TD Insurance’s client-facing chatbot puts generative AI directly into customers’ hands. 

Launched in May, TDI Virtual Assistant answers common insurance questions in plain language, pulling answers from across the TD Insurance website and knowledge base, Kristen Gill, vice president, executive journey product owner, general insurance journey, TD Insurance, told FinAi News.

TD Insurance is a subsidiary of TD Bank. The insurance arm is the only one deploying the TD gen AI chatbot.

Insurance is a natural place to start because it generates a high volume of general questions, from coverage basics to definitions and early-stage exploration, Gill said, adding that AI is well suited for such tasks. 

“Across TD Insurance, about 80% of our clients already engage with us digitally and we know they continuously expect simple, low friction ways to get quick answers,” she said. 

Other FIs with insurance arms have not taken the leap to consumer-facing gen AI chatbots. For example, Bank of America, JPMorgan, Citi and Wells Fargo continue to lean on traditional natural language models and machine learning for their chatbots. 

And Bank of America explicitly states that its chatbot Erica runs on machine learning rather than gen AI. 

The development 

The tool was built in-house by Layer6, TD’s AI research and development center, and runs on a custom model rather than a single off-the-shelf provider, Christopher Cooney, vice president of data analytics and modeling at TD Insurance, told FinAi News. 

“The LLM architecture was built and optimized by Layer6 using a mixture of GPTs, including closed-source models,” Cooney said. 

Before launching, TD Insurance put the tool through layered testing. 

“We took a structured and iterative approach to testing the TDI Virtual Assistant before launch,” Cooney said, including end-user testing to validate functionality and whether responses reflected the bank’s content and business context. 

The bank also ran adversarial testing throughout development “to help identify potential gaps and improve performance over time,” he said.  

Adversarial training is a technique where a model is deliberately exposed to crafted, deceptive, or worst-case inputs during training so it learns to handle them correctly rather than being fooled by them in production, according to Github. 

A monthlong internal pilot with roughly 300 TD Insurance colleagues helped validate the model’s performance “in a real-world setting, providing additional insights to support ongoing improvements,” Cooney added. 

The $1.8 trillion bank’s insurance arm aims to deepen customer engagement with its new AI tool rather than reduce costs or labor, Cooney said. 

“The virtual assistant is intentionally designed with clear boundaries — supporting general insurance questions while directing users to contact TD Insurance when a more personalized or complex conversation is needed,” Cooney said. 

The case for in-house models 

TD’s decision to build rather than buy reflects a broader shift in how financial institutions approach AI. 

Ashish Nagar, founder and chief executive of Level AI, a customer-experience automation vendor whose platform runs on custom-built models, said data security is the biggest driver. 

“For financial institutions, it is data security and privacy first,” Nagar said. “Our own LLMs allow us to take all this data and not send it to any third party and do processing on all of this in-house.” 

Performance and cost follow close behind.  

Nagar pointed to TD Bank‘s scale — which he estimated at around 50 million conversations a year — as exactly the kind of volume that strains third-party providers. 

“Imagine if we had to send all the data to OpenAI. They would rate-limit us,” he said, adding that latencies would result in a subpar product. 

Nagar said he expects financial services to lead the move toward open-source and custom-built tooling.  

Level AI deploys multiple open-source models such as Meta’s Llama, Nvidia’s Nemotron and Alibaba’s Qwen, Nagar said, adding that Level AI’s token costs are nearly 30 times lower than closed source models with the same usage and accuracy. 

Many FIs will pivot to open-source models in the coming years in response to token costs and boost security of their data.  

Register here for the FinAi Lending Summit, set for Oct. 7-8 in Las Vegas.