Beyond the MCP Demo: Building Governed AI Agents for Regulated Banking
About this session
Most Model Context Protocol demonstrations assume flexible cloud infrastructure, broad service permissions, and relatively simple authentication. These assumptions quickly break down when AI agents must operate inside a regulated financial institution.
This session presents a practical, sanitized case study of designing a governed MCP architecture that connects enterprise AI agents with systems such as Snowflake and Microsoft 365. It examines how per-user identity, API management, tool-level authorization, data-source isolation, policy enforcement, PII protection, and end-to-end auditability affect the architecture.
The talk will also discuss challenges encountered during the design process, including OAuth token propagation, separating MCP servers by security boundary, managing dependencies on bank-hardened virtual machines, and promoting containerized services across development, UAT, and production environments. Rather than presenting MCP as a simple integration layer, the session demonstrates why it must be treated as a governed control plane for enterprise AI.
Attendees will leave with a reference architecture, a practical threat-modeling framework, and an implementation checklist for building secure and auditable AI-agent integrations in regulated environments. No prior hands-on experience with MCP is required.
Speaker
Key takeaways
- Identify why standard MCP implementations can create security, identity, and governance risks in regulated enterprise environments.
- Design a reference architecture that connects AI agents to enterprise data sources through API management and isolated MCP servers.
- Apply per-user authentication, tool-level authorization, policy enforcement, PII protection, and auditable tool-call tracing.