The energy sector is digitizing. Smart grids, digital billing, solar panel management — the technology is moving fast. But there is one domain that lags behind: customer contact. And when we applied AI to it, we learned things we did not expect.
Lesson 1: The data is a mess. And that is normal.
Before you can analyze, you have to read. And customer service data is not the clean dataset you see in a tutorial.
Tickets with only a subject line and no body. Tickets forwarded four times with “see below” as the only addition. Tickets where the customer starts with a billing question and switches halfway to a complaint about an outage. Tickets in German. In emoji language. In pure rage.
The natural reaction: “we need to clean the data first.” That sounds logical. But it is a trap. Customer service data never becomes clean. There is no moment when all tickets arrive neatly formatted with a clear subject and a structured body.
The better approach: build a system that can handle messy data. An embedding model that works with half sentences, forwarded emails, and unstructured complaints. A pipeline that treats missing fields not as errors but as data characteristics.
This felt like a compromise at first. In hindsight, it was our best decision. Because the system does not only work with today’s data — it also works with last year’s data, and next year’s. Regardless of how messy it is.
Lesson 2: What you learn is sometimes uncomfortable
The analysis produced insights nobody expected. The largest cluster of customer questions turned out not to be about complaints or outages — it was about simple information requests. Questions whose answers should have been on the website.
But the answers were not there. Or they were, but impossible to find. Buried under three clicks in an FAQ that was last updated when the website launched.
That is an uncomfortable insight. Because it says: a substantial part of your support costs is the result of poor communication, not complex problems. And that conversation — “we need to improve our own communication” — is harder than any technical conversation.
We also discovered that customers express the same frustration in dozens of ways. The ticketing system had five categories. Reality had forty variants. The gap between how an organization thinks customers communicate and how customers actually communicate is larger than you expect.
Lesson 3: The hardest part is not the technology
This is the lesson that took the longest to learn.
The AI model worked after two weeks. The embeddings were good. The clustering produced usable results. The FAQs were factually correct.
But the organization needed considerably more time to actually use the insights. Not out of unwillingness — out of a very understandable combination of hesitation and habit.
“The customer says X” is a different conversation than “the system says the customer says X.” The first, they trust. The second, they have to learn to trust. And you do not build that trust with a nice dashboard. You build it with evidence, repetition, and by bringing people along in the process.
We learned that implementing AI in an organization is not a technical project with an organizational tail. It is an organizational project with a technical component. The technology is the easy part. The change in how people work, decide, and trust — that is where the real challenge lies.
What this means for the energy sector
The energy sector is facing a wave of AI adoption. Not just in customer contact — in operations, in forecasting, in asset management. And the lessons we learned in customer contact apply more broadly:
Your data will be messy. Build systems that can handle it.
The insights will sometimes be uncomfortable. Embrace that. It is information that makes you better.
Technology is not the bottleneck. The organization is. Invest at least as much in adoption as in implementation.
And perhaps most importantly: do not start with the technology. Start with the question. What do you want to know? What would you change if you had the answer? And only then: how do we build the system that provides that answer?