Headquartered in Switzerland, BPC—a global leader in payments solutions—is powering banking payments and e-commerce transactions for over 450 enterprises in over 120 countries. It has been using AWS since 2014. “We continuously innovate to remain competitive,” says Artem Zhukov, DevOps team lead at BPC. “At BPC, our goal is to help our customers drive innovation and deliver the payment experiences of tomorrow. To support this, we began exploring the implementation of AI features to benefit both our employees and customers.”

To improve the customer experience, BPC started to develop an intelligent chatbot that could respond to customer inquiries. “We use generative AI to enhance our operational efficiency,” says Eugene Tochilkin, engineering manager at BPC. “We aimed to develop a chatbot that could be used by both our clients and customer support teams. For instance, an agent can enter a client’s request into the chatbot and pass the generated response to the customer after review.” This chatbot was trained for high accuracy so that it reduces the workload for customer support agents.

To achieve maximum performance and deep AI capabilities, BPC uses large language models (LLMs) as the core of its chatbot. To deepen the capabilities of its solution, it also used retrieval-augmented generation, a technique that fetches data from BPC’s data sources and enriches prompts to provide more relevant and accurate responses.

Storing its data on AWS, BPC unlocked its full potential without sacrificing security using Amazon Bedrock, a fully managed service for building and scaling generative AI applications using foundation models. “To ensure optimal performance, we require an environment with advanced security features,” says Peter Ilosvai, head of operations at BPC. “Using Amazon Bedrock, we can resolve matters quickly and identify the causes of disruptions.”

BPC evaluated several foundation models in Amazon Bedrock. “Our prompts are huge,” says Tochilkin. “We put a lot of information in our prompts to produce highly relevant answers, which can drive up costs.” To balance cost with performance, the company uses Anthropic’s Claude 3 Haiku, which is one of the fastest and most cost-effective models in its intelligence class. (See figure 1.) “As soon as Claude 3 Haiku became available, we changed the settings of our engine to support it,” says Tochilkin. BPC has since reduced the cost of its chatbot by 66 percent compared with other models.

After deploying its chatbot to production, BPC offered it as a rebranded solution to its customers. “We’re getting positive feedback,” says Tochilkin. “Our customers can deploy the chatbot in their own sandbox environment and use it for marketing and advertisements.”

Next, BPC focused on streamlining its developers’ workflows by testing Amazon Q Developer, a generative AI–powered assistant for software development. Before this solution, some of the company’s developers used another AI assistant to help them debug code. But the quality was not good enough for company-wide adoption—the tool had an average acceptance rate of 30 percent according to the tool vendor.

BPC began piloting Amazon Q Developer with a group of eight developers. In 4 months, its developers generated over 4,000 lines of code with an acceptance rate of 46 percent. The group has also used the service to perform security scans on 6,640 lines of Java code, helping improve developer productivity. “It significantly reduces the amount of time spent on simple code and repetitive tasks,” says Stanislav Krainov, developer team lead at BPC. “Now we can focus on the more intelligent, challenging, and high-quality code in addition to our security posture.” Encouraged by the results of this experience, BPC plans to use Amazon Q Developer to help automate the refactoring of the company’s codebase and to debug application logic issues.