Customer experience / Listening
Uses AI in Codex to interpret synthetic feedback, with source excerpts and human review.
Recorded Codex AI analysis · Human review
Explore the full project ↓Customer feedback is only useful when a team can trace a pattern back to the comments behind it.
Codex interprets each distinct comment in context, proposes themes and sentiment, and explains the judgment with a source excerpt. Code validates the output and links repeated text.
180 synthetic records, 17 AI-analyzed texts and 163 linked repetitions. All judgments await human approval; independent accuracy has not been measured.
Portfolio examples illustrate the workflow. They are not claims about real customer outcomes or independently measured production impact.
THE OPERATIONAL QUESTION
Customer feedback rarely arrives neatly organized. One comment can praise a product, report a delay and describe a defect. This workflow keeps those signals together, with evidence a person can check.
Three examples show how AI separates mixed feedback, identifies a sizing issue and flags ambiguity. Each interpretation is linked to the original comment.
FB-0001 / SYNTHETIC FEEDBACK
A costura soltou depois de usar uma vez.
AI-PROPOSED ANALYSIS
The comment describes a seam coming apart after one wear; it does not establish the cause.
“A costura soltou”
Human decision pending.
FB-0002 / SYNTHETIC FEEDBACK
Support answered quickly and solved my question.
AI-PROPOSED ANALYSIS
Both response speed and resolution are praised.
“answered quickly and solved my question”
Human decision pending.
FB-0003 / SYNTHETIC FEEDBACK
It was fine, I guess.
AI-PROPOSED ANALYSIS
Qualified approval gives no operational topic and is too vague for a confident sentiment decision.
“It was fine, I guess.”
Ambiguous sentiment and unspecified topic. Human decision pending.
FB-0004 / SYNTHETIC FEEDBACK
I could not understand how to start a return.
AI-PROPOSED ANALYSIS
The obstacle is understanding the return process, not a confirmed refund delay.
“could not understand how to start a return”
Human decision pending.
FB-0005 / SYNTHETIC FEEDBACK
The color and embroidery look even better in person.
AI-PROPOSED ANALYSIS
The customer compares appearance with expectations; this is not evidence of fabric durability.
“color and embroidery look even better in person”
Human decision pending.
FB-0006 / SYNTHETIC FEEDBACK
The fit is much smaller than the size chart suggested.
AI-PROPOSED ANALYSIS
Reported fit disagrees with the chart; measurements still need verification.
“much smaller than the size chart suggested”
Human decision pending.
FB-0007 / SYNTHETIC FEEDBACK
The design is beautiful, but the price feels high for the fabric.
AI-PROPOSED ANALYSIS
Praise for design coexists with concern about value; no material defect is established.
“design is beautiful, but the price feels high for the fabric”
Human decision pending.
FB-0008 / SYNTHETIC FEEDBACK
Love the style, but delivery was late and the sleeve tore; support was kind though.
AI-PROPOSED ANALYSIS
Positive style and support comments coexist with a late delivery and torn sleeve. Separate owners need to review the issues.
“Love the style, but delivery was late and the sleeve tore; support was kind though.”
Multiple operational issues. Human decision pending.
FB-0009 / SYNTHETIC FEEDBACK
Is this fabric lined and does it stretch?
AI-PROPOSED ANALYSIS
The customer requests lining and stretch facts; neither answer is supplied.
“Is this fabric lined and does it stretch?”
Human decision pending.
FB-0010 / SYNTHETIC FEEDBACK
The measurements matched the size chart exactly.
AI-PROPOSED ANALYSIS
The comment explicitly reports agreement with the chart.
“measurements matched the size chart exactly”
Human decision pending.
FB-0011 / SYNTHETIC FEEDBACK
The package arrived four days after the promised date.
AI-PROPOSED ANALYSIS
A delay is reported relative to a promise, but carrier fault cannot be inferred.
“four days after the promised date”
Human decision pending.
FB-0012 / SYNTHETIC FEEDBACK
Checkout kept refreshing and I had to enter my address twice.
AI-PROPOSED ANALYSIS
Repeated page refresh caused re-entry; no payment failure is stated.
“Checkout kept refreshing and I had to enter my address twice.”
Human decision pending.
FB-0018 / SYNTHETIC FEEDBACK
O pacote chegou quatro dias depois do prazo.
AI-PROPOSED ANALYSIS
The Portuguese comment reports arrival four days after the deadline.
“quatro dias depois do prazo”
Human decision pending.
FB-0020 / SYNTHETIC FEEDBACK
The stitching came loose after one wear.
AI-PROPOSED ANALYSIS
The customer describes a seam failure, not a sizing problem.
“stitching came loose after one wear”
Human decision pending.
FB-0035 / SYNTHETIC FEEDBACK
O tecido tem forro e elasticidade?
AI-PROPOSED ANALYSIS
The Portuguese question asks about lining and stretch, without providing those facts.
“O tecido tem forro e elasticidade?”
Human decision pending.
FB-0052 / SYNTHETIC FEEDBACK
O checkout recarregou e precisei digitar o endereço duas vezes.
AI-PROPOSED ANALYSIS
The Portuguese comment reports refresh and duplicate address entry.
“O checkout recarregou e precisei digitar o endereço duas vezes.”
Human decision pending.
FB-0069 / SYNTHETIC FEEDBACK
As medidas corresponderam exatamente à tabela.
AI-PROPOSED ANALYSIS
The Portuguese comment confirms the measurements matched the chart.
“medidas corresponderam exatamente à tabela”
Human decision pending.
These are observations from a small synthetic dataset, not findings about real customers.
One delivery comment also describes a torn sleeve and helpful support. Routing the entire comment to one team would hide part of the problem.
Next step: review delivery and product-quality evidence separately.
Sources: FB-0008, FB-0011, FB-0018English and Portuguese comments ask about lining and stretch. The AI identifies the information gap without inventing an answer.
Next step: verify specifications before updating product pages.
Sources: FB-0009, FB-0035Repeated text is analyzed once to avoid redundant work. In production, legitimate occurrences from different customers must still count toward volume. Multi-topic theme counts can overlap.
A custom integration could collect new feedback and prepare a review queue automatically. Codex would help build the code; a scheduled service would run it.
Proposed production design · not connected in this prototypeA daily job reads a review or support API. A webhook can trigger processing sooner. Stable source IDs prevent accidental duplicate imports.
Remove unnecessary personal data, then send the comment and versioned taxonomy to a model. Save its proposed labels, excerpts and explanation.
Code checks the schema and cited text. Failed requests retry safely; invalid or ambiguous results go to review instead of disappearing.
A person approves or corrects the analysis, assigns an owner and tracks the action. Later periods show whether the same issue keeps appearing.
Access to the source API, a scheduled backend, persistent storage and a model API or hosted local model. No Make or Zapier is required. Hosting and model usage may have costs; credentials stay on the server.
What exists today: Codex analysis, stored results, source-linked evidence and validation tests. Collection, scheduled inference, review actions and operational tracking are not implemented yet.
All 17 AI judgments await human approval. Automated tests verify source integrity, coverage, excerpts and counts; they do not measure interpretation accuracy. An independently labelled evaluation set is still needed.
LET'S PUT THAT CURIOSITY TO WORK