I designed the controlled taxonomy, evidence fields, and review standard for this workflow. Each classification keeps the original customer comment, a confidence signal, and a clear route for human review when the evidence is mixed or uncertain.
“The dress looks great, but delivery took longer than expected.”
Delivery“The size guide did not match the product I received.”
Review“The product photo was helpful and the checkout was easy.”
PositiveOne visible source per signal. The original wording stays attached to each classification.
Each record keeps its source, receives a controlled classification, and makes low-confidence or mixed feedback visible for a person to check.
Prototype boundary: the case uses 180 synthetic feedback records across a 44-product test catalog. It does not use real customer feedback or make live product decisions.
A team can read comments individually, but that does not scale into a consistent operating view. The risk is a vague summary that hides uncertainty, loses the source, or treats mixed feedback as a clean signal.
Catalog, CX, and operations teams need a shared view of repeat issues and a route for ambiguous feedback.
Comments are unstructured and can contain more than one issue. The LLM maps language into a fixed taxonomy, rather than inventing categories.
No auto-publishing, product change, customer reply, or claim that the feedback represents a real customer population.
Retain the comment, rating, channel, date, and product reference.
Assign theme, sentiment, priority, actionability, and confidence.
Send low-confidence or multi-issue records to a person.
Build counts and insights that link back to visible records.
This is read-only evidence from the completed prototype. “View full table” displays every row in the selected workbook view.
Open the complete workbook →Time to triage feedback, time to identify recurring issues, and adoption of recommended actions by the owning team
Agreement with human reviewers, routing precision, unresolved ambiguity rate, and changes in taxonomy coverage
Taxonomy owner, sampled quality review, versioned prompts, privacy controls, and an escalation path for unsafe or unclear output
The monitoring plan follows the NIST AI RMF Measure guidance ↗: compare with a human baseline, document test sets, and track error handling.