During my master's research in Quantitative Methods (University of Buenos Aires) I analyzed more than 200,000 customer reviews in Spanish using language models and explainable artificial intelligence techniques. The question was academic; the answers, remarkably practical for any SME that sells products or services.
The experiment
I took a large public corpus of reviews in Spanish and trained deep learning models to separate satisfied customers from dissatisfied ones. I then applied topic modeling to the tens of thousands of negative reviews to discover, without predefined categories, what people actually complain about when a purchase goes wrong.
What bothers people most is not the product
The finding that surprised me most: the most prevalent dissatisfaction factor in the corpus was not product quality, but logistics. Orders that never arrived or arrived late account for around 12% of all the complaints analyzed — more than any other single reason.
After logistics, the most recurrent general factors draw a map that anyone who runs a business will find familiar:
- Frustrated expectations: the product was not what the photo or the description promised.
- Value for money: not "it's expensive," but "it's not worth what I paid."
- Durability: it worked fine… for the first two weeks.
- Customer service and refunds: the problem was not the failure, but that nobody responded after the failure.
- Operational details: sizes, colors different from what was ordered, poor packaging.
Three lessons for an SME
1. Your reputation is decided at delivery, not just in the product. You can make something excellent and lose the customer in the last mile. If your sales, your warehouse and your invoicing do not talk to each other, that last mile is invisible to you until the complaint arrives.
2. Dissatisfaction almost never announces itself through official channels. Most of these customers did not open a ticket: they left a review and never came back. Measuring only formal claims is looking at the tip of the iceberg.
3. The factors repeat themselves — and what repeats can be measured. Size, delivery time, refund, expectation: each of these factors can become an indicator tracked every week on a dashboard. What I discovered with language models at the scale of 200,000 reviews, an SME can monitor with the data it already has in its ERP.
From research to your operations
At Inteleqta we apply exactly this logic: connecting the data from the operation —orders, deliveries, incidents, returns— and turning it into indicators that management reviews every week. You do not need an AI lab to get started: you need your systems to talk to each other, and a handful of well-chosen metrics.
Ricardo Borja is the founder of Inteleqta and a consultant in ERP and Business Intelligence. This post summarizes findings from his master's research in Quantitative Methods (UBA) on the analysis of Spanish-language reviews with explainable deep learning.