Chatbots in Banking

How Data Science Drives Success  |  June 1st, 2017

From asking Amazon Alexa for traffic conditions, to receiving helpful tips from Slackbot, to using WeChat to book doctor’s appointments, bots are becoming omnipresent in our lives.

The bot market is hot! There’s a plethora of companies and investments in bots: VentureBeat’s 2016 Bots Landscape shows just under 200 companies ranging from personal assistants to AI tools to messaging, $22B in funding, and a very hefty $159B in valuation.

In this post, we explain why chatbots are rising in popularity with banks, the opportunities and challenges chatbots present, and where data and data science fit into the puzzle.

Banks are turning to chatbots

Increasingly, banking institutions are using chatbots for “conversational commerce” or “Voice-First Banking”—allowing banks to interact with customers (in real-time if desired) via messaging and digital platforms. Both speech-driven and text-driven chatbots can be found in use, depending on real-time constraints or level of acceptable asynchronous behaviors for the task.

Chatbots not only are helping banks with customer service, but also have the capability to impact the bottom line via upselling or identifying ideas for new products. These new data-driven insights and product offerings can potentially help banks differentiate. Some have even dubbed 2017 the year of the bot, specifically when it comes to how chatbots will change digital banking.

Why are banks turning to chatbots? This goes back to how banks engage customers and the costs associated with those channels.

Banks engage customers in a variety of ways, through:

  • Human channels—in-person transactions and customer service calls with a live agent or offshore agent
  • Digital channels—websites, mobile apps, email, online ads, etc.
  • Mixed channels—ATM, online chat, or interactive voice response system

These have varying costs, and customers have their own preferences for different transactions at different times. Keeping in mind that only a small percentage of customers are actually profitable for banks, the banks must balance their need to reduce costs with providing quality customer service.

With in-person and telephone customer service channels being the most expensive—and often the preferred channels for customers for certain services—banks are always on a quest to be more efficient. In the past, these efficiency efforts have included outsourcing and offshoring, both of which have plateaued due to changes in taxation, narrowing wage differences, and other market shifts. In response, banks have started looking at digital labor for cognitive automation, called robotics process automation (RPA). Chatbots are high on the list of promising technologies.

With these business considerations of why banks are turning to chatbots in mind, let’s look at the technology context for chatbots.

Technical challenges and opportunities

Chatbots have been around in some form since the early days of artificial intelligence, sometimes under other monikers, such as conversational interfaces or dialogue systems. Eliza was the first such program, developed in the mid-60s. Eliza used fairly simple techniques to recognize key phrases and then reply with a “matching” response, sometimes with personalization added to make it seem more realistic. Since then, there has been a continual stream of improvement on the basic paradigm.

From a business perspective, chatbots are a natural evolution of the automated online assistants that are already fairly prevalent on customer-facing web pages. In addition, many customers will be comfortable with chatbots, because they may be used to interacting with them in familiar messaging apps such as Facebook Messenger or Siri.

Startups such as Kasisto—which builds KAI, banking on messaging, and MyKAI, a personal banking bot—are building technologies to accelerate and broaden this movement. Kasisto’s AI-powered KAI was made available to banking customers last year by several banks, including Royal Bank of Canada. Kasisto raised $9.2M earlier this year and plans to both “bolster the company’s personal finance bot KAI and expand its AI virtual assistant offerings to markets beyond finance.”

Last October, Bank of America (BoA) and MasterCard announced that they will be releasing chatbots to help customers with basic questions, such as providing the most recent transaction. Both BoA’s Erica and MasterCard’s KAI will be using speech and text modes, with MasterCard’s KAI powered by Kasisto’s KAI banking (same name, but it’s separate from MasterCard KAI), to be available on Facebook Messenger. More recently, Western Union, not a bank but in the adjacent money transfer business space, announced a chatbot allowing customers to digitally wire money via Facebook Messenger. While not all the above chatbots are currently available to the general consumer (although they’re expected to be this year, and there’s likely to be room for improvement), the announcement indicates a desire to increase customer service by leveraging chatbots. BoA’s Erica, at least, is intended to go beyond basic questions, even providing financial advice.

From a technical perspective, what are the risks of turning to chatbots to improve customer service? In prior work on conversational interfaces, we’ve seen false starts in several hype cycles. There have been some much-publicized mis-steps in deploying chatbots in the wild, for example when Microsoft’s AI chatbot learned to respond in racist ways, or Amazon’s Alexa inadvertently ordered very expensive products.

Banking could avoid these types of pitfalls a few ways:

  • The chatbot technology becomes simpler to build when the customer’s range of possible business actions is constrained. The “rules” used by the chatbot are less complex in these scenarios. Since many aspects of banking are already partially automated, adding a chatbot “front end” to existing services such as auto-billing or uploading checks for deposit is an easy, quick win.
  • Chatbots should always allow an option for customers to access other channels of communication with the bank.
  • Flagging customer utterances that may expose the bank to risk or be fraudulent is a good idea, so that the interaction can be routed to a human.

There are several other considerations and tips to keep in mind. First, give customers clues on the types of things they can say, in a manner similar to what Siri does when it doesn’t understand you. However, don’t read out long lists of possible things they can say (think back to phone trees, when there were nine options and you had to listen to them all, because you weren’t sure which one was the best fit to your problem). Finally, today’s banking demographic has a younger generation with a different vocabulary than a formal banker of old. Your bot had better understand the slang and emoticons that come with digital interactions.

Interaction between data science and chatbots

How might data science impact chatbots or vice versa? Assuming that your organization is already leveraging data for business value, there are a multitude of data sources that can improve the quality of chatbot interactions. This includes customer data, such as segments, past account history, web page interactions, and even past customer service phone call logs. Less obviously, institutional data such as regulatory information, reward programs, and loan policies can also feed into the knowledge base that informs chatbot interactions.

For one example, customer segments can be used to tune the bot’s interaction style. For another example, transaction history can be used to predict why the customer might be contacting the bot in the first place, and thus expedite the conversation.

Finally, for banks that are just testing the chatbot waters, enterprise data can be leveraged to ease the transition. For example, banks may use data science and text analytics to make use of artifacts left by past technologies: system logs and their impact on customer retention and upsell can be mined to create bots that matter—that focus in on the most high-value and easiest-to-automate parts of the business to test the waters, then expand from there.

On the flip side, chatbot interactions can feed into data science projects to help inform business decisions. For example, different customers can be presented with different options to test new product ideas. An orthogonal effort could be put into mining chatbot logs for customer utterances that the chatbot didn’t understand, which could in turn indicate ideas for new products and service offerings, as well as simply opportunities to improve the next bot generation.

Conclusion

Banks have been turning to chatbots for cost savings and new product ideas. From a technology perspective, chatbots have come a long way. While there’s still a bit of room for improvement, chatbot technology is mature enough that some sophisticated, human-like conversations and transactions with bots are becoming possible, increasingly with the help of data and data science.