By Kirill Naumov · 2026-09-28

Retailers already have more customer comments than most teams can read in a week: reviews, surveys, service notes, returns reasons and conversations from different channels. AI can help sort this material into themes. The useful output is not a colourful sentiment chart; it is a better question for the team that can fix something.

Start with a decision, not a pile of text

Choose a practical decision: should the size guide change, does a delivery promise need attention, or are customers repeatedly asking staff for information that is missing from the product page? Define the channel, period, product group and decision owner before analysis. If the business cannot act on a theme, it may not belong in the first exercise.

Keep the original comments in view

A language model can group similar phrases and suggest a plain-language label, but its summary can flatten important differences. “The fit is small” may refer to one item, one market or one body type; “delivery was late” may describe a single exception or a repeated route problem. Keep links or IDs back to source comments, and let reviewers inspect examples and counterexamples before naming a theme.

A recent study comparing topic models and GPT on consumer reviews uses a two-stage human evaluation framework and highlights interpretability, reliability and managerial usefulness as validation questions—not simply whether the model produces plausible labels. The [study record](https://www.kci.go.kr/kciportal/ci/sereArticleSearch/ciSereArtiView.kci?sereArticleSearchBean.artiId=ART003325932) is a useful reminder to test categories with people who understand the business and customer context.

Separate frequency from importance

The loudest or most common theme is not automatically the best priority. A rare safety concern can matter more than a frequent request for another colour. Score themes using at least three lenses: how often they appear in the relevant group, the consequence for the customer and whether the team can make a change. Keep “not enough evidence yet” as a valid result.

Check language, channel and who is missing

Review samples from different channels and languages. A short rating prompt may encourage different comments from a post-purchase survey; customers who never complete the survey are absent altogether. Translate only when necessary, preserve the original wording and ask a fluent reviewer to check terms that carry local meaning. Do not treat an AI-generated sentiment score as a direct measure of customer emotion.

Protect customers and make the work accountable

Use only data the team is permitted to analyse, remove identifiers that are not needed and restrict access to raw comments. Set a retention period. Agree who approves a proposed change and who checks whether it helped. NIST’s [AI Risk Management Framework](https://airc.nist.gov/airmf-resources/airmf/5-sec-core/) connects context, measurement and risk management; for this workflow that means keeping a record of the source, reviewer, decision and follow-up—not just saving the model prompt.

Turn one theme into a small test

Suppose a cluster suggests that customers cannot tell whether a product works with an accessory they already own. Review the comments, check the product information, update one page or staff reference, then compare suitable measures such as repeat questions, assisted-service time and returns for the selected items. Do not claim the text analysis caused a change unless the comparison supports it.

That is the practical role of AI in retail feedback: make a large collection of voices easier to examine, while keeping the interpretation and business decision accountable to people. For more applications across retail, sales, learning and loyalty, see the [AI for Retail workshop](https://www.kirillnaumovai.com/workshops/ai-retail-november).

Explore the English AI for Retail workshop → · Start an AI diagnostic

Start a conversation →

AI in Retail Customer Feedback: Find the Decision Behind the Comment