A hiring manager opens an AI interview summary and sees a confident recommendation. One interviewer has not submitted feedback. Another rated a different role. A third left a short concern that disappeared in the summary. The text reads well, but it is not a reliable account of the interview process. Start by defining what the summary is allowed to represent before connecting a model to Lever.
This guide describes a custom workflow around submitted feedback. It separates Lever’s documented form and visibility settings from additional controls your integration should implement. The goal is a readable review packet that preserves uncertainty, rather than a new automated hiring decision.
Start with the feedback form
Lever lets administrators create feedback forms in Settings, then Feedback, using New Form. Questions can use formats such as text, multiple choice, yes or no, checkboxes and scorecards. A question marked required must have a response before the interviewer can submit the form. These settings give your summary workflow a defined input instead of an unstructured pile of comments.
Agree on the questions before adding AI. Ask for evidence tied to the actual role. Separate what the candidate said or demonstrated from the interviewer’s interpretation. Give interviewers a place to describe uncertainty. A required box containing a vague sentence is still weak evidence, so form completion and evidence quality need separate checks.
Lever feedback forms include a four-point rating question that cannot be deleted. Keep that rating attached to its original form. Do not ask the model to manufacture a replacement score or make scores from differently defined questions look comparable.
Define the packet before generating text
Choose a role-specific unit of work. A person can be considered for different positions, so a general candidate identifier is not enough to define the summary. Record the intended opportunity, interview stage and forms included. Our guide to Lever candidates and opportunities explains why person identity and role-specific history need separate handling.
| Packet item | What to retain | What to avoid |
|---|---|---|
| Submitted response | Original wording and its question | Replacing evidence with a model conclusion |
| Interviewer assessment | Who supplied it and the associated form | Presenting one opinion as panel consensus |
| Missing form | Expected interview and incomplete status | Inventing a response or rating |
| AI draft | Input version, review state and audience | Sharing a draft as an approved decision |
Keep the expected interview list separate from the submitted form list. Without both, a summary can look complete simply because every retrieved form was processed. Define who owns missing feedback and how a late submission changes an already reviewed packet.
Preserve disagreement and missing evidence
Give the model a narrow job: organize the evidence under the agreed questions. Each material statement should point back to a submitted response inside the internal review interface. This reference is for your reviewers, not a public citation or a reason to copy sensitive interview text into broad notifications.
Require the draft to distinguish observation, interviewer judgment and unresolved question. If two interviewers disagree, retain both positions and the evidence each supplied. Do not blend them into a neutral sentence that hides the conflict. If no interviewer assessed a requirement, mark it unassessed instead of extrapolating from a different skill.
Use a review rule for unsupported assertions. A statement that cannot be traced to the selected inputs should be removed or rewritten as a question for the recruiter. Treat text inside feedback as source material, not permission for the model to send messages, change settings or perform candidate actions.
Carry visibility rules beyond Lever
Lever’s company settings let administrators configure who can see submitted interview feedback. A separate option can hide others’ feedback from Team Members and Limited Team Members on the interview panel until they submit their own. The documented custom-role permission to view other users’ interview feedback can supersede the default company visibility configuration.
Check the actual roles and settings used in your account. Do not describe the hide-until-submitted setting as a universal barrier for every user. More importantly, do not assume it protects a summary your own integration sends outside Lever. That new copy needs its own access checks.
Apply the intended audience at retrieval and again when a person opens the result. Check notification recipients, search indexes, cached summaries and downloaded review packets. If an interviewer should not see previous feedback yet, a summary containing that feedback should not appear in their chat notification. Have an administrator verify the rule with representative accounts.
Put review before distribution
- Collect the current submitted forms for the selected opportunity and stage.
- Compare submitted forms with the expected interview plan.
- Generate an internal draft with missing evidence and conflicts clearly labeled.
- Have the assigned recruiter compare the draft with the original responses.
- Confirm the recipient list and release the approved version.
- Invalidate that approval if the underlying forms or intended audience change.
Keep approval tied to the exact draft. A recruiter reviewing one version has not approved a later revision that includes a new interviewer’s feedback. Store enough version information to explain what was reviewed without duplicating all sensitive content into ordinary application logs.
Leave hiring decisions in a separate authorized step. Approval to share a summary is not approval to advance, reject or contact a candidate. If the integration cannot determine whether a downstream action succeeded, reconcile the destination before retrying rather than assuming nothing happened.
Test the failure cases operators will meet
Build a test set containing a missing form, conflicting feedback, a changed question, two roles for one person and a permission change after a summary was generated. Include an interviewer who has not submitted their own feedback. Verify the content they can actually retrieve, not just whether a button is hidden.
Check that a corrected response replaces an outdated summary. Verify that removing a person’s access also affects saved links and cached copies. Decide how your external copies follow the organization’s retention process. These are implementation requirements to agree with the team, not guarantees supplied by a summary model.
Common mistakes
- Using a summary as proof that all interviews are complete.
- Mixing feedback from separate opportunities for the same person.
- Converting uncertainty into a score or a confident recommendation.
- Sending hidden feedback to an interviewer through another tool.
- Leaving an approved summary current after source feedback changes.
How Remqa can help
Remqa builds AI agents for recruiting teams around the systems operators already use. A feedback workflow can be scoped around form collection, evidence review, access checks and a controlled handoff to the recruiter. Bring a representative interview plan and your current visibility rules. Start with a reviewed packet before adding candidate actions.