roles they own or collaborate on. Sensitive reads and all changes are written to an activity log.
- Retention. Records are kept for the employer's hiring process. Automated deletion after a fixed period is not yet in place; it is on our roadmap pending the employer's retention policy. A candidate who has not been interviewed can be deleted by the employer's administrator. Deleting the record of an interviewed candidate (recording, transcript and scores) from the console is planned; until then such a request is carried out by Panayam on the employer's written instruction.
6. Security and operations
- Named logins with recruiter, user-admin and super-admin roles. Sign-in is by email and password; a second factor at sign-in is planned.
- Active sessions can be viewed and ended by an administrator. An activity log records sign-offs, sensitive reads and changes.
- Maintenance mode stops new interviews before a software update so that live calls can finish.
- Automated health checks and alerting run on every service.
7. Changes and contact
This statement is versioned. When the software's behaviour changes in a way that affects a clause, the clause and the version number change with it. Questions: info@panayam.net.