AI Consulting in Sydney, Australia

Insurance / Financial Services / AI / Sydney

A Sydney insurance group (AUD 2.1B gross written premium — general insurance, life insurance, and health insurance, 2,200 employees, 1.8 million policyholders, regulated by APRA) needed AI consulting to develop and execute an enterprise AI strategy. (1) Competitive: competitors (IAG, Suncorp, QBE) visibly investing in AI — automated claims, AI underwriting, and personalised pricing becoming table stakes. Board: board requesting AI strategy — "What is our AI roadmap and how do we ensure we're not falling behind?" was a direct board question to the CTO. Fragmented: 12 separate AI/ML initiatives across the group — data science team in pricing, separate chatbot project in customer service, ad-hoc ML work in claims — no coordination, no shared infrastructure, no governance. (2) Data: significant data assets but poor AI readiness. Silos: data across 8 core systems — policy administration, claims, underwriting, CRM, finance, actuarial, digital, and call centre — with limited integration. Quality: data quality inconsistent — customer records duplicated (estimated 15% duplicate rate), claims data missing key fields (12% incomplete), and no enterprise data dictionary. Privacy: 1.8 million policyholders' personal data (health, financial, claims history) — Privacy Act compliance for AI use requiring careful consent and purpose management. Historical: 15 years of claims data (4.2 million claims) — valuable for AI training but requiring cleansing and governance. (3) Regulation: APRA regulated entity with increasing AI scrutiny. CPS 230: CPS 230 (Operational Resilience) requiring identification and management of AI-related operational risks — the group having no AI-specific risk framework. Model risk: no model risk management framework — ML models in pricing deployed without formal validation, monitoring, or governance. Explainability: regulator expecting the group to explain AI-driven decisions — particularly for claims decisions and pricing, where policyholders may challenge outcomes. (4) Talent: limited AI capability. Team: 8 data scientists across the group — spread thinly, working in silos, using different tools and approaches. MLOps: no MLOps infrastructure — models developed in Jupyter notebooks, manually deployed, and not systematically monitored. Training: broader workforce having limited AI literacy — underwriters, claims assessors, and customer service staff unsure how AI would affect their roles. (5) Trust: organisational trust in AI low. Previous: previous ML project in claims (automated liability assessment) produced biased outcomes — higher denial rates for certain postcode groups correlating with socioeconomic disadvantage. This was caught internally but damaged trust in AI within the organisation. Fairness: no systematic fairness testing — the biased claims model had been deployed without checking for disparate impact. Fear: staff concern about AI replacing jobs — particularly among claims assessors and customer service agents.

We delivered a comprehensive AI consulting engagement over 16 weeks — strategy, governance, and pilot implementation. (1) AI readiness assessment (weeks 1-4): understanding the landscape. Assessment: assessed data maturity across 8 core systems — documenting data quality, availability, integration, and governance for AI readiness. Use cases: identified and prioritised 28 potential AI use cases across claims (9), underwriting (6), customer (5), pricing (4), operations (3), and fraud (1). Scoring: each use case scored on business impact (AUD value), feasibility (data readiness, technical complexity), risk (regulatory, ethical, operational), and strategic alignment. Data: data quality remediation plan — addressing customer deduplication, claims data completeness, and data integration requirements. Top 5: top 5 prioritised use cases: (1) claims triage automation (AUD 12M annual impact), (2) fraud detection enhancement (AUD 8M), (3) underwriting risk scoring (AUD 6M), (4) customer churn prediction (AUD 4M), (5) document extraction for claims (AUD 3M). (2) AI strategy and governance (weeks 4-8): building the framework. Strategy: enterprise AI strategy document — 3-year roadmap with phased implementation, investment requirements, and expected returns. Total: total AI opportunity quantified at AUD 45M annual impact across 28 use cases over 3 years. Governance: AI governance framework designed — AI Ethics Board (comprising CRO, CTO, Chief Actuary, and external ethics adviser), AI risk management policy, and model lifecycle management process. APRA: APRA-aligned model risk management — model validation requirements, ongoing monitoring standards, and documentation requirements meeting CPS 230 expectations. Ethics: responsible AI policy — 8 organisational AI principles aligned with Australian AI Ethics Framework, with specific implementation guidance per principle. Fairness: fairness testing framework — mandatory fairness assessment for all AI models, testing across protected attributes (age, gender, postcode/socioeconomic, disability status, cultural background). Transparency: model transparency requirements — explainability standard defined per use case risk level (high-risk decisions requiring individual-level explanations, lower-risk requiring aggregate explanations). (3) AI infrastructure (weeks 6-10): building the platform. Platform: enterprise AI platform on AWS — SageMaker for ML training, Step Functions for ML pipelines, S3 data lake, and Redshift for analytics. MLOps: MLOps pipeline — automated model training, testing (including fairness testing), deployment, and monitoring. Feature: feature store — centralised feature engineering enabling reuse across AI use cases. Monitoring: model monitoring — automated drift detection, performance tracking, and alerting. Data: data integration — connecting 8 source systems to central data platform for AI use. Privacy: privacy-preserving AI — data masking, purpose-based access controls, and consent management integrated into AI data pipeline. (4) Pilot implementation (weeks 8-16): claims triage automation. Triage: AI claims triage model — automatically categorising incoming claims by complexity (simple, standard, complex), expected cost, and fraud risk. Training: model trained on 4.2 million historical claims — learning patterns of claim complexity, settlement amount, and processing time from 15 years of data. Routing: automated routing — simple claims (estimated 35% of volume) routed for straight-through processing, complex claims routed to senior assessors, fraud-flagged claims routed to investigation. Fairness: fairness testing — model tested across postcode groups, age, gender, and cultural background. Initial model showed 8% higher complexity classification for certain regional postcodes — model retrained with postcode-related features excluded, fairness testing passed. Explainability: claims triage explanations — each triage decision accompanied by top contributing factors (claim type, amount, claimant history, policy type) visible to claims assessor. Human: human-in-the-loop — AI providing recommendation, human assessor confirming or overriding with reason captured. Override rate monitored as model quality metric. Integration: integrated with claims management system — triage decisions flowing into existing claims workflow, assessors seeing AI recommendation alongside claim details. (5) Change management (throughout): building AI confidence. Communication: AI communication programme — explaining AI strategy to all 2,200 employees, addressing job impact concerns honestly. Claims: claims team workshops — hands-on sessions with claims assessors showing how AI triage works, how to interpret recommendations, and how to override with feedback. Training: AI literacy programme — 4-hour AI fundamentals training for all managers, 2-day technical training for data team, ongoing monthly AI updates. Ethics: ethics training — responsible AI training for all staff involved in AI development and deployment. Success: success stories — celebrating early wins and sharing positive outcomes from AI pilot with broader organisation.

Results over 12 months (strategy delivery + pilot outcomes). Strategy: enterprise AI strategy adopted by board — 3-year roadmap with AUD 15M investment approved. Governance: AI governance framework operational — AI Ethics Board meeting quarterly, model risk management process in place. APRA: APRA satisfied — CPS 230 self-assessment showing AI risk management as "effective" (upgraded from "needs improvement"). Fairness: fairness testing preventing biased deployment — claims triage model's postcode bias caught and corrected before production deployment. Infrastructure: enterprise AI platform operational — shared ML infrastructure enabling faster AI development across the group. Claims triage: claims triage pilot results (6 months). Volume: 42,000 claims processed through AI triage in 6 months — 35% classified as simple (straight-through), 50% standard, 15% complex/fraud-flagged. Accuracy: triage accuracy 91% — human assessors agreeing with AI classification 91% of time. Time: claims processing time for simple claims reduced from 8 days to 2 days — straight-through processing eliminating manual assessment for straightforward claims. Cost: claims processing cost per claim reduced 28% for triaged claims — automation and correct routing reducing handling time. Fraud: fraud detection improved — AI-flagged claims converting to confirmed fraud at 3.2x the rate of random sampling. Override: assessor override rate 9% — indicating high model quality while maintaining human control. Satisfaction: claimant satisfaction improved — faster processing for simple claims improving NPS from +12 to +24 for claims experience. Staff: claims assessors spending 40% more time on complex, high-value claims — AI handling triage of simpler claims, assessors focusing on cases requiring human judgment. Retention: zero claims staff attrition due to AI — change management programme addressing concerns, staff seeing AI as enhancing rather than replacing their role. Financial: claims triage efficiency AUD 4.2M (processing cost reduction for 42,000 triaged claims). Fraud improvement AUD 2.8M (earlier fraud detection and prevention). Staff productivity AUD 1.5M (assessors spending more time on complex claims). Customer satisfaction (estimated retention improvement) AUD 800K. Total annual impact from pilot alone: AUD 9.3M. Additional use cases in pipeline: AUD 12M+ expected from next 4 use cases. Consulting: AUD 480,000 (16-week engagement). AI platform: AUD 320,000 (first year infrastructure). ROI: 19.4x first-year return (on consulting investment, from pilot use case alone).