An agentic claims desk. One long-lived agent takes a claim end to end: it verifies the policyholder, assesses the loss, grounds coverage against the policy wording, scores fraud, screens for duty of care, then authorises, pays, audits and closes the settlement. What happens next is decided by the claims service in code, never by the model, and every control (coverage, the fraud referral gate, duty of care, the settlement authority cap, audit parity) is enforced by the service itself, so the claim parks for a human whenever a control says it must. The agent runs inside an NVIDIA OpenShell sandbox that can reach the claims system and nothing else, and holds no credentials of its own. A policyholder can start the whole thing by talking: they report the loss to a voice concierge, or to an avatar, from a phone or the insurer's website, and the call itself opens the claim the agents then work. Built on NVIDIA NeMoClaw, OpenShell, Nemotron 3.5 ASR, Magpie TTS, Audio2Face and a self-hosted Nemotron NIM.

AI authoring studio for trainers: state a brief, watch AgentX draft an editable curriculum, review content and quizzes, preview a design-quality deck, package SCORM and publish to an LMS, then let learners rehearse it against a spoken roleplay character.

Agent assist for a live human-to-human call. The companion listens, separates officer from caller in real time, tracks which concerns have been dealt with, and pushes cited answers to the officer’s screen mid-call. When the knowledge base does not settle a question it says so rather than guessing, and every one of those gaps is reported back so the guidance can be fixed. Officers can also practise against a synthetic caller. The AI never speaks to a real citizen.