How to Organize Customer Feedback With AI: Labels and Counts

To organize customer feedback with AI, keep each original comment attached to its labels and count clearly defined units. A short theme list is useful only if you can trace it back to what people actually said.

This walkthrough uses six original, fictional comments about a software application. The classifications are editorial examples, not results from a live AI test or evidence about a real product. The exercise shows how to preserve mixed opinions, distinguish different problems, and avoid inflating counts.

1. Keep an identifiable source for every comment

Give each comment a stable identifier before asking AI to group it. Keep the original wording in a separate field so that an edited summary cannot replace your evidence. For real feedback, remove unnecessary personal details from the working copy and use an approved tool for the material you are analyzing.

F1: I could not find the export option.
F2: Export failed twice when I tried to save my report.
F3: Search is helpful, but the export option is hard to find.
F4: Please add a dark mode.
F5: Search is slow with a large project.
F6: Looks good.

These are six comments. They do not establish six distinct customers, because the exercise provides no author information. F2 describes two failed attempts in one comment; it is not two separate feedback records.

GOV.UK’s guide to analyzing research sessions recommends recording observations separately from interpretation and grouping related observations before developing findings. That distinction is useful here: retain the statement first, then add your classification.

2. Define labels before counting them

Use a small set of labels with clear boundaries. For this example, the working definitions are:

• Export discoverability: difficulty locating the export option.
• Export reliability: a reported failure while exporting.
• Search experience: an opinion or reported behavior concerning search.
• Feature request: an explicit request for a new capability.
• Needs clarification: too little detail to assign a specific product theme. Keep this outside the theme totals.

Discoverability and reliability concern the same feature but describe different problems. Combining them immediately into “export is broken” would erase the difference between finding a control and completing an operation.

Allow more than one theme per comment. Also record the type of statement for each theme: praise, reported problem, or request. A single positive-or-negative label for the whole of F3 would lose part of its meaning.

3. Ask for labels with supporting excerpts

Organize the feedback below using the supplied label definitions.

Keep every source ID and the original comment unchanged.
For each comment, return:
- Theme label or labels
- The exact excerpt supporting each label
- Statement type for each theme: praise, reported problem, or request
- Any ambiguity requiring review

Allow multiple themes in one comment.
Do not infer root causes, customer identities, or severity.
Do not treat repeated attempts as separate comments.
If a comment is too vague for a specific theme, mark Needs clarification.

Count distinct comment IDs within each theme.
List the supporting IDs beside every count.
Also report the total number of comments and the number needing clarification.
Explain that theme counts can overlap.

LABEL DEFINITIONS:
[Paste the definitions.]

FEEDBACK:
[Paste the comments with their IDs.]

Ask for the row-level classifications before accepting a summary. If a label seems wrong, you can inspect the quoted wording and correct that record. For more ways to specify output requirements, see our practical AI prompt examples.

4. Review the six classifications

F1 — Export discoverability; reported problem. The supporting phrase is “could not find the export option.” It describes locating the option, not a failure after selecting it.

F2 — Export reliability; reported problem. “Export failed twice” supports the label. Preserve the two attempts in the evidence, while counting F2 once in the theme.

F3 — Search experience, praise; Export discoverability, reported problem. “Search is helpful” and “the export option is hard to find” support different labels. Keep both; the second clause does not cancel the first.

F4 — Feature request; request. “Please add a dark mode” requests a capability. It does not prove that the current display causes eye strain or that most users want this feature.

F5 — Search experience; reported problem. Preserve the qualifier “with a large project.” A summary saying only “search is slow” would broaden the complaint. The comment does not identify the technical cause.

F6 — Needs clarification. “Looks good” expresses general approval but does not identify what the author liked. Keep that positive wording without assigning it to search, export, or another specific feature.

5. Reconcile the counts with the original IDs

The reviewed theme counts for this sample are:

• Export discoverability: 2 comments — F1 and F3.
• Export reliability: 1 comment — F2.
• Search experience: 2 comments — F3 and F5. One praises search; one reports slowness.
• Feature request: 1 comment — F4.
• Needs clarification: 1 comment — F6, reported separately.

There are six source comments, five with specific themes, and one needing clarification. The five themed comments produce six theme assignments because F3 belongs to two themes. Do not add the theme totals and call the result a count of unique customers.

If you combine the two export themes for a broader view, their distinct source IDs are F1, F2, and F3: three comments mention an export issue. Report that as three of the six comments in this sample, not “half of customers have export problems.” The sample does not establish customer prevalence.

For larger batches, reconcile every source ID against the classification output. Check missing IDs, accidental repeats, and labels without a supporting excerpt. Similar wording alone is not sufficient reason to delete a comment: two people can report the same issue independently.

6. Turn themes into questions you can investigate

The label counts help locate evidence, but they are not a priority score. The number of mentions does not by itself establish severity, business impact, or how broadly an issue occurs.

For this sample, possible follow-up checks are to observe whether people can locate export, investigate the circumstances of the reported export failure, and examine search behavior with large projects. These are proposed investigations, not confirmed diagnoses. The dark-mode request can be recorded without turning one request into a product commitment.

Keep the favorable search comment in the report as well. A useful summary can say that the sample contains both praise for search and a context-specific slowness report. It should not force contradictory experiences into a single verdict.

7. Save the definitions with the result

Save the source comments, label definitions, reviewed assignments, and unresolved questions together. If you change a definition later, revisit affected comments before comparing counts across batches. A change in labeling can look like a change in customer experience even when the underlying comments have not changed.

The final output should let another reader move from a theme to its count, from the count to source IDs, and from each ID to the original wording. That trail makes an AI-assisted feedback report easier to question, correct, and use.

Copy a worksheet structure for your own feedback

Use one row per comment–theme pair when a comment can have more than one theme. Keep the source comments in a separate sheet keyed by source ID. The CSV below contains the reviewed classifications from this fictional example; it is not customer data.

source_id,theme,statement_type,supporting_excerpt,review_status
F1,Export discoverability,reported problem,could not find the export option,reviewed
F2,Export reliability,reported problem,Export failed twice,reviewed
F3,Search experience,praise,Search is helpful,reviewed
F3,Export discoverability,reported problem,export option is hard to find,reviewed
F4,Feature request,request,Please add a dark mode,reviewed
F5,Search experience,reported problem,Search is slow with a large project,reviewed
F6,Needs clarification,general praise,Looks good,needs clarification

Copy the block into a plain-text file with a .csv extension, or import it into your spreadsheet as comma-separated data. When adding your own feedback, use a spreadsheet’s CSV export so commas, quotes, and line breaks inside comments are escaped correctly. Keep identifiers as text if leading zeros matter.

Check the unit before counting: this worksheet has seven rows but only six distinct source IDs. F3 has two rows because it contains two themes. Excluding the clarification row leaves six theme assignments across five comments. Within each theme, count distinct source IDs; for the overall total, count distinct source IDs across the whole source sheet.

Run three checks before sharing the summary

Coverage: compare the input IDs with the classified IDs in both directions. Every input should be represented, and every output ID should exist in the input. Repeating F3 for a different theme is valid; repeating the same source–theme pair by accident can inflate a simple row count.

Evidence: read the excerpt beside each classification. An excerpt must support the assigned theme without losing qualifications. For F5, “with a large project” matters; for F3, preserve both praise and the reported difficulty.

Summary: check the reported totals against the worksheet. A valid summary of this example is: “Six comments were reviewed. Three mention an export issue; two mention search, with one positive and one negative observation. One requests dark mode, and one is too vague for a specific theme. Theme groups overlap.” These are counts within the sample, not estimates of all customers.

If a comment needs a direct response, use our guide to drafting and checking blog comment replies. Keep answering the individual reader separate from counting patterns across the dataset.

Collect evidence for the next step

If a theme needs clarification, draft neutral survey questions with checked branches rather than treating a few comments as proof of prevalence. If a comment describes a reproducible failure, turn the available evidence into a bug report and mark missing details instead of inventing reproduction steps.

For another task, browse the AI guides by workflow.