AI Agent Development in Melbourne, Australia

Health Insurance / Financial Services

A Melbourne-based private health insurer (APRA-regulated, 680,000 members, A$2.8 billion in premium revenue, processing 4.2 million claims annually) was experiencing claims processing bottlenecks. The insurer processed four claim types: medical (hospital, specialist, GP — 45% of volume), ancillary/extras (dental, optical, physio, chiro — 40%), pharmacy (PBS and non-PBS medications — 10%), and hospital excess and co-payment adjustments (5%). The 85-person claims team manually processed each claim: reading the receipt or invoice (paper, photograph, or PDF), extracting provider details (name, provider number, service type), matching to the member's policy (checking cover level, waiting periods, annual limits), calculating the benefit (applying the policy's fee schedule and Medicare benefit schedule), and processing payment (to the member's nominated bank account or to the provider). Average processing time: 8.5 minutes per claim for straightforward extras claims, 15-25 minutes for complex medical claims. During peak periods (December-January, when members rushed to use remaining annual benefits before reset), the claims backlog exceeded 15 business days — causing member dissatisfaction and complaints. Error rates averaged 3.2%: incorrect benefit calculations (applying the wrong fee schedule), policy coverage mismatches (paying for services not covered under the member's level), and duplicate payment processing. Each error required correction, member communication, and sometimes recovery of overpayments — costing the insurer an estimated A$2.4 million annually in correction and recovery costs. The Chief Operations Officer needed to reduce processing time, eliminate the seasonal backlog, and improve accuracy — without proportionally increasing headcount.

We built the claims processing AI agent over 14 weeks. Phase 1 (weeks 1-4): document intelligence and data extraction. The agent ingests claims from all submission channels: member app photo upload (the most common — photos of receipts taken on smartphones, with variable quality, angle, and lighting), email attachments (PDF invoices, scanned receipts), web portal upload, and provider-submitted electronic claims (HICAPS — the Australian health insurance electronic claiming system). Document processing: the agent reads each claim document using OCR (with models trained on Australian healthcare receipts — understanding the specific formats used by Australian dentists, optometrists, physiotherapists, GPs, and specialists), extracts: provider name and AHPRA registration number, service date, service type (mapped to the appropriate MBS — Medicare Benefits Schedule — or extras item code), charge amount, Medicare benefit amount (if applicable), and member/patient details. Validation: extracted data is validated against reference databases (AHPRA provider register — confirming the provider is registered and their profession matches the claimed service type, Medicare fee schedules, and the insurer's provider agreement database). Phase 2 (weeks 3-9): policy matching and benefit calculation. The agent accesses the member's policy details from the insurer's policy administration system: cover level (basic, mid, top), waiting period status (each service type has a specific waiting period — the agent checks whether the waiting period has been served), annual limits (remaining benefit available for each extras category — dental, optical, physio, etc.), sub-limits (some policies have per-service limits within the overall category limit), preferred provider status (higher benefits for services from the insurer's preferred provider network), and excess and co-payment (for hospital claims — the agent calculates the excess payable and co-payment amounts). Benefit calculation: the agent applies the policy's benefit schedule — calculating the exact payment amount based on the charge, the fee schedule, the Medicare benefit (for medical claims), and the member's remaining annual limit. For extras claims: the calculation is typically a percentage of the charge up to the annual limit (e.g., 60% of dental up to A$600 per year). For medical claims: the calculation involves the Medicare Benefits Schedule, the insurer's gap cover policy, and the provider's agreement status. Decision engine: the agent determines the claim outcome — approved (payment processed), partially approved (benefit calculated but charge exceeds the schedule or limit), declined (service not covered, waiting period not served, limit exhausted), or escalated (complex scenarios requiring human review — unusual charges, potential fraud indicators, new provider types not in the training data). Phase 3 (weeks 7-12): integration and deployment. The agent was integrated with: the claims management system (creating claim records, updating status, recording decisions), the payment system (initiating member payments for approved claims — next-day bank transfer), the member communication system (sending claim outcome notifications via email and app push notification — with clear explanations of how the benefit was calculated), and the fraud detection system (flagging claims with potential fraud indicators — duplicate services, unusual provider patterns, charge outliers). Human oversight: the agent processes claims autonomously for straightforward extras claims below A$2,000 (representing 65% of total claim volume). Claims above A$2,000, medical/hospital claims, and all escalated claims are reviewed by human assessors — with the agent providing a pre-assessed recommendation and supporting information that reduces human assessment time. Quality assurance: 10% of all autonomously processed claims are randomly selected for human review — providing ongoing accuracy monitoring and training data for continuous improvement. Phase 4 (weeks 10-14): optimization and scaling. The agent was trained on edge cases identified during the initial deployment: multi-service receipts (a single receipt listing multiple dental services — the agent splits and processes each service separately), handwritten receipts (still common from some providers — the OCR model was fine-tuned on handwritten Australian healthcare receipts), and partially illegible documents (the agent requests a resubmission rather than guessing — maintaining accuracy over speed).

The agent went live with phased rollout over 6 weeks, starting with extras claims (the highest volume, most standardised claim type). Processing time: extras claims processed in an average of 45 seconds (from 8.5 minutes — 91% reduction). The agent processes the claim, calculates the benefit, initiates payment, and sends the member notification — all within a minute of submission. Medical claims: pre-assessed by the agent in 2 minutes (from 15-25 minutes of manual assessment) — the human assessor then reviews the agent's recommendation and approves or modifies, taking 3-5 minutes instead of 15-25. Total medical claim processing: 5-7 minutes. Seasonal backlog: eliminated. During the December-January peak, the agent processed the volume increase without degradation — there is no backlog when claims are processed in seconds. Members who previously waited 15+ business days now receive their benefit within 1 business day (same-day for digital claims submitted before 2pm). Accuracy: error rate reduced from 3.2% to 0.8% (75% improvement). The agent applies benefit schedules and policy rules consistently — it does not make the fatigue-driven errors or fee schedule lookup mistakes that human assessors occasionally make. The remaining 0.8% errors are primarily: edge cases not covered by the agent's training data (handled through the ongoing quality assurance process), and data extraction errors from poor-quality document images (the agent now requests resubmission for low-confidence extractions). Member satisfaction: NPS improved from +24 to +48 within 6 months. The single largest satisfaction driver: payment speed. Members receiving their benefit within 24 hours of claim submission (from 5-15 business days) described the experience as "transformed." Claims team: the 85-person team was restructured to 52 assessors (handling complex medical claims, escalations, and quality assurance) and 8 agent operations staff (monitoring agent performance, handling edge cases, and managing continuous improvement). 25 positions eliminated through natural attrition over 12 months. The remaining team members report higher job satisfaction — handling interesting complex cases rather than repetitive extras claims processing. Annual savings: A$4.2 million (team restructuring, error correction elimination, operational efficiency). Member retention: the claims experience improvement contributed to a 1.8 percentage point improvement in member retention — estimated value A$12 million in retained premium (680,000 members × A$4,100 average premium × 1.8% retention improvement ÷ expected member lifetime). Implementation cost: A$420,000. Annual operating cost: A$85,000 (infrastructure + model API costs). First-year ROI: 10x on direct operational savings.