Innovation
Version 1 provide AI fraud prevention platform
A large public-sector welfare agency disbursing over $300 billion annually partnered with Version 1 to transform fraud detection from a reactive, post-payment process into a proactive, real-time prevention system. Leveraging AWS technology, the solution delivers sub-second risk scoring on 20 million benefit claims each year, safeguarding taxpayer funds and improving service for vulnerable citizens.
Challenges
The agency suffered $13.4 billion of annual financial leakage; $12 billion in over-payments (75 percent fraud) and $1.4 billion in under-payments due to manual, rules-based reviews conducted after payments issued. Legacy batch processes imposed days-to-years detection cycles, exposing claimants to inaccurate debts and necessitating costly, time-consuming recoveries. The solution needed to:
- Prevent improper payments in real time at national scale
- Maintain strict government data privacy and security standards
- Deliver sub-second scoring for 20 million claims, 24×7
- Enable rapid rule updates and model refinements
Solution
Version 1 engineered an AI-powered Fraud & Error Detection Platform featuring:
Dual Processing Pipelines: A “Rule Flow” on Amazon EMR for deterministic checks and a “Model Flow” on Amazon ECS (Fargate) running XGBoost/GBM models retrained in SageMaker Studio; both orchestrated by AWS Step Functions.
- Lakehouse Analytics: Amazon S3 unifies structured benefit records, unstructured case notes, and real-time transaction streams. Collaborative notebooks accelerate feature engineering.
- Real-Time Risk API: Containerized execution engine computes risk scores in < 200 ms, publishing via encrypted Amazon API Gateway for case-worker intervention before funds disburse.
- DevSecOps & IaC: Terraform-managed landing zones, GitLab CI/CD with blue-green deployments, container scanning, and automated compliance guardrails enable weekly releases with zero unplanned downtime.
- Sustainability & Efficiency: AWS Graviton3 and Trainium instances cut compute energy by 80 percent and AI training costs by 40 percent, aligning operations with net-zero targets.
Real differences, delivered
We delivered capabilities that did not exist in any vendor toolkit:
- From Detect to Prevent: Risk scoring before payment stops improper disbursements in real time, eliminating reliance on retrospective recovery.
- Operational Agility: Weekly rule updates empower policy teams to respond to emerging fraud patterns within hours instead of months.
- Scalability & Resilience: Serverless, auto-scaling architecture handles benefit peaks with zero idle capacity and sustained sub-second response times.
- Skills & Innovation: A joint Cloud Centre of Excellence upskilled 200 engineers in DevSecOps and sustainable-architecture, reducing consulting spend by 25 percent.
- Open-Source Impact: Version 1’s risk-scoring library adopted by additional government agencies, extending platform value beyond the original engagement.
Delivered
- $1.66 billion saved in FY 2024 by preventing fraud and error. 10 percent of total annual leakage eliminated in year one.
- $8.4 billion projected cumulative savings by 2028, averaging $1.68 billion per year.
- 100 percent of 20 million annual claims receive real-time risk scoring, with zero unplanned downtime.
- 12× acceleration of release cadence, enabling same-day deployment of new detection rules.
- 80 percent reduction in compute energy use and 40 percent lower AI training costs, delivering a sustainable, cost-efficient platform.