Sentinel Insurance Claims Triage AI

The challenge

Before

Sentinel Insurance writes motor and property cover across Singapore, Malaysia, and Indonesia with roughly 11,000 first-notification-of-loss events a month. Their first-touch triage — assessing severity, allocating to the right adjuster track, requesting initial documentation — was taking an average of six hours during business days and far longer over weekends. Customers complained loudest about silence in the first 24 hours, and the team was losing visible NPS to faster-moving InsurTech competitors.

• 11,000 monthly first-notification-of-loss events triaged manually
• First-touch triage averaging six hours on weekdays, longer over weekends
• Severity assessment varying by adjuster experience and time of day
• Document requests sent in bulk template emails, often missing the right items
• No routing intelligence — claims piled up in a general adjuster queue
• Customer NPS losing visibly to faster-moving InsurTech competitors
• Repeat customer follow-ups consuming adjuster bandwidth before real work began
• Fraud indicators surfaced only at advanced claim stages, not at first touch

The solution

What we built

We deployed a claims triage agent that consumes the FNOL (form submission, voice transcript, photo upload, or partner API feed), assesses severity and complexity, classifies the claim track, requests the right initial documentation in the customer's language, and routes to the appropriate adjuster pod. The agent uses computer-vision models to assess motor damage from photos and applies a structured severity rubric. It flags fraud indicators (mismatched accident narratives, image metadata anomalies, repeat-claimant patterns) for human review without making accusations. The agent operates inside the insurer's claims platform, so adjusters pick up cases fully prepared rather than starting from scratch. Customers receive an acknowledgement and a clear next-step request within minutes of submission. Adjusters keep all final authority — the agent prepares the ground, humans make the calls.

How the system flows

1. FNOL Intake (form / voice / photo / partner API)→Language Detection
2. Severity Assessment→Track Classification→Adjuster Pod Routing
3. Photo Analysis→Damage Estimation→Repair vs Total Loss Indicator
4. Document Request Drafted→Customer Channel→Acknowledgement
5. Fraud Indicator Flagging→Human Review Queue→Investigation Decision
6. Repeat Claimant Pattern Detection→Special Handling Path
7. Adjuster Picks Up→Pre-prepared Case→Decision Authority Retained
8. Customer Status Updates→Channel Preference→Brand Voice Maintained

• Quality assurance sampling of agent triage decisions for retraining
• Computer-vision damage assessment with structured severity rubric