How to use AI to summarize interviews: the right summary for each reader

To use AI to summarize an interview, start from a transcript with correct speaker labels, tell the model who the summary is for (the ATS record, a debrief, a hiring manager or a client), give it a fixed structure, and require the candidate's own words for every claim. Keep the summary descriptive, a record of what was said rather than a verdict on the candidate, and check names, numbers and quotes against the transcript before it goes anywhere. That last check is the step people skip, and it is where most summary errors live.

The distinction between describing and judging matters beyond style. As of October 2026, some AI hiring rules treat a tool that records or transcribes an interview differently from one that classifies or scores the candidate, and a summary that says "strong hire" sits on the wrong side of that line. More on that below.

Which kind of interview summary do you need?

"Summarize this interview" produces a generic page nobody uses. Decide who reads it and what they will do next, then pick one of these.

Each has a home page on this site: the ATS format is in how to write candidate notes in your ATS, the in-house version in the candidate assessment summary template, the debrief in the interview debrief template, and the client version in how to turn interview notes into a submittal. Use those as the structure you give the model.

Why should an interview summary describe rather than judge?

Three reasons, in order of how often they bite.

1. Readers anchor on verdicts. A summary that opens "strong candidate, good culture fit" decides the debrief before anyone has read the evidence. A summary that lists what the candidate said lets each reader form a view.
2. Verdicts hide errors. If the model misattributed one answer, a descriptive summary shows the wrong quote where you can catch it. A verdict built on it gives you nothing to check.
3. Evaluation is regulated differently. New York City's rules on automated employment decision tools say the output of tools that "transcribe a video or audio interview" is not the simplified output the law targets, while a score, classification or recommendation can be. The EU AI Act treats AI used to evaluate candidates as high-risk. The statutes are linked in is it legal to use AI note takers in interviews; the practical rule is to keep labels such as "strong" or "not a fit" out of AI summaries and in the hands of the people deciding.

How do you prepare the transcript?

A summary is only as good as its input. Spend two minutes on these before prompting:

• Speaker labels. Make sure each line is marked interviewer or candidate. A wrong label turns your question into the candidate's claim. Tools that capture the interviewer's microphone and the meeting audio separately get this right more often than tools working from one mixed recording.
• Names and technical terms. Fix misheard names of employers, systems and certifications. The model will copy whatever the transcript says.
• Paused or off-record sections. If the candidate asked to pause, check nothing from that stretch is in the transcript.
• One interview per file. Never combine two candidates in one input.

If you do not yet have a reliable transcript, start with how to transcribe a job interview, which covers phone, video and in-person methods and what affects accuracy.

Prompts you can copy

ATS record

Summarize this interview for our ATS record. Use only the transcript.
Quote the candidate word for word where you use quotation marks, and
only quote lines labeled CANDIDATE. Do not rate, rank or recommend.
Leave out anything about age, family, health, religion, nationality,
accent or appearance.

Format, in this order:
ROLE / DATE / INTERVIEWER:
MUST-HAVES (one line each: shown / partly / not shown, with a quote):
- [must-have 1]
- [must-have 2]
- [must-have 3]
MOTIVATION: "[quote]"
PAY EXPECTATION AS STATED:
NOTICE / AVAILABILITY:
OTHER PROCESSES MENTIONED:
NOT COVERED IN THIS INTERVIEW:
CLAIMS TO VERIFY:

Must-haves: [paste]
Transcript: [paste]

Debrief pre-read

Organize this transcript by competency for a hiring debrief. For each
competency below, list up to three quotes FOR and up to three quotes
AGAINST, word for word, with the question that prompted each. If there
is no evidence, write NOT COVERED. Do not summarize the candidate
overall, do not give scores, and do not use evaluative adjectives
such as strong, weak, impressive or concerning.

Competencies: [paste]
Transcript: [paste]

Handoff to the next interviewer

Write a handoff for the next interviewer in under 120 words. List:
(1) topics already covered, (2) answers that were vague or used "we"
without saying what the candidate personally did, with a short quote,
(3) numbers or claims to verify, (4) up to three suggested questions.
Do not include any rating or opinion of the candidate.

Transcript: [paste]

How do you summarize a long interview?

Long transcripts cause two problems: detail from the middle of the conversation gets dropped, and the model smooths over contradictions between an early answer and a later one. For interviews over about 40 minutes, or panels:

1. Split by question or section, not by an arbitrary length. Most transcripts have natural breaks at each new question.
2. Summarize each section with quotes and timestamps, using the same format each time.
3. Combine the section summaries in a second prompt, asking the model to flag any place where the candidate's answers conflict.
4. Read the middle sections yourself. Spot-check that what you remember from the middle of the call made it into the summary.

What should you check before you share a summary?

1. Every quote appears in the transcript, word for word.
2. Every quote is from the candidate, not the interviewer.
3. Names of employers, systems and people are spelled correctly.
4. Every number matches the transcript; anything unverifiable is marked "as stated."
5. Verbs match the candidate's: "helped", "supported" and "led" are different claims.
6. The summary includes what the candidate could not answer, not only what went well.
7. No evaluative labels crept in ("strong", "excellent fit", "red flag").
8. Nothing about protected characteristics or private life is included.
9. The summary goes only to people who need it for this role.

The checks take three to five minutes. A wrong number in a summary that a hiring manager quotes back to the candidate costs far more.

Where should AI summaries live, and for how long?

An interview summary is a hiring record. It is subject to the same retention rules as your notes and scorecards, it can be requested in litigation, and under data protection laws such as the GDPR and the CCPA a candidate can ask to see it; see candidate data access requests. Practical rules:

• Save it in one place, normally the candidate's ATS record, not scattered across email threads and chat.
• Know where the tool keeps its own copy. Many note-taking services store transcripts and summaries on their servers with their own retention settings; where AI note takers store interview data explains how to find out.
• Apply your retention schedule to summaries too. The federal baseline and state minimums are in how long to keep interview notes.
• Do not paste transcripts into personal AI accounts. Use a tool your organization has approved for candidate data.

Summary mistakes, and the fix

If the summary is feeding a scorecard rather than an ATS note, the prompts and checks in how to use AI to write interview feedback go a step further, with evidence for and against each competency and the rating left to you.