Just over seven years after acquiring Toronto-based artificial intelligence (AI) company Layer 6, Toronto-Dominion Bank (TD) is gearing up to open Layer 6’s first office outside of Canada.
“I think it’s going to be a very exciting future, but also one that we have to navigate carefully.”
Maksims Volkovs, TD
TD has turned Layer 6 into the bank’s AI research centre since the buyout, leveraging its now 100-person team of experts to help improve TD’s operations and customer experience. While many of the people behind Layer 6 have since moved on to launch new ventures, Layer 6 co-founder and head of AI Maksims Volkovs has stuck around to help lead TD’s AI efforts as senior vice-president and chief AI scientist.
BetaKit sat down with Volkovs to discuss what Layer 6 has been up to since then, the new office TD plans to open in New York City (NYC), how Canada’s second-largest bank has been using AI, and what he is both most excited about and afraid of when it comes to AI.
“As a bank, we want to be leaders in this area, and we’re really excited about this Layer 6 expansion,” Volkovs told BetaKit in an interview. “I think it’s going to be a very exciting future, but also one that we have to navigate carefully.”
Volkovs argued that Layer 6 has built “one of the leading AI centres in Canada.” TD recently ranked ninth globally in terms of its AI capabilities on Evident’s 2024 AI maturity index for banks, second in Canada only to Royal Bank of Canada (RBC).
RELATED: TD Bank launches Innovation Partners banking platform for tech companies
Volkovs described Layer 6’s NYC expansion as “a natural next phase.” Located at One Vanderbilt, Layer 6’s NYC office will open later this year and complement its headquarters in Toronto’s MaRS Innovation District and satellite offices in Montréal and Vancouver. To start, it will be staffed by 20 AI experts that include data scientists, applied machine learning experts, and generative AI implementation specialists.
TD will use the NYC office to support the bank’s United States (US) operations and develop AI products for US colleagues and customers. Volkovs said TD chose NYC for its existing employee levels and academic relationships, as well as its role as a major financial hub with strong tech talent.
Since OpenAI’s launch of ChatGPT, Layer 6 has allocated the bulk of its resources to deploying generative AI. TD has used the tech to help contact centres eliminate transfers, shorten calls, and provide customers with relevant info faster. The bank has also used AI to help employees retrieve document details more efficiently, and index unstructured data.
Volkovs said TD has developed over 60 AI solutions to date across its business. He said it took Layer 6 approximately 20 weeks to get its first contact centre AI application to an acceptable level of accuracy, whereas it took four weeks to prepare a follow-up use-case. Volkovs noted that all of TD’s customer-facing AI solutions still have a human in the loop to vet the system’s output.
The next phase of Layer 6’s work will involve exploring agentic AI extensions to improve its back-office operations. “If this [generative AI] system can retrieve information, we think that it can do the task as well,” Volkovs said.
Volkovs said TD will never train its own large language models (LLMs) because the economics do not make sense. However, the bank does see “a lot of opportunity” in fine-tuning smaller, existing LLMs on its internal data to increase their accuracy. He is hopeful that this approach will allow the bank to deploy these AI models more cheaply without sacrificing accuracy.
RELATED: Leaders discuss the right speed for innovation and regulation at Remarkable AI conference
TD is currently using AI models from OpenAI, and has also partnered with OpenAI’s Toronto-based competitor Cohere.
Volkovs said he is particularly excited about recent advances in generative AI, particularly the evolution of large-scale foundation models that are capable of generalizing and completing tasks they have not explicitly been trained on.
“We’ve never seen this before up until this [generative AI] explosion,” Volkovs said. “We now have foundation models in language, vision, and chemistry. I think that’s going to completely change each field.”
“I didn’t actually think we were going to get there this fast.”
Previously, TD and others would build AI models for specific applications. Volkovs noted that foundation models have led to significant jumps in accuracy. “I didn’t actually think we were going to get there this fast,” he added.
AI governance and safety are what keep Volkovs up at night. He emphasized that it is “absolutely critical” to embed safety into AI models and downstream applications.
TD has a team dedicated to AI governance and safety, and is continuously conducting research into this area—important given the inherent risks and regulation in banking. Where many AI developers previously focused on maximizing accuracy, Volkovs said he has seen an uptick in interest in AI safety and governance lately that has been encouraging.
“We can’t just blindly improve the accuracy of these models … We have to make sure that all guardrails are in place and are really firmly incorporated into these applications,” he added.
Feature image courtesy Pexels. Photo by Erik Mclean.