How We Built a GenAI Chatbot in 6 Weeks: Lessons to Share
“In just six weeks, we built and launched a generative AI chatbot that exceeded the core performance metrics of our first in-house solution – a project that had taken us almost a year to develop.”
Discover the full story, shared by Karolis Valaika, Head of AI Labs at CyberCare – a company that provides customer care solutions for global tech giants such as NordVPN, Surfshark, Saily, and others – to see how the team strategically leveraged GenAI advancements to drive measurable business impact.
As a customer care provider for global tech brands, handling over 4 million client requests each year, we saw the need to automate early on. At CyberCare, we started integrating AI and machine learning powered chatbots several years ago. Since then, continuous analysis, model training, and knowledge building have remained at the core of what our AI Labs team does. Over time, however, we realized that our basic chatbots had become static and unengaging – so it was time to look for a new solution. To better understand and address the needs of our customers, we chose generative AI as our tool.
Generative AI offers immediate value – it can communicate in any language, hold empathetic conversations, and lay the groundwork for chain-of-thought reasoning (even if it’s not true reasoning in the human sense). However, we had to acknowledge challenges like hallucinations and bias. Our goal was to overcome these issues without sacrificing the high standards we prioritise, especially accuracy.
After careful planning and full team effort, we set out to build a solution that delivers instant, always-available customer care. The result? A system that didn’t just meet expectations – it exceeded them in record time.
Here’s how we did it.
All-In Commitment
Everyone, from stakeholders to developers, approached this project seriously, viewing it as our big bet and a major step forward. It was meant not only to solve the challenges we had been facing but also to serve as a long-term investment in delivering outstanding customer service. We put together a dedicated team of four developers and one content manager to focus solely on building the new solution.
Defining the Success Metrics
We defined success through measurable metrics and data, primarily focusing on customer satisfaction ratings and the percentage of support cases resolved without escalation. We set an ambitious goal: to reach what our first in-house chatbot had achieved after a full year of development.
Preparation is Key
We explored a range of technical options to find a solution that would work not just for today, but also scale into the future. With key considerations like security, data protection, and long-term maintainability, one thing quickly became clear– we needed to build our own solution rather than rely on an off-the-shelf product.
LangChain, LangGraph, the rise of the Model Context Protocol (MCP) – the AI landscape offered plenty of architectural choices. And that was just one part of the puzzle. We also had to address observability, scalability, multi-tenancy, and more. After thorough planning and careful evaluation of each decision, we finally laid a solid foundation and were ready to move forward.
Start Small and Accelerate Fast
A key part of this architecture is intent management and classification. But with hundreds of possible intents, supporting each one can quickly become a time-consuming task. By identifying what matters most to our customers, we focused on the highest-impact intents first. Starting small and gradually adding new ones every few days allowed us to steadily expand the AI’s knowledge and capabilities – enabling it to handle more over time.
Human Touch is Still Needed
The most important part is having an escalation or handoff system in place from the start. Be ready for the unexpected – complex, emotional, or sensitive cases that AI can’t handle yet. We knew this wasn’t just a temporary measure until our AI became “perfect.” Instead, we accepted that some cases will always require human support. That’s why we built this mechanism carefully and made it a core part of our system. Human help should always be available whenever customers need it.
Go Beyond Troubleshooting
AI can do much more in customer care than just simple troubleshooting or sharing information. Once it’s live, new opportunities begin to surface – like using tools, APIs, or services to provide personalized, actionable support. Instead of asking customers to check their balance, do it for them. This not only saves valuable time but also ensures they get accurate information and clear next steps. And that’s just one simple example – there are many more ways AI can enhance the customer journey and improve the overall product experience.
What’s next?
The journey to deliver excellent customer service is never-ending.
With the rise of generative AI, the potential for transforming customer support is enormous. Our experience showed that building a powerful, effective solution doesn’t have to take a year. We’ve proven that it’s possible to adopt new technologies as they emerge – and do it at a speed that brings value to both our customers and our business faster than ever before.
To make our customer service even more timely and proactive, we’ll continue improving bot knowledge, refining scenarios, and enhancing personalization and memory usage. While chat support efficiency remains our primary focus, we’re also exploring ways to strengthen other areas, such as improving how we analyze and QA support and empowering our agents with AI tools to help them provide faster, more effective assistance.