A powerful integration to transform customer service in Polish banking

In the banking sector, the quality of customer service directly impacts customer satisfaction, loyalty, and the institution’s reputation. Customers rely on fast and effective assistance for their inquiries. However, many routine queries, such as account balances, transaction history, or general account information, do not require specialised knowledge and can be handled efficiently by AI-powered solutions like voicebots or chatbots. These solutions provide fast, 24/7 accessible support, handle a large volume of customer inquiries simultaneously, and significantly reduce operational costs. Furthermore, they can efficiently route customers to the appropriate department based on the nature of their query.

The challenge at hand was finding a tool that could effectively support the Polish language. Our solution involved exploring available options and creating a custom solution that integrated Amazon Connect, which does not natively support Polish, with an AI automation platform that does. 

During the proof of concept phase, we evaluated various solutions available on the market that support the Polish language. We assessed them based on configuration flow, integration with third-party tools, the quality of voice recognition and generation, training and intent recognition, and more. Among the available options, BoostAI best met our requirements. This platform offers high-quality voice generation in Polish, easy flow configuration, sentiment analysis, and requires no installation (the access is via an API). It can handle increasing traffic with great scalability without compromising quality. The solution is intuitive, and no coding is required as maintenance is automated.

The flow between Amazon Connect and BoostAI is designed as shown in the diagram below.

First, a customer dials the customer service centre number through Amazon Connect. The connection is then redirected to BoostAI’s voicebot. At this stage, the AI algorithm recognises the caller’s intent and conducts the conversation following trained conversation paths. It’s worth noting that when redirecting to a call, the calling number is used, not the number assigned to Amazon Connect. This allows us to link the conversation in BoostAI with the ongoing conversation in Amazon Connect.

Now, the path diverges in two ways: either the problem is resolved, or the user requests an agent’s assistance. If the former occurs, the conversation transcript is sent to Amazon DynamoDB. Amazon Bedrock then summarises the transcript and analyses its sentiment, concluding the flow. If the latter takes place, the transcript is also sent to Amazon DynamoDB and analysed for sentiment. In the next step, the conversation is redirected back to Amazon Connect, where it is taken over by an available agent, and a new case is created in Amazon Connect, including all relevant information, such as the summary and sentiment of the conversation with BoostAI’s voicebot and contact history, if applicable.

Once the conversation with an agent is completed, the audio recording is stored in an S3 bucket, set as a trigger for a data processing pipeline. The AWS Conversation ID is extracted from the file name. Audio file transcription is performed using the Amazon Sagemaker service and Amazon Whisper model (a requirement for the Polish language). Finally, sentiment analysis and summarisation of the transcript are carried out using the Amazon Bedrock service.

Data from DynamoDB is presented on dashboards created in Amazon Quicksight, simplifying the presentation of the gathered data.