Download Responsible, Ethical and Trustworthy AI Principles
Download Roots' Responsible, Ethical and Trustworthy AI Principles
Built for Insurance
Responsible AI
Insurance runs on documentation and defensibility. So does our platform. Every Bevaya extraction is sourced, scored, reviewable, and recorded.
Human-in-the-loop
Humans stay in control of what matters.
Our patented Human-in-the-Loop is the governance layer behind every AI Agent. Uncertain extractions and policy-sensitive items route to your experts before anything touches your downstream systems.
- Configurable thresholds. Per field, per flow. You decide what "confident enough" means for each decision.
- No silent automation on low-confidence data. The platform routes, it doesn't guess. Anything below threshold lands in a reviewer's queue with the source passage attached.
- Every correction trains the model. Your accuracy compounds, supervised by your domain experts — not a vendor's labelers.
- Full audit trail. Who reviewed what, when, and why is logged on every record. Defensible for regulators, reinsurers, and internal QA.
Visual flow building
Every automation, visible at a glance.
Drag, drop, connect. Complex insurance workflows — branching submissions intake, multi-document loss runs, FNOL setup, endorsement processing — rendered as a single canvas you can read end to end. Bevaya's team builds and maintains your flows here. You see exactly what the AI does, and where.
- No-code by design. Branching, looping, and routing are all configured visually. No engineering tickets to change a threshold.
- Pre-built insurance templates. Start from FNOL, Underwriting Submissions, SOV, Loss Runs, and more — tuned and ready, not built from scratch.
- InsurGPT™ on the canvas. Every extraction, classification, and reasoning node uses the specialized insurance model best suited to the task.
- Real-time testing. Run a draft against a sample document and see every node's output before publishing.
Decision trail
Built for the audit you haven't been asked for yet.
Regulators, reinsurers, and internal audit want to know how a decision was made. Bevaya captures the trail as it happens, not as a quarterly report.
- Immutable per-item logs. Original AI output, reviewer, correction, timestamp, and confidence — captured on every record and locked once written.
- Versioned everything. Flow version, model version, and training-data version pinned to each run, so a decision from last March can always be traced to the exact system that made it.
- Reproducible decisions. Rerun any historical item against its original model and flow to defend a past outcome — to a regulator, a reinsurer, or your own model risk team.
- Exportable on demand. One-click pulls for compliance reviews, internal audit, MRM, and carrier diligence. No engineering ticket, no CSV stitching.
Behind the platform
How we build and govern the models.
Responsible AI isn't a screen in the product. It's the engineering discipline behind it.
Trained on insurance | Built for the work
InsurGPT™ was trained on 300M+ non-public insurance documents, annotated by dedicated insurance domain experts we call model tutors. The model learns the work because the data is the work.
- 300M+ non-public insurance documents in the training set
- Annotated by dedicated insurance domain experts
- Model tutors with real claims and underwriting experience
- No public scrape, no generic web content
- Domain-specific vocabulary baked in from day one
- Trained on the artifacts your teams actually handle
Bias-tested | Pre-production review
Models are evaluated for accuracy drift across document type, carrier, geography, and line of business before they reach production. Decisions are anchored to document content, never inferred from protected characteristics.
- Accuracy drift evaluated across document type and carrier
- Geography and line-of-business slices tested separately
- Decisions anchored to document content, not inferences
- Protected characteristics never used as signal
- Methodology documentable for regulator review
- Customer model-risk teams can audit the evaluation
Controlled rollouts | No surprise swaps
New model versions release through A/B testing in production with confidence-band monitoring. If a new version performs worse on your data, traffic doesn't shift — and older decisions stay reproducible.
- A/B testing in production with confidence-band monitoring
- Traffic only shifts when the new version wins on your data
- Prior version stays pinned to historical runs
- Older decisions remain reproducible after upgrades
- Version changes are explicit, never silent
- Rollback available if regression is detected post-release
Your data, your AI | Tenant isolation
Customer data is never used to train models for other customers. Tenant isolation is architectural, not a policy line — your corrections improve your AI Agents, on your tenant, isolated from your competitors.
- Customer data never trains models for other customers
- Tenant isolation is architectural, not just policy
- Your corrections improve your AI Agents, not someone else's
- Per-tenant model fine-tuning stays inside your tenant
- No cross-customer learning, ever
- Competitive moat: your edits become your advantage
Behind the platform
How we build and govern the models.
Read across every source in the claim file, not one document at a time.
Triangulate facts between medical records, bills, demands, and correspondence.
Produce role-specific summaries for adjuster, supervisor, attorney, and SIU workflows.
Reconstruct timelines from documents that arrive piecemeal over months.
Flag coverage triggers, statutes of limitations, and missed deadlines.
Surface subrogation and recovery opportunities by connecting facts across sources.
Generate brief, medium, or detailed summaries on demand.
Refresh the summary automatically as new documents arrive.
Show grounding for every claim with X-Ray Mode — every fact traceable to its source page.
Operate inside Guidewire, Salesforce, Duck Creek, or your existing claims system.
Surface what matters across the file so adjusters move on the right work first.
Stay audit-ready with every fact cited, sourced, and reproducible.
More Capabilities
Explore the rest of the platform.
Designed, deployed, and governed together. Powered by InsurGPT™ and accessed through the AI Assistant.
Workflow Canvas
Visual builder and production runtime for every automation.
Current page ReviewHuman-in-the-Loop
Configurable review queues with X-Ray verification and a patented feedback loop.
Current page DocumentsDocument Intelligence
Read any insurance document — hundreds of carrier formats, scanned or digital.
Current page GroundingGrounded Explainability
Every value traceable to its source. X-Ray Highlight Mode brings citations to reviewers.
Current page AnalyticsAnalytics Dashboard
Live accuracy, STP rates, reviewer SLA, and agent performance across every workflow.
Current page GovernanceGoverned Automation
Immutable audit trails, role-based access, flow versioning. Compliance is the architecture.
Current pageFAQ
Questions buyers ask us.
Every decision carries a full trail. The source document and exact location of the data, the model version that extracted it, the confidence score at the time, the reviewer who confirmed or corrected it, and a timestamp on each event. That trail is immutable, queryable, and exportable. If a regulator asks how you arrived at a decision, you answer with evidence.
The platform doesn't pass low-confidence data through. When InsurGPT™ is uncertain, the work item routes to a human reviewer before it touches your claim system, policy admin, or downstream workflow. Confidence scores are calibrated. When the system says 90% sure, it's right 90% of the time.
Bevaya models extract from documents. They don't make underwriting or coverage decisions. We test for accuracy drift across document type, carrier, geography, and line of business before any model reaches production. The methodology and test sets are documentable for regulator review and customer model-risk-management programs.
No. Customer data is never used to train models for other customers, and tenant isolation is enforced at the architecture level. Your corrections improve your AI Agents on your tenant. Our shared base models are improved through controlled training cycles on our own non-public training corpus, never on customer PII or proprietary content without explicit contractual permission.
Bevaya is built for an evolving regulatory environment. The documentation regulators are starting to require, including model behavior records, source traceability, human oversight evidence, and reweighting documentation, already exists in our architecture. As state, federal, and international standards evolve, we adapt the platform without making customers retrofit governance after the fact.
A named team. Every Bevaya customer has a dedicated continuous improvement team assigned permanently, not just at launch. Internally, Responsible AI is governed by our research and engineering leadership in partnership with an Industry Advisory Board of insurance and AI leaders.
Yes. Model versions, flow versions, and source documents are pinned to every historical run. You can rerun any item against its original model and flow to validate or defend a past outcome for internal review, customer disputes, or regulator inquiries.
GET STARTED
Ready to design, deploy, and govern your AI workforce?
Bevaya AI Agents can help you triage, analyze, and recommend across underwriting, claims, and policy servicing. Let's connect and show you how it works.