AI Operator Console

Regulated AI agents in banking and insurance can't just act, every decision has to be explainable, auditable, and confirmable by a human when needed. This console is built as the middle path between a black-box chat and a raw trace dump: a conversation view, a structured reasoning panel, and a scrubbable trace rail, with every step carrying its own confidence score, the SOP section it cited, the tools it called with their structured arguments and results, and a before/after state diff.

It ships with 5 realistic scenarios covering the kind of decisions that actually need this scrutiny: unfreezing an account after a geo-anomaly fraud alert, a $75k cross-border wire under enhanced due diligence, an auto-loan hardship adjustment, an insurance claim SLA follow-up, and a chargeback dispute with a provisional credit. Each trace supports live replay (play, pause, step forward, reset) and exports as audit JSON a compliance reviewer can actually open and read. It's a frontend prototype by design, React 19 and TypeScript strict mode, all mock data local and deterministic, no backend, built to demonstrate the UI layer's information architecture rather than a full agent integration.