An energy company receives thousands of support tickets per month. Billing questions, contract changes, outages, switching requests, questions about solar panels. Each time an individual message, each time individually answered, and then forgotten.
Categories existed. Labels existed. Filters existed in the ticketing system. But when you asked the simplest question — "What do our customers ask most?" — the answer was a shrug.
Not because the data wasn't there. But because nobody had ever read 22,912 tickets.
The blind spot
It's a paradox that plays out in many organizations: you have more customer data than ever, but less insight than ever. Every interaction is stored, categorized with a dropdown menu, and closed. The data grows. The insight doesn't.
At this energy company, the situation was no different. The support team knew from experience that billing came up a lot. They knew that switching requests were seasonal. But that was gut feeling, not data. And gut feeling doesn't scale.
The real questions — what are the patterns, how do topics relate, which questions could you have prevented — remained unanswered. Not from unwillingness, but from inability. You can't read 22,912 messages and expect structure. You need a different kind of reader for that.
What we did
We had AI read every ticket. Not to answer them — to understand them. Using embeddings, we converted each ticket into a representation of its meaning. Not the words, but the intent. "I want to cancel my contract" and "how do I stop my subscription?" are different sentences, but the same question.
Then we let a clustering algorithm find the structure. Which tickets belong together, not based on keywords but based on what the customer means?
The result: 40 thematic clusters. Each with a label, a summary, and a picture of how large the topic is relative to the whole.
Three things that surprised us
The largest cluster wasn't what you'd expect. Not complaints. Not outages. The largest cluster consisted of simple information requests — questions that should have been answered on the website already, but were impossible to find. Customers who called or emailed because they couldn't find the answer themselves.
That's not an AI problem. That's a communication problem. But you can only fix it once you know it exists.
Customers ask the same question in dozens of ways. A question about the monthly amount can come in as "what do I pay per month?", "my monthly installment is wrong", "adjust my advance payment", "I'm paying too much", or twenty other variations. Keyword matching catches three. Embeddings catch them all.
This is why traditional categorization fails: it's based on words, not meaning.
A significant portion of the tickets was preventable. Not all of them — but a substantial percentage was about topics where proactive communication could have made the difference. Rate changes that catch customers off guard. Annual statements that raise questions. Outages with no notification sent.
Every preventable ticket is a missed opportunity for better customer communication. And an unnecessary burden on the support team.
What this means
The ticket analysis isn't the end goal. It's the foundation. The insights lead to concrete actions: better FAQs, more targeted proactive communication, a knowledge base that agents have at their fingertips.
And it's the first step toward something bigger. The same knowledge base built from the tickets becomes the engine behind automated ticket handling, voice support, and a chat assistant. One analysis. Three channels. One integrated platform.
But that's a story for the next article.