As of 13 August 2026, AI can find defect patterns in your customer complaints.

What to actually do

1. Hand it to a person

   The route this page recommends

   A person who owns the outcome does this end to end, worth it when the failure is dear.

2. Use a tool built for this

   Second choice

3. Do it yourself

   The distant third

   A chat interface, chat-fluent skill, and roughly 30 minutes until you can act on the result.

   How to actually do it

   1. Export the complaints from your CRM, helpdesk, email system, or spreadsheet and keep the complaint text together with any available date, product, batch, supplier, site, channel, and outcome fields.
   2. Remove names, addresses, phone numbers, email addresses, order numbers, and other identifying details before copying the data into a new document.
   3. Open a chatbot and paste the prompt, then add your anonymised complaint export, the definitions used by your quality team, and any known exclusions or context.
   4. Ask the model to produce the themed report and retain the complaint identifiers or row references needed to trace each quoted example back to the source file.
   5. Compare every reported count, percentage, quotation, duplicate flag, and field breakdown against the source export, correcting the report where the model has merged or split complaints incorrectly.
   6. Give the report and source examples to a quality or operations colleague, who should test the highest-priority themes against inspection records, returns, batch records, supplier information, or a physical sample.
   7. Record confirmed defects separately from unverified patterns, then send only the confirmed actions into your normal quality-control or corrective-action process.

   Prompt

   Analyse the customer complaints below for possible product, service, packaging, delivery, or process defects. First remove or ignore names, addresses, order numbers, phone numbers, email addresses, and other identifying details. Do not invent facts, counts, causes, or defect categories.
   
   For each complaint, identify only what is explicitly stated: date, product or service, channel, location if supplied, symptom, suspected defect, and outcome. Then:
   1. Group complaints into clear recurring themes, using plain labels.
   2. Give the number of complaints in each theme and the proportion of the supplied complaints, showing your working from the supplied data.
   3. Provide two or three representative complaint excerpts for each theme, keeping them anonymous and changing no wording inside the excerpts.
   4. Separate direct evidence from possible explanations. Do not claim that a theme is a confirmed defect unless the data proves that.
   5. Flag duplicate complaints, vague entries, conflicting information, and themes based on very little evidence.
   6. Look for patterns by product, batch, supplier, site, channel, date, or delivery method only when those fields are present.
   7. End with a prioritised list of checks for a quality or operations colleague. Each check must name the evidence needed, such as inspection records, batch records, returns, supplier data, or a sample of affected items.
   
   Return the result under these headings: Data quality, Recurring themes, Breakdown by available fields, Evidence and excerpts, Possible explanations, Missing or conflicting information, and Recommended checks. Treat the output as a hypothesis report for human checking, not as proof of a defect.
   
   Complaint data:
   [PASTE ANONYMISED COMPLAINT EXPORT HERE]
   
   Available product, process, defect, batch, supplier, and channel definitions:
   [PASTE DEFINITIONS HERE OR WRITE NONE]
   
   Known exclusions or context:
   [PASTE CONTEXT HERE OR WRITE NONE]

   Open it prefilled in ChatGPT or Claude, or copy it into Gemini, which takes no prefill link.