An agile approach allowed TD Bank and IBM to roll out a specialized artificial intelligence (AI)-powered virtual assistant in mere weeks.

Both IBM and the $1.4 trillion Canadian bank deploy an agile model that delivers in iterations, Daniel Cascone, financial services sector leader for IBM Canada, told Bank Automation News. That allowed a fast rollout of the customized virtual assistant for TD Bank’s Precious Metals digital store, which launched in mid-January.

“It literally took us weeks to implement this,” Cascone said. “TD has very much adopted an agile delivery model. It fits nicely within the bank’s model in terms of not only the development, but the training of the AI conversational bot, and it’ll continue to learn as it interacts with TD clients.”

The virtual assistant is already paying off for TD Bank in achieving its goal of reducing calls that otherwise would go directly to the division’s traders, Cascone said. It does so by handling common but basic shop-related questions around, for example, how price is determined, whether there’s a minimum or maximum product count or dollar value when making a purchase, and delivery options. The bot can also answer specific questions about where an order is or if it can be picked up.

TD Bank declined to provide specific metrics for the virtual assistant, referring BAN to IBM’s release. In it, James Wolanski, managing director and head of retail and wealth distribution and product innovation at TD Securities, said the bot was a response to customer demand.

“We know our customers are looking for an enhanced digital experience and the new virtual assistant will provide quick responses to help customers feel confident in their purchasing decisions,” Wolanski said. “Our TD Precious Metals support desk will remain available for any inquiry that may require additional support or a human touch.”

Currently, the TD Bank virtual assistant responds via text, although the IBM Watson Assistant can support speech to text. Customers type their questions into the virtual assistant and receive an instant written response, along with links to help further assist them.

Making AI explainable and customized

The virtual assistant is powered by IBM’s Watson Assistant, which is ranked as a leader in IT research firm Gartner’s Magic Quadrant for Enterprise Conversational AI. Also ranked were Denver-based OneReach.ai; New York-based Amelia, formerly IPsoft; Germany-based Cognigy; and Greece-based Omilia. Solutions from U.S.-based Google and Oracle both rank as challengers.

Watson Assistant is designed specifically for enterprise use in that it allows data to remain with the bank and supports governance with a traceability in how it determines answers, Cascone told BAN.

It also incorporates explainability as part of that transparency, which is created in part by the overall design and product principles of Watson Assistant. It also relies on “a very tight data lineage,” algorithms processing and governance, Cascone said.

“What we look at is why did the AI system reply in a certain way? And what we were able to clearly demonstrate is what was the algorithm used in the training? What were the answers during that training? What did it sell?” he said. “Throughout the entire process, so from inception all the way through the lifecycle, it is constantly tracking every data point, from not only the training model itself, but the data that it’s interacting with — where that data is coming from?”

The virtual assistant project started by training the AI to answer questions the division had historically fielded — essentially, an FAQ.

“You’ll also take the opportunity to conduct design thinking and some empathy mapping in terms of what are some of the features, what are some of the product capabilities coming out in general as part of the overall site,” Cascone told BAN. “And you’ll start to pre-train that just anticipating the types of questions that may be coming down the pipeline as well.”

After that, the AI starts to self-learn by analyzing data and the type of questions it’s receiving, down to the types of questions by the time of day. Then, a TD employee looks at the analysis and comes up with answers to the incoming queries.

“Eventually, as it gets more of a pattern in terms of types of questions that start to align, it will start to make its own judgment as well,” Cascone said.