Multi-Agent Orchestration in Singapore

Insurance / General Insurance

A Singapore general insurer (MAS-licensed, 890,000 policyholders, processing 145,000 claims annually across motor, property, travel, and health lines) operated a claims processing workflow that was sequential, manual, and slow. A typical motor insurance claim: notification (customer calling or emailing to report the claim — the call centre agent recording details in the claims system), assignment (claims manager reviewing the notification and assigning to a claims adjuster based on claim type and workload), investigation (the adjuster reviewing the claim — checking policy coverage, examining photos/documents, potentially arranging vehicle inspection, contacting witnesses for complex claims), assessment (the adjuster determining the claim amount — repair costs, medical costs, third-party liability — and checking against policy limits, excesses, and exclusions), approval (claims above SGD 5,000 requiring senior adjuster approval, above SGD 25,000 requiring claims manager approval, above SGD 100,000 requiring head of claims approval), and payment (approved claims processed for payment — cheque or bank transfer to the policyholder or directly to the repair shop/hospital). Processing time: average 18 working days for motor claims, 23 days for property claims, 12 days for travel claims. Customer satisfaction: NPS 28 — significantly below industry benchmark of 42. The specific problems: sequential processing (each step waiting for the previous to complete — even when steps could run in parallel. Policy verification, coverage check, and fraud screening could all happen simultaneously, but the manual process handled them sequentially. Adding 5-7 days of queue time where the claim sat waiting for the next step), adjuster bottleneck (52 claims adjusters handling 145,000 claims annually — approximately 2,800 claims per adjuster per year or 11 per working day. Adjusters spending 40% of time on straightforward claims (small motor repairs, standard travel claims, simple property damage) that didn't require expert judgement — just verification that the claim was valid, covered, and within normal parameters. The remaining 60% on complex claims where their expertise was genuinely needed. No mechanism to automatically route simple claims through an expedited path), fraud detection gaps (fraud estimated at 8-12% of claims by value — but the current process had no systematic fraud detection. Individual adjusters might notice suspicious patterns based on experience, but there was no data-driven screening. The insurer's fraud investigation unit was reactive — investigating only claims that adjusters specifically referred, missing systemic fraud patterns), and document processing delays (each claim averaging 4.3 documents (accident reports, police reports, medical bills, repair estimates, photos). Documents arriving as emails, WhatsApp photos, and physical mail — the operations team manually sorting, digitising, and attaching documents to the correct claim record. Missing documents: the single largest cause of processing delay — 34% of claims stalled at least once waiting for additional documents from the customer).

We deployed a multi-agent claims processing system over 16 weeks: Agent architecture (weeks 1-6): 7 specialised agents coordinated by an orchestration engine. Document Processing Agent: AI-powered document analysis. When claim documents arrive (email, WhatsApp, portal upload), the agent: classifies the document type (accident report, police report, medical bill, repair estimate, photo evidence), extracts structured data (claim amounts, dates, parties involved, vehicle details, injury descriptions), validates completeness (checking whether all required documents for the claim type are present — motor claims requiring accident report + photos + repair estimate at minimum), and flags quality issues (blurry photos, incomplete forms, missing pages). OCR accuracy: 95%+ on printed documents, 88% on handwritten (Singapore accident reports often partially handwritten). Processing time: 30 seconds per document (vs 15 minutes manual). Coverage Verification Agent: instantly checking the claim against policy terms. When triggered by the orchestrator, the agent: verifies the policy is active (premium paid, no lapse), confirms the incident type is covered (checking policy schedule, endorsements, and exclusions), identifies the applicable excess/deductible (standard excess, voluntary excess, named driver excess), calculates the maximum claimable amount (sum insured minus previous claims in the policy period), and flags any coverage concerns (late notification, excluded drivers, excluded territories). Processing time: 2 seconds (vs 30 minutes manual — adjusters previously reading through policy documents to verify coverage). Fraud Detection Agent: pattern-based fraud screening. Every claim automatically screened for: claim frequency (multiple claims from the same policyholder, address, or vehicle in a short period), timing patterns (claims filed immediately before or after policy changes — common in staged claims), amount patterns (claims consistently just below the threshold requiring additional documentation or approval), network analysis (connections between claimants, repair shops, medical providers, and witnesses — identifying fraud rings), and document anomalies (metadata analysis of photos — creation date, device, GPS location compared to claimed accident details). Risk scoring: each claim scored Low (0-30), Medium (31-60), High (61-80), or Critical (81-100). Low-risk claims: proceed without delay. Medium: flagged for adjuster awareness. High/Critical: referred to the fraud investigation unit with detailed analysis. Assessment Agent: calculating claim amounts for straightforward claims. For motor claims: repair cost assessment using a database of standard repair costs by vehicle make/model/year and damage type — the agent comparing the submitted repair estimate against market rates and flagging significant deviations. For travel claims: flight delay compensation calculated automatically from flight data (FlightAware API), medical expense assessment based on submitted bills with standard rate comparison, and lost baggage compensation per policy terms. For simple property claims: damage assessment based on submitted photos and descriptions, compared against standard repair/replacement costs. The assessment agent handling "simple" claims end-to-end — the approximately 40% of claims that are straightforward with clear coverage, reasonable amounts, and no fraud indicators. Complex claims: routed to human adjusters with the agent's preliminary assessment as a starting point. Approval Agent: automated approval routing. Claims within the assessment agent's confidence threshold and below SGD 5,000: auto-approved (with audit logging). Claims SGD 5,000-25,000: routed to senior adjuster with agent assessment and recommendation. Claims above SGD 25,000: routed to claims manager. The approval agent: tracking approval status, sending reminders for pending approvals, and escalating if approvals exceed SLA. Communication Agent: proactive customer updates. WhatsApp integration: the customer receiving automated updates at every milestone — "Your claim CL-2026-78432 has been received. We have your accident report and photos. We still need the repair estimate from the workshop — please submit when available." When documents are missing: specific, clear requests — not generic "please submit outstanding documents" but "We need the repair estimate from the workshop. You can take a photo and send it here on WhatsApp." Status enquiries: the customer asking "What's happening with my claim?" and receiving an immediate, accurate status update — no waiting for a call centre agent to look up the information. Payment Agent: processing approved claim payments. Bank transfer via FAST for immediate payment (small claims under SGD 1,000 — paid within hours of approval). GIRO for larger amounts. Direct payment to repair shops and medical providers where the policyholder has authorised. Payment confirmation sent to the customer via WhatsApp. Orchestration Engine: coordinating all agents through the claims lifecycle. When a new claim is filed, the orchestrator launches parallel processing: Document Processing Agent starts immediately on submitted documents. Coverage Verification Agent runs simultaneously (using claim notification data — doesn't need full documents). Fraud Detection Agent begins screening with available data (enhanced as documents are processed). As each agent completes, the orchestrator evaluates: all documents received + coverage verified + fraud score low → trigger Assessment Agent. Assessment complete + amount within auto-approval threshold → trigger Approval Agent (auto-approve). Approval complete → trigger Payment Agent + Communication Agent (payment initiated + customer notified). The entire flow for a simple claim: potentially completing in hours rather than weeks.

Results after 6 months: Processing time: simple claims (40% of total — clear coverage, low fraud risk, below SGD 5,000): average 4.2 hours (from 18 days for motor). These claims processed end-to-end by the multi-agent system with no human intervention. Medium-complexity claims (35% — requiring some adjuster review but with agent pre-assessment): average 5.3 working days (from 18 days). Agents handling document processing, coverage verification, fraud screening, and preliminary assessment — the adjuster reviewing and confirming rather than building the assessment from scratch. Complex claims (25% — high value, coverage disputes, complex liability, fraud referrals): average 14 working days (from 18 days). Agents still handling document processing and preliminary analysis — but human expertise driving the investigation and decision. Overall average: from 18 working days to 6.8 working days (62% reduction). Fraud detection: fraud detection rate improved from estimated 35% (only cases noticed by individual adjusters) to 78% (systematic AI screening of every claim). Claims referred to fraud investigation: increased from 120/year to 340/year (more detection, not more fraud). Confirmed fraud savings: SGD 4.2M in the first 12 months (claims that would have been paid without the fraud detection agent). False positive rate: 8% (92% of fraud referrals confirmed as genuine concerns — acceptable given the financial impact of missed fraud). Customer satisfaction: NPS from 28 to 61 (above the 42 industry benchmark). Primary drivers: faster processing (simple claims resolved in hours), proactive communication (WhatsApp updates at every step — customers knowing exactly what's happening without calling), and document handling (clear, specific requests for missing documents — not generic follow-ups). Customer effort: the number of times a customer had to contact the insurer about a single claim reduced from 3.2 (average — multiple calls to check status, respond to document requests, chase payment) to 0.8 (most customers receiving proactive updates and payment without needing to initiate contact). Adjuster productivity: the 52 adjusters now focusing on complex claims requiring expert judgement. Simple claims: handled entirely by agents. Medium claims: adjusters reviewing agent pre-assessments rather than building from scratch — 60% time saving per claim. Adjuster capacity: effectively doubled — handling the same claim volume in half the time, with additional capacity for quality improvement on complex claims. Financial impact: fraud savings: SGD 4.2M/year. Adjuster productivity (equivalent of 20 additional adjusters at SGD 85,000 fully loaded): SGD 1.7M/year. Operations team reduction (document processing automated): SGD 480,000/year. Total annual benefit: SGD 6.38M. Implementation cost: SGD 1.2M. Annual operating cost: SGD 180,000. First-year ROI: 432%.