TD Bank Group is no stranger to artificial intelligence.

It began using Kasisto’s conversational AI platform to communicate with customers in 2017 and a year later bought AI technology startup Layer 6. It also started working with Flybits to use AI to personalize its mobile banking offering.

But lately, the bank has been kicking those efforts up a notch — deploying AI in every business line throughout the bank, conducting research to find out what consumers think about AI, and holding a recent roundtable with 18 experts from different industries to speak on responsible AI.

Its experience could serve as a model for other institutions, some of which have struggled to implement AI.

Despite its commitment to the technology, however, TD executives said they are being careful about how it is deployed because they are aware of the potential pitfalls, including that human biases can be built into AI.

“Rushing to deploy AI can be fraught with danger if it’s not done in a very measured and careful way,” said Tomi Poutanen, chief AI officer at TD Bank Group and co-founder of Layer 6. “This is about people’s financial health and it’s our responsibility to be careful and measured in the application of AI. To me personally, responsible AI means that I can sleep at night. It means that when we build and deploy an AI solution, we need to be sure that it improves the world and isn’t unfair to any population.”

Data challenges

One of the biggest hurdles to adopting artificial intelligence software is having a data environment that’s not ready for it. A prediction or pattern recognition engine that is applied to information that is out of date, for example, could lead to disastrous consequences: an investor being given bad advice, a borrower being approved for a loan who clearly shouldn’t be.

In a recent survey conducted by Harvard Business Review, 31% of executives from large companies surveyed said outdated and inadequate technology for quickly deriving insights from data was one of the biggest challenges they face to adopting AI.

TD has been working on this, according to Michael Rhodes, the bank’s group head for innovation, technology and shared services.

“When you step back and think about what’s required to be successful in AI, you need to get all your data sorted and available in a way that can be used,” he said. “You need to have strong analytic capabilities and then you have to have the ability to deploy the insights into the marketplace.”

Most banks have separate applications for things like managing customer interactions, managing risk and handling operations, he said. It’s not uncommon to have data siloed in 200 or more systems. Getting all that data into a common place where an AI system can analyze it and derive insights is a big task.

“Once the data is sorted, you can do great things with it, but that effort to pull the data together is nontrivial,” he said.

Though the bank has vendors it works with, “the real hard work is internal,” Rhodes said. “It’s understanding your data.”

The bank has made “tremendous investments” in making its data safe, secure and accessible in the past couple of years, Rhodes said; he declined to share any specific numbers.

Applying AI to every line of business

The bank’s purchase of Layer 6, an AI technology company, was unusual. At the time, Layer 6 had 17 data scientists who had developed an AI prediction engine that could ingest a variety of data types, including customer records, transactions, phone calls, photos, documents and video and forecast things like a next best action for a customer or which transactions are most likely fraudulent. All 17 have stayed on, the bank says.

In part, the goal behind the purchase was to deploy Layer 6 technology across multiple lines of business. In many cases, that technology is replacing existing statistical models.

“We’ve found that no matter what business line or what problem the AI system was looking to solve, it greatly exceeds the statistical approaches in terms of prediction and accuracy,” said Poutanen.

In some cases, accuracy has improved more than 30%.

“It really is very impactful improvement in in accuracy that drives a business decision,” he said.

One area of the bank where TD is applying AI is to what it calls the “homeowner’s journey.”

Here the AI engine examines the bank’s data for signs that a customer may be interested in buying a home. “If we have some perspective in advance, it helps us reach out when and have a conversation with the right advice and the right person,” Rhodes said.

Knowing that a customer is likely to buy a house in six months, the bank rep might talk about savings products or debt capacity, to help get the customer set up for a mortgage.

Another area where the bank is using AI is to serve up everyday advice to customers. Other places the bank is starting to use AI include personalization, risk management, and fraud detection.In TD’s investment bank, for example, it’s using AI to predict price moves over seconds.

Responsible AI

Rhodes and his team had their own point of view on what responsible AI meant, but they wanted to calibrate that with leaders outside of the bank and the financial services industry. They put together a roundtable of people from banking, technology, fintech and academia “so we could form a more holistic view,” Rhodes said.

Three themes emerged from the roundtable: explainability, bias, and diversity.

“AI is unlike a statistical model where the explainability kind of leaps out at you,” Rhodes said.

TD Bank Group has developed model validation tools to understand the explainable variables in an AI model. To confront bias, it’s back-tested all its models to ensure it’s getting the right types of output and controlling for any unintended bias. To tackle diversity, the bank has put together an AI team that is diverse in gender, race and ethnicity, Rhodes said.

Jodie Wallis, managing director for AI for Accenture in Canada, who attended the roundtable, said there was a lot of discussion about monitoring and management.

“To date, the focus on AI for large organizations has been largely, can we develop and implement AI solutions?” she said. “The working group was really strong on the fact that not only do we need to design and develop them effectively, but we need to make sure that over time we are managing them.”

Wallis breaks responsible AI down into a slightly different list of issues: fairness (the opposite of bias), ethics (consistent with an organization’s code of business ethics), accountability (making sure that for each critical decision there is a human ultimately responsible), and transparency or explainability (making sure that important decisions made by AI can be explained appropriately to the people who were most affected).

Customers seem to be on board

TD Bank Group recently conducted a survey of Canadians to gauge their feelings about AI.

The majority (72%) said they are comfortable with AI if it means they will receive personalized services; 87% said they want banks to be innovative; 81% said they are willing to try new forms of technology from financial services companies.

“I was pleasantly surprised to see that customers see a lot of upside in AI,” Rhodes said.

They also had concerns: 70% say customers should have control over how their data is used; 55% say companies should be transparent about how they use the technology; 28% say decisions made using AI should be easy to explain and understand.

“We learned that people expect banks to innovate, people are very comfortable with AI, and they get the utility of AI in their everyday lives,” Poutanen said. “But when it comes to financial services, I think they have great greater concerns around what does this mean to their access to credit, what does this mean to insurance premiums, and so forth.”