Meet the AI-powered debt collection bot that’ll speak to, and text customers on your behalf
What does the future of voice and chatbots look like?
By Kiran Kumar Kuppili
Chatbots have proven to be a powerful channel that can be deployed with minimal effort. In the banking industry, chatbots are typically developed as 24/7 virtual assistants that respond immediately to frequently asked questions (FAQs). In more advanced implementations, different application programming interfaces (APIs) are also integrated into the backend for customers to complete various bank transactions at any time of the day.
Chatbot-powered channels play a major role in reducing time and cost spent on customer service. In 2016, we launched our first chatbot for consumer banking customers. A year after, we kicked off our conversational artificial intelligence (AI) journey with voice assistants to match human needs with AI capabilities.
Since then, we have rolled out different bots across multiple regions, with support for various languages. For example, our Joy virtual assistant assists DBS IDEAL (institutional banking) customers with checking their account balance and retrieving transaction details, including the status of issued cheques and trade applications. Separately, the DBS digibot for consumer banking customers is capable of furnishing information about mortgage products and preferred payment plans, along with a customer’s own personal loans, existing car loans, and applying for a credit card.
At DBS, chatbots don’t just empower a customer to self-help and complete their tasks. As the old saying goes, “Conversation is food for the soul”. We decided to emulate this and harness the same magic of meaningful conversations to better serve our customers and to build trust with them.
In this piece, I outline our journey to build an AI-powered debt collection voice-and-chat bot for our consumer banking customers in Singapore.
Automating debt collection with AI-powered voice bots
Debt recovery can be expensive, complex, and hyper-sensitive. A seamless, personalised, end-to-end conversational experience is required when the debt collection team contacts a customer about their payments, while educating them on the fees and penalties for delays in payment.
Our ground research showed that customers prefer text chatbots for complex transactions, and voice bots for simple interactions. A further study revealed a lack of trust as the main reason why customers avoided voice instructions when performing complex financial transactions. However, this could be overcome if a functional voice recognition technology was built to cater to this target segment.
Native language support was also essential as we scaled this use case. While there are available solutions in the market, vendors quoted us eight weeks to develop the AI-powered chatbot to automate debt collections. This led us to build our own conversational AI platform from scratch and resulted in us having a working solution in just one week.
A team of four engineers at DBS Tech India (DTI) built the in-house framework that could segregate customer actions, product system decisions, and validation components. Each of these components was designed to be a standalone deployable. With backend APIs made available for consumption, customer inputs could be configured via a flow designer user interface (UI), enabling users to design their own conversation flow. This resulted in reducing the end-to-end design and build cycle to one week, and the entire project from conceptualisation to completion, taking just four months.
We also implemented voice technology to directly communicate payment delays with consumer banking customers and educate them on related fees and penalties, to prevent their late fees from escalating.
The debt collection bot was developed to:
1. Deliver reliable performance using the latest AI technologies like ASR (automatic speech recognition), TTS (text to speech), and NLP (natural language processing).
2. Automate debt collection tasks with a collections bot to conduct personalised, human-like conversations that are seamlessly transferred to a human customer service officer (CSO) where necessary.
Humanising Technology
A core component of the strategy involved implementing a solution that could be quickly prototyped with the existing framework. While the intention was to initiate change, it was equally important to streamline internal workflows, establish a dedicated technical team, and integrate data sources.
To accelerate solution deployment, we collaborated with regional teams to build the speech-to-text (STT) engine. All other integration components were similarly built in-house. A diagram of the high-level component architecture can be seen below.
Our text-to-speech and speech-to-text components were built over multiple iterations; this enabled the system to understand DBS products and services. Other components, such as flow orchestration, natural language understanding, and NLP required enhancements to add empathy and sentiment as attributes.
Extensive research was done to compare multiple libraries available. Most libraries could not be retrained to support mixed languages, and only performed well when the conversation was confined to a single language. Other libraries offered the service only through an API. As this solution required DBS to share customer voice or bot responses to an external network, we ultimately did not proceed with it due to concerns around customer data protection.
Below is a summary of our findings:
Our Solution: The AI voice bot
Upon evaluating the open-source technologies available, we concluded that it was essential to use a custom-trained voice model for debt collection, as this would support dialect and dual language requirements. We built a prototype with Wiz.ai, and the results were encouraging. Our debt collection bot could converse in Singaporean English in a natural, human-sounding way. It could also make phone calls to multiple customers at the same time, and detect the customers’ intent to make payments in real-time. This was made possible using in-house trained natural language understanding, with common banking terms.
The AI debt collection bot makes calls to customers and can detect when a customer is speaking. At the same time, it can identify if the customer is digressing, and send an alert to a customer service officer who’ll take over the call. After each call, the chatbot summarises the conversation and prepares a report for the debt collection manager who would assess if the customer has to make a late payment fee.
While the primary channel of our voice bot is via calls, it can also be used via chatbots, as seen from the above example.
Other key features of our AI voice chatbot include the ability to:
· Extract names, numbers, and dates to feed flow orchestration
· Detect and handle user pauses efficiently
· Negotiate for a promise to pay
· Negotiate follow-up calls
Business Impact
By automating the existing debt collection process, our debt collection bot reduced the time taken to complete 200 calls from eight hours to just 45 minutes. By leveraging data-driven decisions and by pushing outcomes in real-time to our engine, our bot achieved a “Promise to Pay” of SGD 2.3 million in the first six months after its launch in Q1 2022. This is a rather healthy ROI, considering that our work was further optimised with reusable frameworks for additional bots, and brought down the cost of designing the conversational chatbot from SGD 90,000 to SGD 1,500.
With the success of this debt collection bot, we will be rolling out other loan products, which would be completed by the end of 2023.
Kiran is a Principal Engineer at DBS Tech India’s Future-Ready Technologies (FRT) department. He has close to 20 years of experience in the banking and finance industry. He is responsible for platform engineering and delivery for various platforms within FRT. Kiran is also responsible for hiring and nurturing talent to build resilient application frameworks for the bank.
