Key takeaways
- Why agentic AI differs from generative AI: agents write and execute code to call tools and take action, rather than just retrieving information, per Sri Ambati.
- TD Bank's "human first" principle: a bot never talks to a customer in the middle of a fraud incident; agentic tools work behind the scenes to support the humans who do.
- How TD used agentic AI internally to compress a weeks-long data-mapping and migration design process, by reasoning over both legacy context and future use cases.
- How TD is connecting fraud, cyber, and financial crime signals across fragmented legacy systems (mainframe, Java platforms, Oracle, Databricks) using agentic AI to unify the picture.
- Cost and latency tradeoffs in production: distilling and fine-tuning small language models, like H2O Danube, for narrow tasks such as call center classification instead of running large general-purpose models.
- Why AI ROI often shows up in unglamorous back-office use cases, and why an MIT study cited by Sri found two-thirds of AI efforts succeeded when paired with a trusted partner versus purely homegrown builds.
Overview
In this webinar, TD Bank's Annie Yan, VP and Head of Protect Product, and H2O.ai founder and CEO Sri Ambati discuss how agentic AI is reshaping fraud, cyber, and financial crime protection in banking. Annie explains TD's human first approach to protecting customers and how the bank is using generative AI today, with agentic AI as the next step. Sri unpacks the shift from prediction to action, the cost and latency tradeoffs between large and small language models, and what separates successful AI programs from stalled ones.