Google and CoverGo have published a new whitepaper on how insurers could move from isolated AI projects to an operating model built around software agents, authored by Maxim Afanasyev of Google and Tomas Holub of CoverGo, looks at what changes when AI agents start working across distribution, underwriting, claims and servicing rather than inside single use cases.
The paper proposes an Agentic Insurance Operating System where specialised agents access insurance data and tools, interact with core systems, coordinate with other agents and escalate decisions when human judgement is required.
That shift creates new problems around agent authority, conflicting objectives, data access and oversight. An underwriting agent, claims agent and servicing agent might each optimise for different outcomes, forcing insurers to define who controls decisions and when people step in.
Drawing on perspectives from Google and CoverGo across AI, technology and insurance modernization, the paper argues that the impact of Agentic AI could extend far beyond automating individual tasks or improving productivity.
As AI agents increasingly operate across distribution, underwriting, claims and servicing, insurers will need to consider how their businesses, operating models and technology foundations evolve with them.
At the center of the paper is the concept of an Agentic Insurance Operating System: an environment in which specialized AI agents can access insurance context and tools, collaborate with other agents, interact with core systems and workflows, and involve people when human judgment is required.
The whitepaper examines the new challenges this creates, from agent conflicts and authority boundaries to data readiness, governance and human oversight.
Data is another constraint. The authors point to vector databases, structured extraction pipelines and master data management as necessary infrastructure for agents working across large volumes of policy, claims and customer information.
Human control remains part of the model. The paper proposes confidence scoring, mandatory review for higher-risk decisions, audit trails and feedback from human corrections.
Google and CoverGo also include a reference architecture and a five-step modernisation approach for insurers moving from separate AI initiatives toward coordinated agent-based operations.
Autonomous systems making decisions that affect customers need to remain auditable, explainable and interruptible. The authors refer to IAIS guidance from 2024 when discussing this requirement.
The proposed controls include confidence scoring, where low-confidence outputs are sent to employees for review. High-value claims, unusual underwriting profiles and regulatory exceptions also pass through mandatory human approval.
Audit trails record what an agent decided, its confidence level and whether a person later changed the recommendation. Human corrections then feed back into future system performance. This structure gives insurers a way to automate more work without giving agents unrestricted decision-making authority.
Alongside its reference architecture, the whitepaper presents a five-step approach for insurers moving toward agent-based operations. The direction is different from the current pattern of buying or building AI applications for separate departments.
Google and CoverGo are instead describing an insurance environment where agents operate across functions, access shared data and work within defined authority limits.
