AI interview note summarisation.

What a good interview summary contains

An interview summary is a record of evidence, organised by the criteria the role was advertised against. For each criterion it lists what the candidate said, with a short quote and a timestamp, and notes where the interview did not cover the criterion at all. It does not say whether the candidate is good. That distinction is the whole design: AI is reliable at organising what was said and unreliable at judging people, so the workflow keeps it on the first job.

Recruiters gain twice. Write-ups that used to take 25 minutes take 8, and the notes sent to a client or hiring manager are consistent from one consultant to the next. The interviewer also stops writing from memory at the end of a long day, which is where most inaccurate notes come from.

How the workflow runs

Inputs

The interview recording or meeting transcript (with consent), the interviewer’s own notes, the job description and scorecard criteria, and the candidate’s CV for reference only.

Systems

Teams, Zoom or a phone recording app, a transcription model, a language model, and the applicant tracking system or CRM where approved notes are stored.

Steps

Transcribe with speaker labels; map candidate answers to each criterion; extract supporting quotes with timestamps; list criteria not covered; remove content touching prohibited grounds; produce a draft for the interviewer.

Human review

The interviewer checks every point against the recording, corrects speaker errors, adds their own rating separately, and approves. Panel members rate independently before seeing each other’s notes.

Outputs

An approved evidence summary per candidate, a list of gaps to probe in a second interview, and client-ready submission notes for agencies.

Measures

Write-up minutes per interview, corrections per summary, criteria left uncovered, and complaints or access requests about notes.

Risks

Misattributed quotes, smoothed or exaggerated answers, verdict language creeping in, and sensitive personal details recorded unnecessarily.

Illustrative example: a Tauranga agency with 8 consultants

Illustrative example, synthetic numbers: a Tauranga recruitment agency runs about 160 interviews a month, averaging 45 minutes. Consultants spend about 25 minutes writing up each one.

Assumptions, at prices listed on 19 October 2026: transcription with OpenAI’s gpt-transcribe at US$0.0045 a minute (OpenAI pricing). Summaries with Claude Sonnet 5.5 at US$2 per million input and US$10 per million output tokens (Anthropic pricing), assuming 15,000 input tokens (a 45-minute transcript plus scorecard and instructions) and 1,200 output tokens. If interviews run on a platform that already transcribes, the transcription line disappears.

NZ privacy and fair hiring rules

• Purpose and confidentiality. The Privacy Commissioner’s recruitment guidance says to share applicant information only with people directly involved, to protect applicants’ identities (especially internal candidates), and not to use recruitment information for any other purpose without the applicant’s express consent. A summary forwarded to the wrong client breaks all three.
• Access. Under the Privacy Act 2020, candidates can request the personal information held about them, including interview notes. The same guidance warns that referee comments may have to be disclosed on request unless they were given in confidence, so keep referee material out of interview summaries.
• Prohibited grounds. Section 22(2) of the Human Rights Act 1993 makes it unlawful for anyone procuring employees for an employer to treat a job seeker differently on a prohibited ground, and section 23 covers inquiries that suggest an intention to discriminate. If a candidate mentions age, family plans or health, the summary should leave it out unless the candidate raised it to ask for support, such as an accommodation.
• Retention. The guidance says unsuccessful candidates’ applications should be securely destroyed unless they gave prior consent to be kept on file. Recordings should usually go first, once notes are approved.

What interviewers keep

Interviewers decide what an answer shows, rate candidates against the criteria, and make or recommend the hiring decision. They also decide what goes to a client. Our guide to summarising documents with citations explains why every claim in the summary should carry its timestamp: it makes checking fast enough that people actually do it. Our AI governance work covers the policy that sits around it.

Where summaries mislead

• Wrong speaker. Diarisation errors put the interviewer’s description of the role into the candidate’s mouth. Check every quote the summary relies on.
• Smoothing. Summaries remove hesitation and caveats, so “I helped with the budget once” becomes “managed budgets”. Instruct the model to quote, not paraphrase, for experience claims.
• Verdict drift. Phrases like “strong culture fit” appear unprompted. Ban evaluative adjectives in the prompt and reject drafts that contain them.
• Names and te reo Māori. Transcription often mangles names, iwi and place names. Correct them before notes go anywhere, because a misspelt name in a client submission reflects on the agency.
• Silent gaps. A criterion never discussed can be filled with CV content. Keep the CV as reference and mark uncovered criteria explicitly.

Summaries feed candidate communication drafting, and old notes are part of what recruitment CRM data cleanup has to manage. The recruitment sector guide and our workflow automation service cover the build.

Written by the Levy Street team in Wellington. Last updated 19 October 2026.