Using a language model to read what customers write, at any volume

The situation

Customers tell a retailer a great deal in their own words: in reviews, questions, support messages and survey answers. It is some of the most honest information the business has, and almost none of it is ever read in full.

The problem

Earlier methods for sorting comments worked well for discovering what people talk about, but they had limits. Short and messy comments were hard to place, and changing the list of themes meant starting over. The business also wanted to track a fixed set of topics over time, and compare like with like from one month to the next.

What we did

Newer language models can sort text into themes you define without being trained on your data first. So the business wrote down the themes it cared about, in plain words, with a short description of each.

The model then assigned every comment to those themes. Before anything went live, people who knew the products checked a sample against the model's choices, and the instructions were tightened wherever the two disagreed. We also ran the same comments twice to see how much the labels moved, because consistency matters when you are tracking change.

Results went into a simple report. Teams could pick a product range, see which themes were rising or falling and read real comments behind each one.

The outcome

The whole stream of customer comments could be sorted into a stable set of themes that anyone in the business could understand and filter, with a person's check built into the process. Changing a theme became a matter of editing a sentence, not rebuilding a model.

What it meant for the business

Teams stopped sampling and started seeing everything. The model did the reading and the people did what only people can: deciding what to fix and checking that the labels made sense.

Could this work for you?

This suits any business with more text than it can read: reviews, tickets, emails, survey answers or call notes. We also help weigh up the trade-offs, including cost at high volume and where the text is sent, which can be the deciding factor for sensitive comments.