Someone on a tactical gear maker’s team asked me how AI could analyze customer interactions for product development and quality control. It came up at an AI lunch-and-learn I gave for their company.
My answer was about patterns. “AI is great at finding patterns, and it doesn’t get tired.” A customer on the phone says, “I wish this knife had this extra feature.” The rep hears it once, and it’s gone. If five reps each hear it once, nobody catches it. AI does.
It was a good answer. I had never tried it on my own store.
The most-returned product in my Amazon store is a mat. You lay it on your car roof under a cargo bag so the bag doesn’t scratch the paint. Across 2024 and 2025, customers sent back 18% of the ones we sold. The store as a whole runs about 4%.
Amazon records why each return happened, in a field called the reason code. For the mat, the answer looks settled. Most of those returns are coded “unwanted item” or “ordered wrong item.” The customer changed their mind. Nothing to fix.
Every seller has a product like this, and a return coded “unwanted” reads like weather: annoying, and nobody’s fault. Before that question, I would have left it there.
So I tried my own answer. Underneath the reason code, Amazon also keeps whatever the customer typed in the comment box. The mat had 494 of them. Nobody had ever read them all, because nobody has an afternoon to spend on 494 return comments about a mat.
I had my AI read every one.
It comes back with one sentence, over and over.
“The roof bag I ordered came with a protective mat.”
“Didn’t know roof carrier came with protective mat.”
“I did not realize the roof cargo bag I ordered already came with a protective mat. This has not been opened.”
More than a hundred of them. About one in five of every comment anyone bothered to write. Two customers even described how it happened. One wrote, “Inaccurate recommendation by Amazon.” Another: “We thought we bought all things bought together. The original kit already came with it.”
Almost every one of those returns is coded “unwanted item” or “ordered wrong item.”
The reason code says the customer made a mistake. The comments say the mistake was in how the mat was being sold. Shoppers were buying it alongside a cargo bag that already had one in the box, and they didn’t find out until the box arrived. When I pulled the returns, nothing on my listing told shoppers to check their bag first. The title said the opposite: put this mat “under any rooftop cargo bag.”
A reason code is a dropdown. The comment is the customer.
A reason code is a short list a customer picks from on the way out the door. “I didn’t need it because my bag came with one” has no button, so it becomes “unwanted item.” Any report built on reason codes files it as noise.
The signal is in the words. And the words are exactly what nobody reads, because one person can’t read 494 comments, and one rep who hears a complaint once has nothing to compare it to. Every complaint gets handled once and thrown away. The pattern only exists across hundreds of them.
That is the job AI is built for. Not deciding what to do. Reading everything, without getting tired, and telling you what keeps coming up.
The same comments, read for the product
The duplicate mat was the loudest pattern. It wasn’t the only one, and the others answer different questions.
Product development. About two dozen customers said the mat was too small. One wrote, “The product description says it is universal but it’s not.” A few said they expected it to be thicker. That is a spec for the next version, written by the people who would buy it.
Quality control. Only a handful described an actual defect: it slid around, it smelled, it left marks on the paint. That small number is a finding too. The product mostly works. The sale is where it goes wrong.
Noise you can set aside. About two dozen came back because the cargo bag they bought it for went back. Those say nothing about the mat. Knowing that is worth something on its own, because it stops you from fixing the wrong thing.
None of this needed new data. It was all sitting in a comment box.
Three questions to ask your own complaints
Point AI at a quarter’s worth of returns, reviews, support tickets or call notes, and ask:
- What are people complaining about that we haven’t fixed? Not the top complaint. Everybody already knows the top complaint. Look for the one at number six that is quietly growing.
- Which complaints cluster to one product, one batch, or one vendor? That is quality control. A complaint that traces back to a single production run is a finding you will never get by handling tickets one at a time.
- What do people ask for that we don’t sell? That is your product roadmap, written by customers, for free.
The same move works on sales calls. Put a call that closed next to one that didn’t, and ask the AI what the first rep did differently.
One honest caveat. This only works if the complaints live somewhere the AI can read them. Amazon return comments already sit in one report. If your complaints are spread across an inbox, phone notes and someone’s memory, step one isn’t AI. It’s getting them into one place.
Read the comment before you believe the code
I used to treat a return coded “unwanted” as the customer’s problem. I don’t anymore. I read what they wrote before I believe the code.
The fix the comments point to is not complicated. It’s one sentence on the listing: check whether your bag already includes a mat. No dashboard would have told me to write it.
Every one of those 494 customers took the time to tell us what went wrong.
For years, the only thing listening was a dropdown.
