Use an AI agent to draft interview scorecards from the transcript
An interview scorecard agent reads a finished interview and drafts a scorecard against the role's agreed criteria. Each criterion gets the candidate's own words as evidence and a suggested rating. The useful version stops at a draft: the interviewer checks it, changes what's wrong and owns every score and the hiring decision.
The job: a draft scorecard, not a verdict
Interviewers often fill in scorecards hours later, from memory and a few notes. By then the overall impression has replaced what the candidate actually said.
This agent drafts the scorecard from the transcript instead. For each criterion, it quotes what the candidate said and suggests a rating. The interviewer reviews it, fixes it and submits their own scores.
Inputs, output and what it must not do
| Part | What it should be |
|---|---|
| Input | The finished interview's transcript and notes, plus the role's criteria and rating scale |
| Criteria source | Pasted into the agent's instructions, or a scorecard file the agent can search |
| Output | One row per criterion: evidence quotes, a suggested rating, and gaps where no evidence came up |
| Must not | Recommend hire or no-hire, rank candidates, or comment on anything outside the criteria |
| Human step | The interviewer checks each quote against the transcript and sets the final ratings |
Start from agreed criteria
The agent is only as fair as the criteria you give it. OPM describes structured interviews as measuring job-related competencies, with every answer judged on the same rating scale and standards.
So agree the criteria and the scale before the first interview, usually at the intake meeting. Write what a weak and a strong answer sounds like for each one. Without that, the agent invents its own standard.
Keep it to the criteria
A transcript holds far more than job evidence. Write these limits into the instructions, and check every draft against them:
- No hire or no-hire recommendation, and no ranking against other candidates.
- No comments on age, family, health, accent, appearance, nationality, religion or other personal characteristics.
- No ratings for "culture fit", confidence or likeability unless they are written criteria with defined answers.
- No rating without a quote. If a criterion never came up, it says so and leaves the rating blank.
- Small talk and anything said before or after the interview questions is ignored.
Instructions you can adapt
This is a starting point, not a built-in template. Test it on interviews you remember well.
- If the transcript is not a job interview, reply "Not an interview" and stop.
- Use the criteria and 1–4 scale in the file "Scorecard – <role>". Score nothing that isn't on it.
- For each criterion, quote up to two things the candidate said, with the speaker and rough time.
- Suggest a rating only when a quote supports it. Otherwise write "Not covered".
- List follow-up questions for criteria with thin evidence.
- Never recommend a hiring decision or comment on the candidate personally. Return a draft for the interviewer.
Set it up in Bridge
Custom agents need a paid plan (Professional or Business). You create one in four steps: Identity, Abilities, Tools and Triggers. Paste the instructions in Identity, which takes up to 4,000 characters. Every agent can read transcripts.
For longer criteria, upload the scorecard to your Knowledge Base in Account settings → Context → Files. It accepts PDF, DOCX, CSV, TXT, MD, JPEG and PNG up to 10 MB. In Tools, turn on Files search and give the agent only that file. File storage is not on Starter.
Keep the interviewer in control
"Return a draft" is an instruction, not a hard limit. The hard limits are the abilities and tools you turn on.
Tell candidates the interview is recorded and transcribed, and that an AI drafts the first scorecard. Follow your company's rules on consent and how long candidate data is kept. Check with HR or counsel whether local rules on AI in hiring apply.
- Leave Connectors off. The agent then has no route to an applicant tracking system, email or chat.
- Leave Web search off, so it can't look the candidate up.
- Leave Hearing access off. Tone and sentiment are not criteria, and the scorecard should rest on words you can check.
Pick a trigger
Custom agents and agentic workflows such as the after-every-meeting trigger both come with Professional and Business, not Starter.
- @mention: tag the agent in the interview's Assistant once it ends. It runs only when you ask, so use only this trigger until you have tested the agent.
- After every meeting: the agent runs automatically against the recap and transcript as soon as each meeting ends, interview or not. Before turning it on, run the agent on a non-interview meeting and confirm it replies "Not an interview" and stops.
Example: one criterion in a draft
This example is illustrative. Acme is hiring a support lead, and one criterion is prioritization. The draft quotes the candidate: "I sorted the queue by customers blocked from working, then by account size". It suggests a 3 of 4.
The interviewer remembers a follow-up answer the draft missed, where the candidate explained the trade-off to her manager. She adds that quote and changes the rating to 4.
Check the first drafts
For the first few interviews, read each draft beside the transcript before you submit scores. Watch for four failures:
- A quote that is paraphrased, trimmed out of context or said by the interviewer.
- A rating with no quote behind it.
- A comment on the candidate that isn't tied to any criterion.
- A criterion marked covered when the question was never asked.
