In recent years, there seems to be a sense of urgency for banks to go digital and expand into new communication channels. In ten years time, physical brick and mortar banking might not be the preference of the majority of customers. To attract younger millennial customers, banks seem to be realizing the need to understand their preferences and interact with them in the way they want to be communicated with.
New digital communication channels, such as chatbots and virtual assistants available through banking web portals or mobile apps, seem to be gaining popularity among the millennial customers. Natural language processing (NLP) plays a key role in making these channels work. That said, NLP is a far broader term that includes many other capabilities that could work for banks.
We spoke with Peter Hoopes, VP WW Sales at Gamalon, Inc., to better understand how NLP can be applied to banking. Hoopes laid out some of the value that it might bring to customer service in banking when he said:
“We see a lot of customers wanting to talk over free form text, social, voice or text, chat messengers. The whole ‘Google effect’ means that customers don’t want to talk to a person anymore or don’t want to visit the branch. Research firms like Gartner and IDC seem to agree that in the near future a majority of the new-age customers will want to communicate in natural language voice or text to computers and expect the systems to understand them.”
Banks might need to adapt to the new expectations of younger customers, who might not want to visit the branch unless absolutely necessary. Some of the larger banks have been forerunners in adopting new communication channels, and many of them have launched chatbots, virtual assistants, or conversational interfaces of other kinds.
For the larger banks, this poses an additional challenge of having to assess millions of messages coming in from additional communication channels.
Apart from these channels, most banks also keep record of more traditional customer interactions that happen over calls, texts, or website forms. Banks seem to be collecting increasing amounts of such data, which further supports the idea that banks might focus less on maintaining physical branches in the future.
That said, banks that have launched new customer communication channels will also need to adapt to new regulatory compliances relevant to chatbots and conversational interfaces. Banks might need to monitor and analyze customer complaints to identify cases where the bank may be at fault, thereby potentially breaking regulations.
Today, with customers leaning toward digital banking, a large bank might receive millions of customer messages through online portals, and it might be almost impossible to read through all these messages manually and identify issues that customers seem to be facing.
Processing this volume of the incoming messages could cost a considerable amount of human effort, time, and resources to analyze.
NLP could also help banks manage new communication channels (conversational interfaces) by automatically reading through a large number of customer interactions. Furthermore, NLP-based software might help banks identify and prioritize customer complaints that might need action from their regulatory compliance team.
In this article, we’ll discuss the impacts of AI and NLP on brick-and-mortar banking, specifically two applications:
We’ll then discuss the challenges that come with adopting AI and NLP at Established banks. First, however, we’ll start with an analysis of chatbots in banking:
One of the more common applications for NLP is in customer-facing chatbots and conversational interfaces. Most banks seem to be phasing out of large scale brick and mortar operations, and conversational interfaces might be what a majority of banking customers prefer in the near future.
There seems to be a general dissatisfaction among customers about conversational interfaces and their ability to accurately deliver useful information or respond to queries by “understanding” the context of the conversation in a way that humans could. Banks will need their conversational interfaces to improve in order to meet customer needs without needing to escalate the customer to a human customer service representative.
Hoopes considered the example of chatbots to explain what might be driving the need for better NLP capabilities:
“When we look at the early chatbots, such as those developed by most large banking financial firms, they get criticized over being unable to respond to clients accurately. People are underwhelmed by what chatbots can do and get frustrated with the interface as they feel they are not getting their voice heard.”
Hoopes noted that customers right now tend to speak or type to chatbots robotically. If a customer were inquiring about their account details over a call with a customer service rep, they might say, “Can you tell me my account balance?” In this case, customers might simply say “Need account details.”
Understanding that both these requests mean the same thing is easier for humans with financial context. Allowing humans to categorize these types of customer inquiries might help accelerate the “learning” for NLP algorithms. In this case, a banking subject-matter expert might indicate a list of such phrases that all might essentially mean the customer wants details on their account.
Periodically allowing these experts to add more of these word or phrase associations might help the algorithm improve the accuracy of categorization.
It is possible that the NLP algorithms might discover these associations automatically, but this could require much more training data and time than if a human expert helped the algorithm categorize the requests.
However, Hoopes also stated that not all the categorizations require a human in the loop. For example, if a financial firm engaged with customers through a social media messaging platform, the algorithm might identify the sentiment in a particular message by detecting the use of commonly used positive or negative terms.
Such relatively simpler categorizations might be in the realm of being completely software-driven. Given enough examples to train on, the software could automatically identify which patterns show up again in any future customer interactions.
Hoopes states that while NLP algorithms could automatically learn to identify how to tag or categorize messages, this usually involves massive amounts of data and a long time to get the software to work the way it is supposed to.
In Gamalon’s case, their algorithm categorizes messages on its own first. Then, subject-matter experts at the bank tweak the categories the algorithm comes up with.
The benefits that conversational interfaces might bring to banking customers might make it easier for them to access information or file a complaint. These conversational interfaces will get better over time and the number of events that necessitate a branch visit from the customer might decline.
With the emergence of chatbots and other conversational interfaces, the banking regulations around the implementation of these new communication channels have also emerged. For example, the GDPR regulations state that banks that have implemented chatbots need to define policies and procedures for customer data protection, assess potential data risks, and adhere to codes of conduct.
Further adoption of conversational interfaces mean that banks collect even larger volumes of customer interactions. These new communication channels require banks to follow additional regulations. NLP-based regulatory monitoring tools might offer a way for the larger banks to manage new communication channels and ensuring that banks are complying with mandated regulations.
For example, customer complaints might contain cases where a customer might claim compensation for a fee that was wrongly charged by the bank. As Hoopes explains it:
“Considering the case where a customer calls in and says that the bank charged him an overdraft fee that they shouldn’t have, since he has overdraft protection. If it turns out that this is the banks fault, they might be breaking regulations leading to lower ratings from supervisory bodies.”
Hoopes also quotes an example from a client his firm worked with. The large bank had over 60 million customer complaint messages coming in from several channels every year. A small number of these complaints, where the bank was at fault, needed to be identified in order to firstly resolve the customer’s issues and also ensure that they didn’t attract any regulatory fines or rating cuts.
The bank needed to read through all the messages and understand their intents to categorize them as relevant for immediate action. The bank worked with Gamalon to develop an NLP-based categorization tool that helped classify complaints with regulatory importance by using input from the bank’s regulatory experts.
Banks might also find previously undiscovered patterns to identify customer support tickets that lead to regulatory violations. For instance, the software might identify that customers who file a complaint about misallocation of funds usually have a certain type of tone and use phrases or words that might be similar.
Banking regulatory staff could then identify and address more such customer issues earlier to avoid any non-compliance.
Regulatory compliance is a persistent issue for banks requiring constant monitoring to ensure that the bank is adhering to local, national and international rules for conducting financial transactions and handling customer data. As regulatory requirements increase (such the GDPR), the costs to serve an individual customer often go up for banks and AI software might help cut these costs down and allow banks to serve more customers.
In a real-world use case, Deutsche Bank claims they developed an AI software to monitor financial regulations. Their software can purportedly sort through large volumes of interaction data between customers and employees to ensure the bank’s employees are complying with rules and regulations.
The bank was finding it challenging to meet regulatory standards since a majority of their customers preferred communicating through online channels and the volume of incoming messages was too huge.
According to Deutsche, the NLP system could automatically search for terms that compliance auditors might look for, a task which previously meant manually going through tape and listening to several hours of audio recordings.
A few of the incoming messages might contain patterns in conversations that correlate to fraud or money laundering cases. For instance, historical customer conversations regarding fraudulent claims for stolen credit cards could be input to NLP-based software.
When the software finds new messages which have been tagged as suspicious, it can alert the bank’s fraud detection team.
A key friction for many businesses might be that they can’t seem to identify every scenario for customer search queries in chatbots. Teaching a machine which has no preconceived knowledge of human speech to make associations like humans is an incredibly difficult task.
NLP and machine learning offerings today support varying levels of visibility into the AI system. Many AI vendors offer software that works on supervised learning. Once the algorithm is trained on labeled data, it’s unclear how it “processes” that data, so to speak, to come to the conclusions it does.
In other words, an NLP software might be able to categorize certain messages as a request for an account balance, but there’s no way to really figure out why the algorithm categorized the message that way.
Other NLP vendors offer software that could double up as diagnostic tools and allow even non-technical subject-matter experts at banks to sift through customer data and possibly tweak the NLP algorithms to suit their application better.
For example, Hoopes claims that Gamalon’s systems allow users to rank and compare words that have been categorized to mean the same thing or look at probabilities for whether two sentences mean the same thing. He adds that this can be done for millions of words and phrases, and users can edit these lists to help improve the accuracy of the algorithm in “understanding” and categorizing free-form text messages.
The dichotomy of banks having to deal with growing customer interactions every year that they cannot handle and the fact that younger millennial customers expect a lot more in the customer service department than older customers seems to be driving banks towards adopting NLP software.
It might be crucial for banking leaders to identify ways to adopt to this technology, and the push towards digitization might render the brick-and-mortar banking traditions somewhat obsolete in the near future.
NLP software can help banks manage newer channels of communications. With more customers expected to prefer online banking experiences to visiting a branch in the future, NLP might play a key role in the full transition from brick-and-mortar banking to digital banking.
This article was sponsored by Gamalon, and was written, edited, and published in alignment with the transparent Emerj sponsored content guidelines.
About The Author
Raghav serves as Content Lead at Emerj, covering our major industry areas and conducting research. Raghav has a personal interest in robotics, and previously worked for research firms like Frost & Sullivan and Infiniti Research.
About The Founder, Emerj
Daniel Faggella is the founder of Emerj, which serves as an objective source for business leaders looking to make productive use of AI in their industry. Stay in touch with Daniel on LinkedIn or Twitter.