How TD wants to bridge the confidence gap with AI-based tools | Bank Innovation | Bank Innovation

How TD wants to bridge the confidence gap with AI-based tools


Canadians may accept that AI is a part of their lives, but many express concerns about the potential risks that the application of the technology presents, including biases. That’s according to a recent TD Bank survey of 1,200 Canadian adults carried out by Environics Research.

According to Matt Fowler, vice president of enterprise machine learning at TD, the bank is taking steps to build confidence among customers, including an emphasis on ‘explainability’ and building capabilities to detect bias. The survey found that 72% of Canadians are comfortable with AI if it means they’ll receive more personalized services, while 77% are concerned about the risks that AI poses to society.

“It comes back to trust,” Fowler told Bank Innovation. “We are not a technology or social media [company]; we’re not going to deploy something and then quickly take steps back.”

Just over two-thirds of survey respondents said they believed the lack of diversity among people working in the field of AI could lead to biases in how the technology is being developed. It’s a risk other large banks, including Bank of America, are trying to address. Indeed, Cathy Bessant, chief operating and technology officer, recently said, as AI and machine learning tools are deployed to assess customer behavior or creditworthiness, understanding where bias is introduced is key.

See also: BofA’s Bessant: Need to guard against bias in AI banking tools

From TD’s point of view, the issue isn’t that the data itself, or the models, are biased. “When you remove data that could present bias in the model’s output — for example, men, women or address — the models are so smart that they can introduce biases through other information [considered], including where you shop, what you buy, etcetera,” Fowler said.

To address bias, TD has developed model validation tools to understand the “explainable” variables in an AI model. In addition, the bank has back-tested all its models to ensure it’s getting the right types of output and controlling for any unintended bias, Fowler noted. TD’s efforts have been supported by Layer 6, the AI company it acquired in early 2018.

Another component to address AI biases is ensuring a diversity of people who develop and test AI tools. The bank, which this summer convened an expert panel to look into the survey findings, sees diversity as a powerful tool to help fight bias associated with deployments of AI tech. “Diversity in the team that builds AI models is key,” said Fowler, noting that 90% of the Layer 6 team actually are from outside Canada. “Diverse backgrounds, diverse approaches and [diverse] thought processes gives us a better end product.”

TD also has been working to better explain the capabilities of AI to customers, including opportunities for more personalized customer experiences. “We don’t hear about a lot of positive [aspects of AI], which is where we can offer hyper-personalized service and where branch managers would know an individual customer.”