Reading 1,182 customer conversations in a day

The monthly Voice of Customer analysis took a week, so it was the first thing to go when the month got busy. Some analysis was not slow, it simply never happened.

I built a pipeline for it. A 6.8 MB raw chat export goes through a rule-based cleaning pass, then a semantic filter that drops conversations with no real human intent, then a thematic synthesis that comes out with quantified pain points. What arrives is a document you can act on the same day.

The output is not the point on its own. That analysis is what surfaced the two main problems of the year, a slow site and a pricing page nobody understood, and both got fixed. The initiatives were not picked in a room. They came out of the reading.

Speed is the smaller half of the story. The larger half is that this analysis was not being done at all, because it cost more hours than it was worth. Cheaper analysis changed what the team could afford to look at, and what we saw changed what we built.