Before using AI to write an employee-feedback summary, build a map from proposed themes to the comments that support them. Keep mixed views and less common concerns visible, verify what the counts describe, and only then draft the management brief. A readable paragraph should be the final presentation of the analysis, with an evidence trail behind it.
This matters when HR or learning teams decide what to change after a survey, training pilot or onboarding programme. “People want more training” could conceal several different requests: time to attend, more advanced practice, or materials they can actually use. Each points to a different response.
The workflow below is our editorial recommendation, illustrated with fictional feedback. It is not a validated survey-analysis method or a report of an independent product trial. It gives a team a concrete way to inspect AI assistance before relying on its summary.
Research supports checking which opinions survive compression
In Polarity Calibration for Opinion Summarization, published at NAACL in June 2024, Lei and colleagues found that tested summarization systems could amplify the dominant positive or negative orientation and neglect less common opinions. Their experiments concerned product reviews and political articles. This identifies a failure mode worth examining; it does not estimate errors in today’s employee-survey products.
Wang and colleagues’ November 2025 FacSum paper explicitly organized opinion dimensions before generating summaries. Its evaluations included election and COVID-19 microblog datasets, with additional paper-review experiments. The authors reported improved opinion coverage, but also noted that overlapping dimensions could create redundancy or skew attention. Making categories explicit helps inspection; the categories still need scrutiny.
The same research group’s EvioSum paper at WSDM 2026 connected supporting and divergent evidence to summarized opinions. It used microblog benchmarks and a Vicuna-7B backbone, with automated and human evaluations. It provides another research design for retaining disagreement, rather than independent replication of FacSum or proof of workplace effectiveness. None of these papers validates the particular workflow proposed here.
Practitioner guidance offers a complementary caution. Nielsen Norman Group’s Accelerating Research with AI, first published in September 2024 and reviewed in January 2026, recommends treating generated codes and clusters as an initial pass. Its UX research guidance describes omissions and mistaken groupings. That is practice guidance, not a controlled employee-feedback experiment or a guarantee about a named product.
Define the question and the permitted evidence
Start with the decision the feedback can inform. For a training pilot, ask which changes would make the next session more usable. Avoid asking AI to identify the organization’s “real culture” from a small selection of optional comments. A collection of statements can support observations about those statements without establishing what everyone believes.
Record the survey question, collection period, inclusion rules and number of usable comments. Keep blank, excluded and unavailable records accounted for separately. Preserve a read-only source copy under the appropriate access controls, and give each comment a stable reference such as C01. An internal reference allows authorized reviewers to recover the source; it need not expose the employee’s identity.
Use only an organizationally approved tool and data-handling arrangement. Remove unnecessary identifiers before analysis, then check that preparation has preserved relevant meaning. A job title, location or distinctive incident can identify someone even when their name is absent. Access to the detailed evidence map should be narrower than access to the final brief.
These are proposed operating rules for this workflow. They do not replace the organization’s privacy, employee-listening or research procedures. If the team cannot establish that the tool and use are approved, resolve that before submitting real comments.
Code statements before drafting the story
Ask the assistant to propose short labels for what each comment actually says. A label, often called a code, marks a relevant idea in the text. Define what qualifies for each label, retain the supporting passage and allow one comment to receive several labels.
For example, “The recording helped, but I had no time to practise” contains both a useful feature and a barrier. Forcing it into one positive or negative bucket discards information. Preserve the relationship between the two observations instead of averaging them into a neutral sentiment.
A bounded instruction could be: “For each supplied comment ID, propose relevant codes and copy the exact supporting span. Allow multiple codes. Mark uncertain interpretations for review. Do not infer identity, motives, prevalence outside this collection or recommended actions. Return unassigned comments separately.” This is a suggested instruction, not a prompt we have independently tested.
Review proposed labels against the source text. Merge labels only when they represent the same idea; shared vocabulary alone is insufficient. “More training” and “time to use existing training” should remain distinguishable. Keep an unassigned category visible so uncategorized material does not quietly disappear.
Build a theme map that preserves mixed feedback
Consider these six invented comments about a fictional training pilot:
- C01: “The replay helped me repeat the steps.”
- C02: “The replay was useful, but I had no protected practice time.”
- C03: “The replay helped; I wanted harder practice tasks.”
- C04: “The live walkthrough helped, but I had no protected practice time.”
- C05: “My shift had no protected practice time.”
- C06: “I could not read the subtitles; I need an accessible version.”
The following map is our interpretation of this fictional set. In a real analysis, retain supporting spans and reviewer decisions alongside the references.
| Proposed theme | Definition and supporting IDs | Qualification to preserve |
|---|---|---|
| Useful instruction | Explicit value in a replay or walkthrough: C01, C02, C03, C04 | Different formats helped; this does not establish approval of the whole programme. |
| Protected practice time | Explicit lack of time reserved for practice: C02, C04, C05 | Two comments also valued the instruction. |
| More advanced practice | Request for harder tasks: C03 | One comment; it does not establish demand across all learners. |
| Material accessibility | Reported difficulty using the subtitles: C06 | A separate concern, even though it appears once. |
There are nine theme assignments across six comments because some comments contain several ideas. Neither nine assignments nor six comments establishes how many employees hold each view. If the collection permits repeat submissions or contains several answers per person, counting unique people requires a separately justified respondent mapping. Do not reconstruct identities merely to obtain that count.
A defensible summary would say: “Four of the six comments describe useful instruction. Three mention a lack of protected practice time, including two that also value the instruction. One requests harder practice, and one reports difficulty reading subtitles.” All numbers describe this invented comment set. They are not company-wide survey results.
“Most employees support the programme, with minor requests for improvement” would overreach. It changes the denominator, expands the claim and dismisses the accessibility concern without examining it.
Separate frequency, significance and action
Count each comment once within a theme, however often it repeats the point. Keep overlap across themes explicit. If you report percentages, name the denominator and explain exclusions. Optional text responses do not establish the views of people who supplied no text; absence of a complaint is not evidence of satisfaction.
Frequency answers how often an idea appears in the analyzed collection. Significance asks why it matters to the decision. A single report that someone cannot use the materials may warrant investigation even when other comments are favorable. This does not prove the issue affects everyone, and the brief should not imply that it does.
Keep proposed actions in a separate section. In the fictional pilot, the finding is that three comments describe missing practice time. Allocating time in the next pilot is a proposed response. Whether scheduling caused the difficulty, and whether a particular change will solve it, remain questions to investigate.
Assign an owner to investigate the accessibility report through an appropriate confidential route. Avoid reproducing distinctive employee wording in a widely circulated brief when a careful paraphrase will convey the finding. Preserve the original evidence for authorized review rather than distributing it with every presentation.
Review from the source as well as from the summary
Check in both directions. For every sentence in the brief, inspect its supporting comments and qualifiers. Then return to the original collection and look for material that never reached the brief: disagreement, mixed comments, ambiguous wording and unassigned items. A summary can contain only supported sentences while still omitting something consequential.
For a small set, read every comment. For a larger collection, record which records received detailed review, how they were selected and what remains unreviewed. Include an independently selected sample as well as uncertain and less frequent categories. Sampling offers limited assurance; escalate consequential omissions rather than calling an unchecked remainder verified.
Keep a review record of changed labels, corrected counts and removed claims. Our guide to reviewing AI answers against evidence explains the broader distinction between checking fluent prose and checking support. Here, the additional question is whether the process preserved the range of supplied views.
Start with a bounded learning decision
Try this process on an approved, manageable feedback set before integrating it into routine reporting. Retain the source version, coding definitions, tool details, evidence map and final brief. Compare review effort, unsupported claims, missed themes and corrections with the team’s existing process. Measure useful findings and required rework, rather than speed to the first paragraph alone.
Use qualified employee-listening or research support when interpretation is consequential or contested. Route allegations and individual employment matters through established human processes; an aggregate AI summary cannot investigate them. For a training pilot, begin with one actionable question and publish a brief whose important statements can be traced back to the feedback. Keep uncertainty visible where the evidence cannot resolve it.