At the third annual Evident AI Symposium in New York City this past October—sponsored by Teradata—agentic AI was front and center. Senior leaders from major U.S. banks shared how they’re deploying this technology and what they’ve learned along the way.
Goldman Sachs: building the right environment for AI
Marco Argenti, CIO of Goldman Sachs, stressed that success starts with the right conditions for AI to thrive. For Goldman, that means a strong data foundation—especially the semantic layer that helps AI interpret information.
Argenti put it simply: “Data quality is the single most important factor in the quality of AI.” He compared the data lake to a road map, while the semantic layer acts like traffic signals, guiding AI along the right paths.
He also highlighted the need for detailed documentation of business processes. By codifying best practices and “tribal knowledge” (or tacit knowledge), Goldman ensures its agentic systems—especially those using reinforcement learning—can learn effectively from examples and institutional expertise.
Citigroup: one platform, global scale
David Griffiths, CTO at Citigroup, explained how Citi replaced a patchwork of tools with a single global platform. This move makes AI easier to govern and harder to ignore.
Griffiths emphasized that agentic systems need strong foundations. Citi is investing in private cloud infrastructure to scale efficiently and relies on open-source frameworks to stay flexible and avoid vendor lock-in. As agents take on more complex, customer-facing roles, Citi is doubling down on reliability.
Mastercard: fighting fraud with context
Greg Ulrich, Chief AI and Data Officer at Mastercard, focused on fraud prevention at the point of sale. The key? Add more context to transaction data and build a merchant graph for each cardholder to spot unusual patterns.
Ulrich shared that Mastercard successfully integrated this merchant-graph signal into its fraud models—without adding latency—helping validate cardholder intent and reduce fraud risk.
Capital One: reducing cognitive burden
Prem Natarajan, Chief Scientist and Head of Enterprise Data and AI at Capital One, described their approach: use reasoning to break down complexity and specialize through task-specific models.
He noted that inference costs have dropped by a factor of 1,000 in the past year—and further declines are expected. For Natarajan, the goal of agentic AI is clear: “Transfer the cognitive burden from the human to the system.” Agents at Capital One are semi-autonomous, understand context, and act on behalf of their creators.
Teradata solutions for agentic AI
Teradata’s Autonomous AI + Knowledge Platform uses the Financial Services Analytics Schemas to give agents the banking-specific context they need when working with large language models.
With Teradata’s Enterprise Vector Store, agents can retrieve accurate, up-to-date information at request time, preserving document lineage for traceable citations and reducing hallucination risk.
Teradata also offers MCP Server and Bring Your Own LLM (BYO-LLM) capabilities for secure, flexible orchestration. Agent actions can integrate with enterprise systems through secure connectors, while teams maintain control and avoid vendor lock-in by choosing the right model for each workflow—including those requiring long-context handling.