Agentic AI in Production: Architecting an Enterprise Treasury Copilot
About this session
AI demos are easy. Shipping something a CFO will actually rely on is not. I spent the last year building an agentic copilot for corporate treasury — the kind of system that has to answer "what's our cash position in EMEA right now?" without making anything up, without leaking data, and without falling over when an Azure region has a bad afternoon. This talk is about what I learned along the way. I'll walk through the real architecture: how we wire up tool-calling LLMs, why we ended up with a mix of RAG and direct database queries, how multi-region failover on Azure OpenAI actually works in practice, and what it took to plug all of this into Microsoft 365 Copilot as a declarative agent. I'll also spend time on the boring-but-critical stuff — auth, audit logging, observability — because that's usually where enterprise AI projects quietly die. If you've ever built a slick prototype and then watched it struggle the moment it met production, this one's for you. You'll leave with a blueprint you can steal, a list of traps to avoid, and a clearer sense of where today's agent frameworks help and where you're still on your own.
Speaker
Key takeaways
- A reference architecture for agentic AI in regulated enterprises — how to combine tool-calling LLMs, RAG, and structured data access without ending up with a Frankenstein you can't debug.
- Multi-region Azure OpenAI failover that actually works — the deployment topology, routing logic, and failure modes we hit in production (and the ones the docs don't warn you about).
- How to make an LLM trustworthy enough for finance — practical patterns for grounding, citations, guardrails, and the "show your work" UX that turns skeptics into power users.