InsurGPT is an ensemble of specialized AI models, each trained on a specific insurance document type — loss runs, SOVs, ACORDs, demand letters, medical bills, and more — and validated by a secondary verification layer. The training corpus spans 300M+ real insurance documents annotated by insurance experts.

The Workflow Canvas includes Guidewire-native nodes for Create Claim, Search Claim, and Update Claim against ClaimCenter, InsuranceSuite, and InsuranceNow. Each node performs one operation for auditability. Credentials are stored in Bevaya's secret manager and never exposed in workflow definitions.

A document arrives via the configured trigger. Document Intelligence runs OCR, classifies, and splits multi-doc submissions. InsurGPT extracts each field with a confidence score and source citation. High-confidence fields proceed automatically; low-confidence fields route to HITL review with X-Ray Mode. Output writes to your systems of record. Every step is logged.

In Azure regions per your contracted deployment region. Each tenant runs in a dedicated environment with strict logical and network isolation. Federated learning improves models without sharing data between tenants. Full details in the Trust Center.

Yes — every deployment is configured to your workflows by Bevaya's team. The Workflow Canvas supports custom branching logic, sandboxed code execution, and bespoke connectors. Custom AI Agents can be built for customer-specific document types. Your IT team doesn't write workflow code — Bevaya does.

More than a model. Production insurance AI needs insurance-trained models, workflow orchestration, a reviewer experience for confidence-scored work, audit trails, integrations to Guidewire and the rest of the insurance stack, and continuous management once it's live. Bevaya brings all of that in one platform — designed for insurance, delivered in days, by our team, with a performance commitment.

Continuous monitoring for accuracy drift across all production deployments. Updates are released by Bevaya's ML team on a managed cadence, validated against held-out evaluation sets. Workflow versioning lets you lock configurations to specific model versions and roll back when needed. Reviewer corrections feed back into training without data leaving your tenant.