Governance and Observability of Agentic AI in the Insurance Sector
When an AI agent makes the wrong decision and no one can explain why, the issue stops being an IT matter and becomes one of governance, accountability, and corporate culture. This was the central theme of the second episode of Insurance in the Mirror, Insurzine's format dedicated to insurance innovation, which also featured Davide delle Cave, S2E's Business Line Manager for Data Analytics and Observability.
Joining him on the panel were Raffaele Avila from Zurich Italy, Luca Magnoni from AXA Italy, and Alberto Dominici from Bene Assicurazioni. The discussion revolved around a single key word: observability. Yet behind that concept lies much more.
Delle Cave framed the topic by highlighting a distinction with significant practical implications: AI agents can be divided into those designed to support internal productivity and those embedded within insurance products, directly interacting with customers. For both categories, the challenge is not technical but semantic. Unlike traditional software, AI agents do not operate through deterministic logic. They reason using natural language, process context, and generate responses through chains of thought that cannot be reduced to a sequence of programmable instructions. When errors occur, they are not found in a system log but in the reconstruction of a cognitive process.
As Delle Cave explained, monitoring platforms traditionally focused on infrastructures and applications. Today, they must also monitor AI agents, and the key question has changed: no longer "Is the system working?" but rather "Is the agent reasoning in the way I intended?" This represents a major shift that requires entirely new tools, KPIs, and skill sets.