How to Use AI to Organize Customer Feedback Into Useful Themes

Begin with a question you can act on, such as which instructions people find unclear or what they mention about contacting your business. Asking AI to tell you everything about your customers is much less focused.

Sorting comments by theme means grouping them by subject. Positive and negative comments can belong to the same subject, so keep the topic separate from the opinion expressed about it.

GOV.UK's guidance on analysing user research distinguishes observations from interpretation and recommends grouping related observations into themes. For this small business exercise, the useful lesson is to retain what people said while examining possible patterns.

• GOV.UK: Analyse a research session

Use a small fictional set first, like the example below. You can learn how the task works without sharing customer records. For real work, use only information your business is permitted to share in the chosen tool.

Keep the original feedback in its existing location and work from a separate, permitted copy. Give each comment a simple ID, such as F1 or F2, so you can trace a label back to its source.

Record the question people were answering and the period or channel the comments came from. A complaint about finding an address means something different if the question concerned collection instructions rather than delivery.

Do not upload names, contact details, order numbers or confidential information just to try the exercise. Removing a name does not necessarily remove every identifying detail. If the information is sensitive or its permitted use is unclear, use fictional material instead.

Imagine a small business that lets people order online and collect their purchase. These six comments are invented solely for this exercise. They are not Latitudes client feedback, real customer quotations or results from testing a particular AI tool.

The practice question is: Which parts of the experience might need clearer information?

• F1: I could not find the collection hours.
• F2: Placing my order online was straightforward.
• F3: The collection instructions did not tell me which entrance to use.
• F4: My email question received a quick reply.
• F5: I waited three days for a reply through the contact form.
• F6: I was unsure what I needed to bring when collecting my order.

Ask for suggested topics with supporting IDs, then review the result. A useful illustrative grouping of this fictional set would look like this:

• Collection information: F1, F3 and F6. The comments concern hours, location and what to bring.
• Response timing: F4 and F5. One describes a quick email reply; the other describes a wait through the contact form. Preserve both experiences and their different channels.
• Online ordering: F2. This is positive feedback about placing an order, rather than a problem with collection instructions.

Read the source comments for each label. Does the label describe the actual subject? Has the summary turned a specific concern into a broader claim? Has a small or positive observation disappeared?

In the example, a label such as poor customer service would blur several different issues. Unclear collection instructions do not establish that staff were unhelpful, and two different response experiences do not establish that replies are always slow.

SurveyMonkey's November 3, 2026 guide to AI analysis of open-ended responses also stresses reviewing themes against the underlying comments and checking for distinct viewpoints absorbed into larger groups. That guidance supports treating the grouping as a draft.

If a comment genuinely discusses two subjects, you can assign two labels. Keep a note of that choice so you do not later mistake the number of labels for the number of comments.

• SurveyMonkey: Reviewing AI-generated themes in open-ended responses
• Check an AI answer before using it for work

Give the AI a narrow question and ask for evidence beside each proposed theme. The prompt below is a starting point for practice, not a guarantee that the output will be correct.

Use the fictional comments above, or another permitted practice set. Inspect the suggested labels yourself before using the summary.

Count from the corrected list yourself. In our fictional example, three of the six comments concern collection information. That describes this exercise only. It is not a satisfaction score, a percentage of all customers or evidence that a proposed change will work.

Use the themes to identify a question or a small improvement to investigate. The collection comments suggest checking whether the instructions clearly state the hours, entrance and required items. The response comments suggest looking separately at email and contact-form handling.

Keep a simple note with the theme, supporting IDs, your interpretation and the next question. Separating these fields makes it easier to explain why you are considering a change.

For real feedback, consider how it was collected. A few comments from one channel or period may not describe everyone using your business. Read serious or unusual concerns individually, even if they appear only once.

After making an appropriate change, review new feedback about the same question. Compare the actual comments and their context rather than assuming a cleaner AI summary means the customer experience improved.

Practising with a small example helps you see which parts of the task you understand and where you need help. You might be comfortable asking AI for a list but less sure about choosing a useful question or judging whether a label fits.

Latitudes Training, Coaching & Consulting, led by Marcela Arenas, teaches practical AI skills connected to everyday business tasks. If you would like guidance, describe what you want to learn and use a non-sensitive example to explain where you are getting stuck.

The learning goal is to understand the grouping well enough to make your own decisions. You remain responsible for what you conclude from the feedback and what you change in your business.

• Explore practical AI training with Latitudes
• Find a manageable first task for learning AI