AI Consulting in Singapore

Financial Services / Insurance

A Singapore insurance company (general insurance — motor, property, travel, health) with SGD 800 million annual premiums had deployed 6 AI models over 3 years with no formal AI governance: (1) Claims automation: an ML model auto-approving low-value motor insurance claims under SGD 2,000 — processing 40% of motor claims without human review. (2) Fraud detection: a model flagging potentially fraudulent claims for investigation. (3) Pricing: ML-augmented pricing models for motor and property insurance. (4) Customer churn prediction: a model identifying customers likely to not renew. (5) Chatbot: an NLP-based chatbot handling first-level customer inquiries. (6) Document processing: OCR and classification for incoming claim documents. None of these models had: documented model cards (what data the model was trained on, what its limitations are), fairness testing (whether the models produced equitable outcomes across demographic groups), governance oversight (no one was formally responsible for model performance or bias), drift monitoring (no systematic tracking of whether model accuracy degraded over time), or incident response (no defined process for what happens when a model makes a consequential error). The CRO (Chief Risk Officer) recognized that MAS was increasing scrutiny of AI in financial services — several industry consultations on AI governance had signaled regulatory expectations. The company needed to retroactively implement AI governance for existing models and establish a framework for future AI deployments.

We delivered a 12-week AI governance programme: (1) Week 1-3: AI inventory and risk assessment. We catalogued all 6 AI models with: model purpose and business context, training data sources and vintage, performance metrics (accuracy, precision, recall), downstream impact (what decisions does the model influence? what happens if it is wrong?), and risk classification (using IMDA Model AI Governance Framework risk tiers). Risk classification results: claims automation — high risk (autonomous financial decisions affecting customers), fraud detection — high risk (flags can lead to claim denial and customer investigation), pricing — high risk (determines pricing offered to customers, potential fairness concerns), churn prediction — medium risk (influences marketing, not customer-facing decisions), chatbot — medium risk (customer-facing but limited to information, not decisions), and document processing — low risk (operational efficiency, no customer impact from classification errors). (2) Week 3-6: Fairness and bias audit. For the 3 high-risk models, we conducted comprehensive fairness testing: claims automation — tested approval rates across customer demographics (age, gender, vehicle type, postal code as proxy for socioeconomic status). Finding: claims from customers aged 18-25 had a 12% lower auto-approval rate than other age groups, even controlling for claim value and policy type. The model had learned age-based patterns from historical adjuster decisions that reflected human bias. Pricing model — tested premium quotes across demographics. Finding: postal code was a significant pricing factor, creating de facto socioeconomic pricing differences. While legally permissible in Singapore, this created reputational risk and potential MAS scrutiny. Fraud detection — tested flag rates across demographics. Finding: claims from certain nationalities were flagged at 2.3x the rate of others — the model had learned patterns from historical fraud investigator focus areas that reflected investigator bias. (3) Week 6-9: Governance framework implementation. We established: AI Governance Committee (CRO as chair, CTO, CDO, Head of Actuarial, and Head of Compliance as members — meeting monthly), model risk management policy (defining model lifecycle: development approval, validation, deployment, monitoring, retirement), model cards for all 6 models (standardized documentation: intended use, training data, performance metrics, known limitations, fairness assessment results), monitoring dashboards (tracking model performance metrics, drift indicators, and fairness metrics monthly — with automated alerts when metrics cross defined thresholds), and incident response playbook (what happens when a model makes a consequential error — who is notified, what is the investigation process, what is the communication plan). (4) Week 9-12: Remediation and AI Verify assessment. Claims automation model: retrained with bias-corrected training data (removing age-based decision patterns). Post-correction: age-group approval rate variance reduced from 12% to 2.1%. Fraud detection model: nationality feature removed and model retrained with behavior-based features only. False positive rate reduced by 15% overall (the nationality feature was adding noise, not signal). We ran AI Verify testing on the 3 high-risk models — producing transparency, fairness, safety, and accountability reports suitable for regulatory examination.

MAS readiness: the company is now the most AI-governance-mature general insurer among its peer group in Singapore. When MAS conducted an industry-wide AI governance survey (which insurance industry participants expect to lead to formal guidelines), the company was one of the few able to demonstrate comprehensive governance — a significant competitive advantage in regulatory perception. Bias remediation: the claims automation model's age bias correction increased customer satisfaction scores for the 18-25 demographic by 18% (measured through NPS surveys 6 months after correction). The fraud detection model's nationality-blind retraining reduced false positive investigation costs by SGD 120,000/year (fewer unnecessary investigations of legitimate claims from previously over-flagged demographics). Governance efficiency: the AI Governance Committee processes model change approvals in 2 weeks (vs. the previous ad-hoc process that took 1-3 months because no one was formally responsible). New AI deployments now follow a standardized approval process — the company deployed a 7th AI model (customer risk scoring) in 8 weeks from concept to production, vs. 4-6 months for previous models with no formal process. Cultural change: the governance programme created awareness among data scientists that model fairness is not just a compliance requirement but a product quality issue. The data science team now proactively tests for bias during model development, not just when asked. Consulting cost: SGD 180,000 for the 12-week programme. The fraud detection remediation alone (SGD 120,000/year savings) provides ongoing ROI, and the regulatory readiness position is invaluable as MAS AI governance expectations formalize.