Background
A national general-insurance carrier was facing a familiar pressure point: first-line claims triage consumed a disproportionate share of agent time, most of it spent on deterministic-but-tedious clarifications, document checks, and policy lookups. The carrier wanted to know whether an agentic AI assistant could compress the triage step without opening compliance holes that the regulator and the internal audit function would not accept.
Challenge
- Compliance posture must be preserved. Every model output and every tool call had to be traceable to an authenticated identity and persisted in an audit trail.
- No bypass paths. The agent had to authenticate against the same identity provider as human agents, with the same SCIM lifecycle.
- Inspectable behaviour. Reviewers needed to see exactly what the agent saw, what it recommended, and what changed in the case record.
- Rollback at every step. If the assistant produced an unacceptable recommendation, the case had to revert to a fully-human flow without data loss.
Approach
The agentic assistant was built on OpenShift AI, with NVIDIA NIM and vLLM serving the inference layer behind the same OpenShift fleet that hosts the carrier’s other production workloads. Tool integration was done via MCP (Model Context Protocol) servers — each tool (policy lookup, claims-history retrieval, document classifier, mail send) is a separately-deployed MCP server with its own audit log.
The identity boundary is WSO2 Identity Server: the agent runs under a service identity provisioned via SCIM from the same directory of record as human agents, with OAuth2 token exchange for downstream tool calls. The agent cannot reach any backend system except through the IS-mediated boundary.
Search across policies and prior claims uses OpenSearch, with results filtered by the agent’s session identity. State for in-flight conversations is held in Redis with a short TTL and full audit-event emission.
Outcome
- Production agentic assistant integrated into the carrier’s claims front-end
- Measurable reduction in first-touch handling time on the cases routed through the assistant
- Full audit trail accepted by the carrier’s internal audit and compliance functions
- Documented rollback path tested in production and in DR drills
Engagement shape
Approximately 14 weeks, sequenced across identity integration, MCP tool wiring, inference platform tuning, audit and compliance review, and production rollout under change-management discipline.