AI Voice Agent Development in Melbourne, Australia

Health Insurance / Private Health Fund

A Melbourne-based private health insurer (A$2.1 billion premium revenue, 1.4 million members, predominantly hospital and extras cover) was processing 2.8 million inbound calls annually through a 340-seat contact centre. Call volume had grown 12% year-over-year while the ability to recruit and retain contact centre staff had deteriorated — annual agent turnover was 42%, and time-to-fill for new positions averaged 8 weeks. The contact centre was chronically understaffed: average wait times exceeded 8 minutes during peak periods (Monday mornings, claims periods, annual premium increase announcement weeks). Analysis of call types revealed that 62% of calls were routine — membership enquiries (what am I covered for, what is my excess, when does my waiting period end), claims status checks (has my claim been processed, when will I receive payment), provider search (is this dentist a preferred provider, what will I get back for this procedure), and payment management (update direct debit, request payment holiday, change payment frequency). These calls were predictable, data-lookup driven, and well-suited to AI resolution — but the existing IVR (Press 1 for Claims, Press 2 for Membership...) had a 28% abandonment rate, and members who navigated the IVR still waited for a human agent to perform what was essentially a database lookup. The Chief Operating Officer set a target: resolve 50% of routine calls through AI voice agents within 12 months, reducing wait times to under 60 seconds while maintaining member satisfaction.

We developed and deployed the AI voice agent in three phases over 16 weeks. Phase 1 (weeks 1-5): membership enquiry agent. The first voice agent handled the highest-volume call type: membership enquiries. Architecture: the caller is greeted by the AI agent ("Hello, this is your health fund's virtual assistant. I can help with membership enquiries, claims, finding providers, and payments. How can I help you today?"). The AI agent authenticates the caller using: member number (spoken or entered via keypad), date of birth, and Medicare number — matching the existing human authentication protocol. Once authenticated, the agent has real-time access to the member's policy data through API integration with the health fund's core membership system. The agent handles: coverage enquiries ("Am I covered for dental?" — the agent checks the member's specific product and responds with their actual coverage details, excesses, and any waiting periods), benefit utilisation ("How much of my dental benefit have I used this year?" — the agent checks the member's claims history and reports remaining benefits), waiting period status ("When does my waiting period for major dental end?" — the agent calculates and reports the specific date), and product comparison ("What would I be covered for if I upgraded to the next level?" — the agent compares the member's current product with the upgrade option and highlights differences). Phase 2 (weeks 4-10): claims status and provider search. Claims status agent: members call asking about claims they have submitted. The AI agent: looks up recent claims (by date range, provider name, or claim type), reports claim status (received, being assessed, approved, paid — with dates and amounts), explains any claim rejections (with the specific reason — waiting period, annual limit reached, service not covered under their product, provider gap), and initiates claims reviews when the member disagrees with a decision (collecting the member's reasons and routing to the claims review team). Provider search agent: members call to find preferred providers before treatment. The AI agent: searches the provider network by: location (near the member's registered address, near a specific suburb, or near their current location), specialty (dentist, physiotherapist, optometrist, psychologist, etc.), and gap status (no-gap providers where the member pays nothing above their excess, known-gap providers with predictable out-of-pocket costs). The agent provides: provider name, address, phone number, and whether the provider offers direct billing (the member pays nothing at the point of service — the provider bills the fund directly). Phase 3 (weeks 8-16): payment management and full deployment. Payment agent: handles direct debit updates (new bank account or credit card details — with PCI DSS tokenisation of card numbers), payment frequency changes (weekly, fortnightly, monthly, quarterly, annual), payment holiday requests (temporary suspension of cover and payments — with clear explanation of the implications for waiting periods and coverage gaps), and arrears management (members behind on payments — AI explains the situation, offers catch-up payment plans, and processes immediate payments). Full deployment: all four agent capabilities deployed behind a single entry point. The AI understands multi-intent requests ("I want to check if I am covered for physio, find a physio near work, and also update my direct debit details") and handles them sequentially within a single conversation. Escalation to human agents for: complex complaints, hardship applications, policy cancellation (retention team), and any request where the member explicitly asks for a human ("I want to speak to a person" — immediate transfer with conversation summary). Voice characteristics: the TTS voice was selected in consultation with the health fund's brand team — a warm, professional Australian female voice at a slightly slower pace than standard (recognising that many callers, particularly elderly members on hospital cover, appreciate a measured speaking pace). The AI adjusts: speaking more slowly when explaining complex coverage details, pausing to let the member process information, and offering to repeat or explain in simpler terms when it detects confusion.

The AI voice agent launched across all inbound phone lines after a 4-week pilot with 10% of call traffic. Pilot results (4 weeks, 10% traffic): 71% of routed calls resolved by the AI agent without human escalation. Average handle time: 2 minutes 14 seconds (compared to 4 minutes 38 seconds for human agents on the same call types — the AI accesses data faster and does not need to navigate multiple systems). Member satisfaction: 4.1/5 for AI-resolved calls (compared to 4.3/5 for human-resolved calls on the same call types — close enough to proceed with full rollout). Common AI resolution failures: members with complex multi-product families (joint policies with dependents on different products), edge-case coverage enquiries requiring clinical judgment, and members who simply preferred speaking to a human regardless of AI capability. Full rollout results (6 months): call containment rate: 64% of total inbound calls resolved by the AI agent (exceeding the 50% target). For routine call types: 78% containment for membership enquiries, 82% containment for claims status, 74% containment for provider search, and 61% containment for payment management. Wait time: average wait time reduced from 8+ minutes to 22 seconds (the AI agent answers immediately — the only wait is the initial telephony routing, typically 3-5 seconds, plus the greeting). During peak periods (previously 15-25 minute waits), the AI agent handles the surge — AI capacity scales instantly while human capacity is fixed. Human agent reallocation: 85 contact centre positions (25% of the workforce) were redeployed: 40 positions to a new "complex resolution" team (handling the calls the AI escalates — complaints, hardship, retention, clinical coverage queries — higher-skill, higher-satisfaction work), 25 positions to a proactive member engagement team (outbound calls for preventive health, benefit utilisation reminders, product review — work that the contact centre had never had capacity for), and 20 positions eliminated through natural attrition (not replacing departing staff). Remaining 255 human agents handle escalated calls with shorter queues, better tools (AI conversation summaries), and more complex, interesting work — agent satisfaction improved and turnover decreased from 42% to 28%. Annual financial impact: A$4.2 million in contact centre cost reduction (salary, recruitment, training, facilities), A$1.8 million in member retention improvement (members who previously abandoned calls during long waits — now immediately served by AI — had measurably higher retention rates), and A$600,000 in after-hours call handling (previously, after-hours calls went to voicemail with next-day callback — now the AI agent resolves them immediately). Total annual benefit: A$6.6 million. Implementation cost: A$420,000 (development, integration, pilot, deployment). First-year ROI: 14.7x. Ongoing cost: A$8,000/month (telephony, AI inference, maintenance). Member Net Promoter Score: overall NPS improved from +18 to +29 within 6 months — the single largest driver was the elimination of long wait times. Members specifically commented on: "I called at 11pm and got my answer immediately," "It knew exactly what product I was on and gave me a straight answer," and "It found me a no-gap dentist near my office in 20 seconds." Regulatory compliance: APHI (Australian Private Health Insurance) regulatory review confirmed the voice agent met all disclosure requirements for AI interaction. The Privacy Commissioner's office reviewed the implementation as part of a broader health sector AI assessment and noted it as a positive example of transparent AI deployment — the agent clearly identifies as AI, explains data handling, and provides easy escalation to humans.