Use ChatGPT to Turn Customer Feedback into an Evidence Tagged Action Table

Customer feedback is easy to summarize badly. A neat paragraph can hide which comments were common, which were unusual, and which action came from an actual customer statement.

This is an illustrative ChatGPT workflow with fictional feedback. It is a documentation based guide, not a captured ChatGPT execution and not a claim about customer demand.

What you will make

You will turn a small CSV into an action table with five columns:

Theme
Evidence rows
What customers actually said
Possible next action
Question for a human owner

The last two columns are deliberately separate. A comment can support an observation without proving that a particular product change is the right response.

What you need

ChatGPT with file uploads available on your account. OpenAI documents CSV and spreadsheet analysis as a supported file use case, subject to plan limits and account settings. See the File Uploads FAQ.
A CSV containing approved, low risk feedback. Use fictional or redacted content until your organisation has reviewed its data handling.
A list of allowed themes, so the model does not invent a new category for every row.

Create a file named feedback.csv with this complete fictional input:

feedback_id,customer_segment,submitted_at,comment
F-101,Small team,2026-09-03,"The weekly export is useful, but I cannot tell which rows changed since last week."
F-102,Small team,2026-09-04,"I spend too long checking whether an alert is still relevant before I share it."
F-103,Mid-market,2026-09-05,"The export arrives on time. We need a clearer owner field for each open item."
F-104,Mid-market,2026-09-06,"The dashboard has the numbers I need, but the labels are unfamiliar to new team members."
F-105,Small team,2026-09-08,"Please keep the export. It is easier to send around than a screenshot."
F-106,Mid-market,2026-09-09,"I could not find the date of the last update in the report."
F-107,Enterprise,2026-09-10,"Our review process requires a named person to confirm each exception."
F-108,Small team,2026-09-11,"The alert wording is clear once I open the detail page."

These rows are practice material. They are not measured results.

1. Set the analysis boundary

Open a new ChatGPT conversation, upload feedback.csv, and paste this prompt:

Review feedback.csv as a small qualitative feedback sample.

Allowed themes only:
1. Change visibility
2. Alert review
3. Ownership
4. Report clarity
5. Positive signal
6. Other

Create a table with these columns:
- Theme
- Evidence IDs
- Exact evidence quotes
- Observation
- Possible next action
- Question for a human owner

Rules:
- Use only the rows in feedback.csv.
- Keep every feedback ID attached to its quote.
- Quote comments exactly. Do not combine two comments into one quote.
- Assign one primary theme per row. If a row could fit two themes, choose the closest allowed theme and explain the choice in the observation.
- Do not count a theme as a market finding or customer demand. Call it a sample observation.
- Do not infer urgency, revenue impact, sentiment beyond the words used, or a requested feature that is not stated.
- Put any suggested product or process change in Possible next action, not in Observation.
- If the feedback does not answer a question, write NOT STATED.
- Add a final section called Review checks.

The prompt uses a fixed label set and keeps observations separate from actions. That separation is the useful part. The table is not a product roadmap.

2. Check the evidence links

For each output row, compare the evidence IDs and quotes with the CSV. A practical review table should look like this for the fictional input:

Theme	Evidence IDs	What the rows support	Human check
Change visibility	F-101	One respondent cannot see which rows changed	Check whether this is a repeated issue in a larger sample
Alert review	F-102	One respondent spends time checking alert relevance	Ask whether the review step is understood before changing the alert
Ownership	F-103, F-107	Two respondents mention an owner or named person	Confirm the required owner field and process
Report clarity	F-104, F-106	Labels and update dates are unclear to these respondents	Check the report terminology and last update field
Positive signal	F-105, F-108	Two respondents describe parts they want to keep	Preserve these parts while reviewing changes

This is an editor written illustrative target. It is not a model output. Compare it with the uploaded rows before using it.

3. Run the acceptance checks

The workflow passes when:

Every evidence ID exists in the CSV.
Every quote matches the source row.
F-105 and F-108 are not rewritten as complaints.
A possible action is not presented as a customer request unless the row says that directly.
The table does not claim that eight rows represent all customers.
Missing urgency, business impact, and frequency remain NOT STATED.
A human owner has a question to answer before anyone changes the roadmap.

Troubleshooting

ChatGPT invents a theme. Repeat the allowed theme list and ask it to classify the row as Other when none fits.

It merges evidence rows. Ask for one table row per primary theme and require a comma separated list of original IDs. Then check the quotes manually.

It turns a suggestion into a finding. Add this rule: “Observation must describe only what the rows say. Possible next action may contain a hypothesis, clearly labeled as a hypothesis.”

It produces a polished but vague summary. Ask for the table first, then the summary. A summary without evidence IDs is difficult to audit.

Limits and safer use

A small feedback file can reveal personal information or confidential product details. Use approved or fictional input, remove unnecessary identifiers, and follow your organisation’s data policy. File upload limits and account settings can change, so check the current OpenAI File Uploads FAQ.

TRY: eight to twenty approved, low risk feedback rows and a fixed theme list.

SKIP: regulated complaints, sensitive personal data, or roadmap decisions that require a representative research sample.

USE: the evidence table as a review handoff. Do not use it as an automatic prioritisation or product decision.

For practical AI for work tasks with an honest TRY, SKIP, or USE decision, get The Weekly Verdict.