Most AI success stories sound too good to be true. This one’s different because it started with something we’ve all experienced: being put on hold while someone searches for an answer they should already know.
Essent is the Netherlands’ biggest energy provider, and like most large companies, it had customer service agents drowning in calls and struggling with clunky internal systems. What makes their story worth telling goes beyond the impressive savings they achieved. They got there by paying attention to what was broken and fixing it properly.
The problem was hiding in plain sight
Kellin Sjoerds, Conversational AI Engineer at Essent, did something most people in AI have forgotten – observing end users. In this case, it was customer service agents. What he saw was painful but familiar: agents spent nearly five minutes hunting through a massive knowledge base called Guru whenever a customer asked a question.
Guru had everything – thousands of articles covering every possible energy-related query, but finding anything was like searching a library where someone had removed all the signs and mixed up the card catalogue. There was no search function, no intelligent suggestions, just endless clicking through categories and hoping you’d remember the keyword that might unlock the right article.
“Agents had to know exactly what they were looking for and where to find it,” Sjoerds explains. Meanwhile, customers sat on hold, probably wondering why a simple question about their solar panels took longer to answer than installing them.
Two days to build something better
Essent runs regular innovation sprints, those brief periods when everyday work stops and teams tackle big problems. This knowledge base nightmare seemed like the perfect target. The team included AI engineers, conversation designers, and developers, and they gave themselves 48 hours to build a solution.
They created Bolt, an AI assistant that sits next to agents and instantly surfaces relevant information when they type in a question. Simple concept, but the execution required some serious behind-the-scenes work.
The biggest challenge was the content. Years of different people writing articles in different styles created conflicting information and inconsistent formatting. Before Bolt could be useful, someone had to clean this up and make it AI-ready.
“We found articles that contradicted each other, information buried in the wrong sections, and pieces that were so long they’d overwhelm the AI,” says Ricky Rekkers, Conversational AI Engineer at Essent. The team worked closely with content managers to restructure everything, breaking long articles into logical chunks and adding dividers that helped the AI understand where one topic ended and another began.
A slow and steady rollout
The response was immediate when the team demonstrated Bolt after their 48-hour sprint. One manager asked: “How fast can we implement this?” But instead of rushing to company-wide deployment, Essent took a careful approach.
The team started with one customer service team, whose supervisor had been closely involved in Bolt’s development. The transformation was dramatic. Within weeks, this team went from bottom performers to the top of the leaderboard. The results were convincing enough that other supervisors asked when their teams could access Bolt.
The rollout happened in two teams simultaneously, with careful monitoring and training at each stage. One crucial rule was introduced: agents couldn’t just copy and paste Bolt’s responses. They had to read, understand, and reframe the information in their own words. This kept humans in the loop and ensured customers got thoughtful responses, rather than robotic readouts.
Impressive results
Bolt saves at least a minute per call, which might not sound like much until you multiply it by hundreds of agents handling thousands of calls daily. Those seconds add up quickly.
The total savings reached €1.8 million in the first year, not from people losing their jobs, but from needing fewer expensive outsourced agents to handle overflow. More importantly, the remaining agents could focus on helping customers instead of playing hide-and-seek with information.
The improvements went beyond speed and cost. Agents reported feeling less stressed because they weren’t frantically searching while customers waited. First-time resolution rates improved because agents could provide complete, accurate answers instead of partial information that led to follow-up calls.
Improvements for the customer-facing chatbot
While building Bolt, the team also found a smart way to improve their existing customer-facing chatbot. Like many NLU-based bots, it sometimes struggled to understand what customers were asking. The old solution was pretty crude: when the system wasn’t confident about a query, it would disambiguate by offering the three closest match intents and hope one of them was close enough.
The team replaced and improved this approach by bringing-in AI for a single, clarifying question. If someone types “I can’t turn on my lights,” the system might respond, “It sounds like you might be experiencing a power outage. Is that correct?”
A clearer response will return to the chatbot if the customer confirms. If not, they immediately connect to a human agent without endless loops and frustrations.
This small change reduced unnecessary escalations to human agents by 12%. Again, it might not sound huge, but every improvement matters when handling thousands of conversations weekly.
What made it work
Essent’s success came from doing several things that many companies skip:
- Starting with a real problem. Not a vendor pitch or a boardroom mandate, but a pain point observed firsthand.
- Involving the right people early. Content managers, legal teams, and compliance officers were part of the conversation from day one.
- Testing carefully. Instead of a big-bang launch, they rolled out gradually, gathered feedback, and made adjustments.
- Keeping humans in charge. The AI made information easier to find, but agents still had to think, interpret, and communicate.
- Measuring what matters. Not just technical metrics, but real outcomes: call times, resolution rates, agent satisfaction, and cost savings.
The bigger picture
Essent didn’t begin its journey with the need to use AI. They began with the problem. The technology was important, but what made the difference was understanding what needed to be fixed and taking the time to fix it properly.
In a world where AI projects often promise revolutionary change, Essent’s approach might seem modest at first. But sometimes the most transformative changes are the ones that make difficult work easier, one conversation at a time.
Listen to the full episode on the VUX World podcast also available on Apple Podcasts, Spotify and YouTube.
