AIWorkDex
Ecommerce Practice

Turn return feedback into one prioritised fix

Create a reviewable cause table by product, then choose one change with traceable customer evidence and a measurement plan.

Preparation
A focused analysis session, followed by a later measurement review.
Difficulty
Moderate
Task
Find product return causes
Back to the Task: Find product return causes

Before you start

Prepare the data, boundary, and owner.

  • Choose one comparable date range and the products or category you want to investigate.
  • Export physical return line items, sales quantities, reason codes, and available free-text comments.
  • Gather relevant support excerpts or reviews only when your privacy policy and access controls permit their use.
  • Use an organisation-approved ChatGPT workspace and confirm its current data controls before upload.
  • Assign a product or operations owner who can inspect evidence and approve the chosen change.

Actionable steps

Build a narrow route, then test it.

Each step names its input, output, supporting Tools, and inspectable evidence.

  1. Define the return cohort

    Choose the period, market, products, and return state. Separate physical returned quantity from cancellations, edits, and refunds without returned goods.

    Why it matters: A mixed cohort can assign costs or causes to the wrong event.

    Input
    Shopify return records, order records, and the analysis question.
    Output
    A written cohort definition and the exact filters used.
    Tools in this stepShopify

    Evidence for this step

  2. Build the minimum useful export

    Create one row per returned line item. Keep product, variant, quantity, reason, comment, date, and a non-identifying evidence key. Remove direct customer identifiers.

    Why it matters: The row needs enough context for diagnosis and review. Extra customer data creates avoidable privacy risk.

    Input
    The filtered return cohort and approved related feedback.
    Output
    A cleaned file with one returned line item per row and no direct customer identifiers.
    Tools in this stepShopify

    Evidence for this step

  3. Write a cause taxonomy

    Start from the store’s structured reasons. Add category-specific causes found in the comments. Keep Unknown and Other as valid outcomes.

    Why it matters: Generic labels can hide product-specific problems. Forced labels can create a pattern that the evidence does not support.

    Input
    Structured reason codes and a human-read sample of comments from each product group.
    Output
    A cause list with definitions, examples, exclusions, Unknown, and Other.
    Tools in this stepShopifyChatGPT data analysis

    Evidence for this step

  4. Classify each row and keep its evidence

    Ask ChatGPT to apply only the approved taxonomy. Require cause, confidence, exact supporting excerpt, evidence key, and a short rationale for every row.

    Why it matters: A summary without row-level evidence is hard to audit and easy to over-trust.

    Input
    The cleaned export, taxonomy, and a fixed classification instruction.
    Output
    A labelled file that retains each original comment, evidence key, confidence flag, and rationale.
    Tools in this stepChatGPT data analysis

    Evidence for this step

  5. Audit the labels

    Review every low-confidence, Other, and Unknown row. Check a varied sample of the remaining labels. Correct the file and refine unclear definitions.

    Why it matters: Language models can overgeneralise. Human review tests whether the taxonomy fits this store and product mix.

    Input
    The labelled file, original comments, and the taxonomy.
    Output
    A corrected file, audit notes, and a recorded review sample.
    Tools in this stepChatGPT data analysis

    Evidence for this step

  6. Quantify causes by product

    Group reviewed labels by product, variant, category, and cause. Show returned units, sold units, return rate, and available return value. Keep Unknown visible.

    Why it matters: Counts favour high-volume products. A sales denominator shows whether a cause is unusually frequent for that product.

    Input
    The audited labels and sold quantities for the same cohort.
    Output
    A cause table with counts, denominators, rates, values, and links to representative evidence.
    Tools in this stepShopifyChatGPT data analysis

    Evidence for this step

  7. Choose one fix

    Select one product-cause pair with repeated reviewed evidence and a change you control. Write the proposed fix and the evidence against competing explanations.

    Why it matters: One bounded intervention makes ownership clear and gives the team a result it can inspect.

    Input
    The cause table, representative comments, product page, support context, and operational constraints.
    Output
    A one-page decision with the chosen cause, evidence, owner, change, risks, and rejected alternatives.
    Tools in this stepShopify

    Evidence for this step

  8. Measure the change

    Record the launch date and affected products. Compare the same cause and denominator over a suitable later window. Note promotions, price, stock, and policy changes.

    Why it matters: A return-rate movement can have several causes. A recorded comparison reduces false confidence in the chosen fix.

    Input
    The decision record, launch date, baseline cohort, and later return and sales data.
    Output
    A follow-up table and a decision to keep, revise, expand, or reverse the change.
    Tools in this stepShopifyChatGPT data analysis

    Evidence for this step

Success checks

Check facts and handovers before expanding.

  • The cohort contains physical returns and uses the same filters for return and sales data.
  • Every AI label retains the original evidence key and supporting customer language.
  • A person reviewed all low-confidence, Other, and Unknown rows plus a varied sample of other labels.
  • The ranked table shows returned units and a valid sold-units denominator for the same period and product scope.
  • The selected fix names one product-cause pair, representative evidence, an owner, a launch date, and a follow-up measure.
  • The decision record states the limits of the evidence and plausible competing explanations.

Failure modes

Stop and repair these conditions.

  • Refunds and physical returns are mixed. Rebuild the cohort with the physical-return fields.
  • Free-text comments contain names, emails, addresses, or order numbers. Remove them before upload.
  • The model forces every comment into a neat label. Restore Unknown and Other, then review ambiguous rows.
  • The top cause comes from one high-volume product. Add sold units and compare rates by product.
  • The summary has no customer excerpts. Return to row-level output and attach representative evidence.
  • A category-specific issue sits inside a generic code. Refine the taxonomy and reclassify the affected rows.
  • Several fixes launch together. Record the overlap or narrow the intervention before claiming a result.
  • The later rate changes during a promotion or policy shift. Record the confounder and extend or repeat the comparison.

Relevant Tools

Use each product for a defined part of the method.

Sources

Inspect the material behind the method.

Product capabilities come from official documentation. Scope, testing, monitoring, and human oversight use the cited risk guidance.