The enterprise AI landscape is littered with impressive demos that never make it past the proof-of-concept stage. While the tooling around large language models has matured rapidly, the gap between a working notebook and a reliable, governed production system remains one of the biggest challenges facing analytics teams today.

This talk draws on practical experience building and deploying agentic AI frameworks across financial services, healthcare and public sector clients. Using real-world examples from Rodan Analytics' Eclipse framework, a Python-based agentic architecture designed for enterprise deployment. I will walk through the key design decisions, failure modes and hard-won lessons from putting LLM-powered systems into the hands of non-technical users at scale.

Topics covered will include how to structure agentic workflows that balance autonomy with human oversight, practical approaches to grounding LLM outputs in organisational data without compromising governance requirements and strategies for managing the messy realities of enterprise integration, from legacy data warehouses to inconsistent APIs.

I will also share findings from deploying LLM-enabled business intelligence tooling across multiple client environments, exploring what works when you move beyond single-prompt interactions towards multi-step reasoning chains that need to be auditable, reproducible and explainable to stakeholders who do not care about your model architecture.

The talk is aimed at data scientists, analytics leads and engineering managers who are past the excitement phase of generative AI and are now grappling with the practical question of how to ship reliable AI products within organisations that have real compliance, security and governance constraints.

Attendees will leave with a concrete framework for evaluating agentic AI readiness within their own organisations, along with practical patterns for building systems that survive contact with production workloads.