Financial firms bank on AI as pilot projects head to production

The financial services sector is pouring money into artificial intelligence (AI), with banks, for example, expected to spend $5.6 billion on AI in 2019 – second only to the retail sector.

Until now, the vast majority of AI projects have remained pilots, and in many cases those projects led to tech deployments without a clear business use.

Simply put, it’s been trendy.

Most AI projects today are aimed at improving customer service efficiency and security by introducing chatbot technology, or by deploying machine-based learning to uncover trends across business lines in customer behavior and what they need.

“It’s about ensuring banks are able to retain the memory of a customer’s journey across bank services,” said Sankar Narayanan, chief practice officer at analytics service provider Fractal Analytics.

Initial systems aimed at fundamental issues

As those test projects mature, there’s a sea change coming, according to Narayanan,  and beginning this year companies are expected to start rolling out production systems. At first, they’ll be focused on solving fundamental problems, such as customer relationship management through conversational chatbots that have advanced in their capabilities.

“The bigger idea is reducing friction,” Narayanan said. “Most banks, when providing lending to businesses, require lots of documentation. One significant friction point is the vetting of those documents. It’s a hassle for businesses seeking funds and bank officials who need to go through checks and balances for risk.”

For example, something as simple as automating credit line increases has typically relied on simple financial calculations: a client who uses a card regularly and pays on time gets offered a credit line increase. But very little research is done on how a consumer perceives an increase in credit.

“Will they be happy? Or will they think you’re giving them a longer noose to hang by,” Narayanan said. “It’s a simple question. So, it’s easy to see if a client qualifies, but genuinely is it the right thing for them? So, [by knowing their history more completely], you can offer them an increase versus automatically increasing it. That’s humanizing AI.”

Most financial services’ lines of business are currently compartmentalized, relying on proprietary or legacy computer systems adopted through the acquisition of other businesses that are not integrated with other CRM or ERP systems. For example, a bank may be able to see a client’s checking, savings and credit card history but not necessarily their mortgage background through a single, integrated view. Integrating that information would make it easier to offer up additional products based on their complete financial profile.

Chuck Monroe, head of AI Enterprise Solutions at Wells Fargo, said many organizations get stuck early on by viewing AI through a narrow lens of either data science or technology rather than as a strategic business tool that can be applied across the company.

“AI technology truly has the potential to drive transformational change. It’s critical to organize data in a way that allows you to pull meaningful insights across your company,” Monroe said. “I think it’s also important to clearly define opportunities and understand the end-to-end process; standalone AI solutions rarely meet business goals.”

A year and a half ago, Wells Fargo created an AI Enterprise Solutions team that  partnered closely with the bank’s data management and IT teams to accelerate adoption of AI throughout the organization. The team touched everything from customer experience to operations and risk management.

The Wells Fargo experience

Wells Fargo began its foray into AI with a Facebook chatbot pilot that began in April 2017 and ran for a year so the bank could evaluate how it enhanced and simplified customer interactions. The bank more recently conducted a short-term Banking Assistant pilot within the Wells Fargo Mobile app to learn about how conversational banking capabilities can improve customer experience and deliver banking information using AI.

The bank, however, hasn’t yet ramped up any production systems.

“We’ve learned a lot about how our customers prefer to use chatbots, which will help inform potential future experiences,” Monroe said. “For example, customers have appreciated the ability to access account information and analyze transactions, and we received very useful feedback on a number of additional capabilities they would like to see in future chatbot experiences, such as the ability to transfer funds and make payments.”

Getting started on AI development

Finding AI developers isn’t easy, as talent in the nascent field is scarce. It’s easier to find developers of rules-based technology than actual AI or machine-learning tech, said Sridhar Rajan, a principal in charge of robotic and cognitive automation at Deloitte consulting.

In creating an AI Development team and deploying AI, a company should first be clear on its business objectives and realize AI developers are rarely home grown; it’s a complex field that requires a lot of education and training. What’s needed, said Rajan,  is a developer who has a good grasp of the technology married to business acumen.

“The center of gravity is shifting toward business knowledge,” Rajan said. “You don’t want to say, ‘I have machine learning, where do I apply it?’ Look for business problems first to solve. Hire a small set of talent. Create a small core team through a center of excellence… like an incubator project.”

In August, a report from Deloitte pointed to the major sticking point for enterprises eyeing AI projects. They include: disparate legacy systems that do not talk to each other; a general lack of AI developers and programmers; and a lack of understanding about what AI can – and can’t – do.

Deloitte also noted that AI does not live in a vacuum but must be intertwined with the development of other technologies, such as blockchain or quantum computing.

AI and machine learning is primarily used for pattern detection to recognize  irregularities or regularities in data; foresight to determine the probability of future events; customization for generating rules from specific profits and applying general data to optimize outcomes; decision-making from generating rules for general data and applying policies against those rules; and interaction or communication with customers through digital or analogue media.

“When business people talk about AI, they typically are not talking about a particular technical approach or a well-defined school of computer science,” the report said. “Rather, they are talking about a set of capabilities that allows them to run their business in a new way.”

At their core, those capabilities are almost always a suite of technologies, enabled by adaptive predictive power and exhibiting some degree of autonomous learning, that have advanced the ability to automate and enhance services or internal processes.

There are four distinct areas where AI is now being used in pilots or production systems, according to Rajan:

• Chat bots and virtual assistants used by retail banks to answer mundane customer questions.
• Robotic process automation or rules-based scripts that can pull data from multiple systems to generate forms or invoices.
• Natural-language processing and generation, enabling systems to read text in contracts to pick out key clauses (and determine the implications of that text) as well as enabling the system to write in plain language.
• And cognitive analytics, which can find customer trends to determine which products they’re more likely to purchase.

Regulatory compliance gets more automated

Anti-fraud, anti-money laundering and know your customer (KYC) rules have also prompted companies to deploy investigative AI, which combs through internal and external resources to paint a more complete picture of potential customers.

When a client logs into a banking site, for example, an AI script would search for a client record, identify any missing data required for regulatory compliance, email the bank relationship manager and subsequently update the information received by the customer – whether it’s a person or a company.

Today, manually searching for missing client data to fulfill KYC and other rules can take as much as six weeks to onboard a new corporate client, according to Rajan. AI and machine learning can cut the time to onboard a new client to a few days, he added.

“The process to investigate and clear somebody is a very well-defined process. The time-consuming part is usually a function of gathering and aggregating data from disparate sources within your shop and outside,” Rajan said.

AI technologies of different capabilities are being used to more efficiently manage client onboarding, offer a more intuitive line of questioning tailored to them while also being able to cull their preferences in order to offer future products to them, Rajan said.

“A lot of financial services firms grew organically and have multiple systems and it takes a tremendous effort to bring the data on all those systems together for clients quickly,” Rajan said.

Wells Fargo looked into an AI model that can detect and continuously re-prioritize potential fraud cases, which would greatly reduce the number of high-risk cases handed over to employees to investigate.

“We have hundreds of thousands of debit card transactions marked as potentially suspicious,” Monroe said. “This…cuts down on the number of false positives for cases that aren’t actually fraud, which helps keep the customer experience [be more] seamless and secure.”

One “global financial institution” client of Deloitte’s, which the company declined to name for privacy reasons, faced a significant manual effort in reviewing each expense report and the supporting receipts for validity and accuracy. The company used AI to automate the reading of the reports and receipts, validating key fields and providing a summary. Additionally, the AI program stores all inputs and reports in a central location enabling audit trail.

The system saved the company “thousands of hours per year” in manual reviews of expense item and receipts. The system also identified rules non-compliance on a daily basis, saving out-of-policy reimbursements. And, the streamlined verification process made it easier to look for potential fraud.

The challenge of integrating legacy data systems with AI remains for companies seeking a unified view of client information. The ideal would be to have one or two data platforms where information flows smoothly, but that doesn’t happen often, Rajan said. As more companies move toward placing their data in the cloud – and away from a fixed infrastructure with multiple systems – AI can enable more automated management of that data.

Even so, AI technology can still help a business cut across different systems and bring data together.

“If the systems aren’t integrated today, which they are not, can I use a technology that will get customer information from a bank file and from a mortgage file, bring it together and present it to you internally so you can then talk to your customer with a unified view?” Rajan said. “I think that’s where we’re seeing AI bridging that gap today while the overall integration is going on. That’s where I think the acceleration of AI technology is happening.”