How to use AI to write interview feedback without losing the evidence

To use AI to write interview feedback, give the model three things: the interview transcript or your notes, the scorecard with its agreed competencies and rating anchors, and instructions to quote the candidate for every point and to write "not assessed" wherever there is no evidence. Then check every quote against the source, choose each rating yourself, and remove anything that is not job-related. The AI drafts and organizes; the interviewer rates, recommends and signs the feedback.

This page covers the AI part of the job: what to hand over, a prompt you can copy, how to check the draft, and where the model should stop. The underlying method of writing feedback from observations is in evidence-based interview feedback, and weak and strong examples by competency are in internal interview feedback examples.

What can AI do in interview feedback, and what should stay with you?

AI is good at the tedious half of feedback: finding every place a competency came up in a 45-minute transcript, lining up quotes, and turning fragments into readable sentences. It is bad at the half that matters: deciding what the evidence is worth. Split the work before you open any tool.

Keeping the rating and recommendation with a person is not only good practice. In several places, software that scores or recommends candidates triggers notice or audit duties, and how much weight the score carries is part of the test. The guide to using AI in recruiting responsibly sorts these uses by risk.

What do you need before you start?

• The source. A transcript is best, because the AI can quote from it and you can check the quotes. Notes work, but the AI can only organize what you wrote down.
• The scorecard with anchors. Competencies, the questions for each, and what each rating level looks like. Without anchors the model will describe the candidate in its own terms, which nobody agreed. The interview scorecard template has the format.
• The candidate's agreement to capture. If the transcript exists, the candidate should have been told and asked first, before the first question.
• An approved tool. A tool your organization has cleared for candidate data, with terms on retention and model training you have read. Not a personal chatbot account.
• Your own recollection, written first. Three lines on what stood out, written before you read any AI draft. More on why below.

The workflow, step by step

1. Write three lines from memory before you open the AI

A fluent draft quickly overwrites what you remember. Before you paste anything into a tool, write down the two or three moments that struck you, positive or negative, and anything you meant to probe and did not. This takes two minutes and gives you something to check the draft against: if the AI never mentions the answer you found most telling, you will notice.

2. Give the model the scorecard and the transcript, with strict rules

Paste the competencies and anchors first, then the transcript, then the instructions. The prompt in the next section asks for evidence for and against each competency, verbatim quotes only, and no ratings. Run one candidate per session so details from two interviews cannot blend.

3. Check the draft against the source

This is the step that makes the feedback yours. Every quote gets searched in the transcript, every number confirmed, every summary verb compared with what the candidate actually said. The checks are listed in a table further down.

4. Rate against the anchors yourself

Read the evidence the draft collected, read the anchor for each level, and pick the level the evidence matches. Then write the "not higher because" and "not lower because" lines in your own words. If you disagree with how the draft framed something, the framing goes, not your judgment.

5. Edit for what does not belong

Transcripts capture everything, including the candidate mentioning their children, a health appointment or their age. A model asked to summarize may carry that into the feedback. Delete it, along with comments on accent, appearance or personality that are not tied to a competency. The full list is in what not to write in interview notes.

6. Submit before you hear anyone else's view

Feedback written after a hallway conversation with another interviewer is no longer independent. Submit your scorecard first, then go to the debrief. Do not paste other interviewers' notes into your prompt for the same reason.

A prompt you can copy

Adapt the bracketed parts. Keep the rules section as it is; each line prevents a specific failure.

You are helping an interviewer organize evidence for their interview
feedback. You do not rate or recommend.

ROLE: [job title]
COMPETENCIES AND ANCHORS:
[paste each competency with its rating anchors]

TRANSCRIPT:
[paste the transcript with speaker labels: INTERVIEWER / CANDIDATE]

RULES
1. Use only the transcript. Do not add facts, context or assumptions.
2. Quote the candidate word for word, in quotation marks. Never
   paraphrase inside quotation marks. Only quote lines labeled CANDIDATE.
3. For each competency, list evidence FOR and evidence AGAINST.
4. If a competency has no evidence, write NOT ASSESSED. Do not guess.
5. Where the candidate says "we", note it and quote what they said about
   their own part, if anything.
6. Do not give ratings, scores, rankings or a hire recommendation.
7. Do not mention age, family, health, religion, nationality, accent,
   appearance or anything else not related to the competencies. If the
   transcript contains such details, list them separately under
   "REMOVE BEFORE SHARING" so the interviewer can check.
8. Keep each evidence line under 40 words.

OUTPUT FORMAT
[Competency]
  For: "[quote]" — [one-line description of what they did]
  Against: "[quote]" — [one-line description]
  Gaps: [what was not asked or not answered]
...
REMOVE BEFORE SHARING: [list]
OPEN QUESTIONS FOR NEXT ROUND: [up to three]

If you only have notes, replace the transcript block with your notes and change rule 2 to: "Quote only phrases that appear in quotation marks in my notes. Everything else is a description, not a quote."

How do you check an AI draft before you submit it?

Run these checks in order. Most drafts fail at least one.

Some interview tools do the first check for you. Interview Signal, for example, removes any scorecard quote that does not appear in the transcript. That catches invented quotes; it does not catch a real quote placed under the wrong competency or a summary that overstates it, so read the draft anyway.

What does good AI-assisted feedback look like?

An invented example. The candidate, Priya, answered a question about a project that slipped.

The AI draft, before checking:

Ownership
  For: "I led the recovery plan and got us back on schedule" — took
  charge of a failing project and delivered on time.
  Against: none found.

Searching the transcript shows two problems. The candidate said "I helped put together the recovery plan," not "I led" it, so the quote was paraphrased inside quotation marks. And four minutes later she said the launch "went out about three weeks late in the end," which the draft missed entirely.

The feedback the interviewer submitted:

Ownership: 2
Evidence: "I helped put together the recovery plan" (her part was the
revised test schedule, when asked); "it went out about three weeks late
in the end."
Not higher because: her own role was one part of a team plan, and the
date was missed; she did not describe what she would change.
Not lower because: she named her specific contribution without being
pushed twice, and was open about the slip.

The final version is shorter than many AI drafts and far more useful in a debrief. Another interviewer can disagree with the 2, but they are arguing about the same facts.

When should you not use AI to write the feedback?

• The candidate declined capture. Then there is no transcript, and you should not reconstruct one. Write from your handwritten notes.
• The tool is not approved for candidate data. Draft from your notes by hand rather than pasting a transcript into an unapproved service.
• You did not run or attend the interview. Feedback is a first-hand account. Summarizing someone else's transcript and submitting it as your view breaks the independence of the panel.
• The feedback is for the candidate. Candidate-facing feedback carries legal and relationship risk that internal feedback does not. AI can help with tone, but the content needs care; see interview feedback emails to candidates.
• You are tempted to let it decide. If you find yourself adopting the draft's framing because it sounds confident, stop and reread the anchors.

Common mistakes, and the fix

The same approach works for summaries that go elsewhere, such as the ATS record or a debrief pre-read; see how to use AI to summarize interviews. If your team's ratings still vary a lot between interviewers after this, the problem is usually the scale rather than the writing, and how to score interviews consistently covers the fixes.