Human review
How to review an AI-generated employee feedback summary.
A summary can make dozens of conversations easier to navigate. Before sharing it, work through the claims that could affect the next decision. This review rubric helps an authorized reader check coverage, trace claims, preserve dissent and resolve errors.
- Review unit
- Each claim likely to affect the next decision.
- Common correction
- Replace a broad conclusion with the group, observed behavior and limits the accounts actually support.
- Release decision
- Mark each claim ready, revise or hold, then record the authorized reviewer’s decision.
Start with coverage
Write down how the group was selected, who was invited and who took part. Check which roles, shifts or locations could be missing. Participation counts and what people chose to discuss describe the available accounts. They cannot establish the view of everyone invited.
If the summary says “managers regularly reported that training was too short,” identify the group and period, then check how many distinct managers actually described the training length. If one participant described it in three answers, count that as one participant. Preserve a low count or missing group where it changes what someone could reasonably decide.
Trace the claims back to the original accounts
Use a short worksheet. Give every major claim a row with its scope, source references, counterexample, risk and decision. Check whether each source describes the same event, a comparable event or a different issue altogether. Mark each source as a direct account, something the participant learned from another person or an interpretation. Keep internal references within the agreed access boundaries.
As you verify each row, read enough before and after the relevant passage to see what question was asked. An answer to “what would you change?” may describe a suggestion. It is evidence that someone proposed an idea. It may say little about whether the problem occurs.
Try a correction on a fictional claim
Take the draft claim “New hires could not get the equipment they needed.” The source material contains two examples of delayed badges, one example of a laptop ready on day one and no responses from the night shift. One delayed badge prevented an employee from reaching the relevant work area. The other had no reported impact on planned work.
A revised claim could read, “Two starters described delays receiving a badge. One said the delay prevented access to their work area during the first shift. One starter reported their laptop was ready. Night-shift starters did not take part.” That gives the owner a route to check the badge handoff and keeps separate categories from becoming a universal claim.
Inspect the summary for sensitive or identifying details
Look at names, exact titles, unusual incidents, a combination of dates and locations, and quotes a colleague might recognize. Check the audience, permission to share, and the appropriate channel for sensitive matters. A summary about a small group may need fewer details or a different audience even when all names are removed.
Check for evaluative labels the original participant never used. If a summary describes an employee as resistant, trace that characterization before it is reused. Document the behavior described and the circumstances. The established people process determines how to handle any individual matter.
Apply a simple review disposition
Mark a claim ready when the source supports the wording, the scope is clear and the proposed audience can receive it. Mark it revise when stronger context or narrower wording can address the issue. Mark it hold when the source cannot be checked, the material conflicts on an important point, or it needs an authorized route before sharing.
Record the reviewer, date, original wording, edited wording and reason in the agreed workspace. When a held claim affects the decision, ask what further evidence could change the disposition. NIST’s generative AI profile identifies plausible false output and over-reliance as risks worth evaluating in context. A review record makes those risks concrete for the team responsible for the use.
Frequently asked questions
Questions that come up
How much of the summary should I verify?
Start with every claim that could change the decision, harm someone or expose sensitive details. Use additional sampling to check the rest of the summary and increase scrutiny when you find a pattern of errors.
Can I use another AI tool as the reviewer?
Automation may help organize references or flag inconsistencies. An authorized human reviewer still needs to examine decision-relevant claims and access, and decide what can be shared.
What should I do when the accounts disagree?
Preserve the difference if it changes the interpretation. Check whether the people described the same event, could see the same facts and used terms the same way.
Talk to Lontra HR
Bring one real HR question.
See how a first campaign could work for your team.