Analyze customer feedback and identify recurring themes

AI can help group comments, but a theme summary should still let you find the records behind it. In this exercise you will build a coding table, count distinct comments per theme, and propose questions to investigate.

Use a text-capable AI chat and a document or spreadsheet. No coding or upload is needed. All comments below are original fictional September 2026 feedback. Hands-on timing has not been measured.

1. Preserve the comments and IDs

These eight records are a teaching sample, not a representative survey. There is no customer-population denominator or evidence of sales impact. Count comments, not unique customers, because respondent identities are not supplied.

2. Agree a counting rule

Use themes Setup, Pricing, and Support. A comment can receive more than one theme. Count each ID at most once within a theme. Keep positive, negative, and mixed observations visible. A relevant positive comment still belongs to its theme; “mentions” do not mean “complaints.” C08 stays General because it does not identify a specific issue.

3. Ask for record-level coding first

Code these eight synthetic comments using themes Setup, Pricing, Support,
and General when no specific theme applies.
[Paste C01–C08 and the counting rules above.]
Return ID, assigned themes, evidence wording, and interpretation for each record.
Allow multiple themes, but count each record once per theme.
Then list theme IDs, distinct-comment counts, and counts divided by eight.
Separate positive and negative evidence. Do not infer population prevalence,
customer identities, churn, revenue impact, or causal explanations.
Propose two investigation actions, clearly labelled as proposals.

4. Compare with a reproducible example

Checked editorial coding:

The counts sum to ten and shares to 125%, because comments can have multiple themes. That is consistent with this counting rule. Setup has three concern records and one positive record, so it would be wrong to say four customers complained about setup.

5. Write a bounded conclusion

Illustrative editorial summary:

Setup is mentioned in four of eight comments, with three concerns and one positive experience. Support appears in three, including two positive experiences and one delayed reply. Pricing appears in two, raising distinct value and clarity questions. Proposed next steps are to inspect checklist visibility and missing instructions using C02/C07, and review pricing explanations against C03/C05. This sample does not establish how common these issues are across customers.

Verify, recover, and reuse

Match every coded row to the original comment. Recount unique IDs per theme. Check that overlapping categories and the denominator are explicit. Inspect mixed comments instead of forcing each into a single sentiment.

If the tool duplicates C01 within Support, deduplicate that theme’s ID list and recount. If it says “most customers,” replace the population claim with the exact sample count. If a code is debatable, retain the comment and record your coding decision rather than treating the model’s label as fact.

For practice, remove C05. The denominator becomes seven; Setup becomes 3/7, Pricing 1/7, and Support remains 3/7. Save the coding rules and version with the result. For real feedback, use approved data handling, preserve consent/context where relevant, and have another reviewer challenge consequential interpretations.

Source: original fictional comments C01–C08; theme assignments and output are editorial examples. No real customer research or named AI service execution is claimed.