Fanatics Betting and Gaming Implements Multi-Agent Customer Support on AWS

Fanatics Betting and Gaming (FBG) has built a multi-agent customer support system on AWS to address the unique challenges of sports betting support, including jurisdiction-specific rules, real-time responsible gaming monitoring, and traffic surges during major sporting events. As their user base expands rapidly across multiple U.S. states, FBG needed a scalable, accurate, and compliant support infrastructure.

Traditional chatbot solutions based on decision trees proved insufficient for the complexity and volume of queries, which range from account issues to betting regulations and responsible gaming. To meet these demands, FBG designed a multi-agent AI system leveraging AWS services, including Amazon EKS, Amazon Bedrock, and custom retrieval pipelines.

The architecture features an orchestrator pattern where a primary agent manages customer messages, coordinating specialized sub-agents and tools. This modular design allows for independent updates and scalability, supporting new features and jurisdictions without system overhaul. Customer messages flow from the FBG app through Salesforce Einstein, then to a Spring AI service on Amazon EKS, passing through Amazon Bedrock Guardrails and responsible gaming classifiers before reaching the supervisor agent.

The supervisor agent, running Anthropic Claude on Amazon Bedrock, determines which specialized tools to invoke—such as retrieval augmented generation (RAG) pipelines for FAQs, account and transaction information, or escalation to human agents. The system employs a custom RAG pipeline with Amazon Titan V2 embeddings, enabling precise retrieval of jurisdiction-specific and general documents, which improves response accuracy.

Responsible gaming compliance is managed via Amazon Nova 2 Lite classifiers, which evaluate interactions for signs of problem gambling and trigger escalations to human agents when necessary. Amazon Bedrock Guardrails help prevent prompt injection and keep conversations within appropriate boundaries, balancing security with customer experience.

The multi-model architecture uses Amazon Bedrock models like Nova 2 Lite for classification, Anthropic Claude for orchestration, and Titan V2 for embeddings, with workload distribution strategies to ensure high throughput during peak events. The system has demonstrated significant improvements within two months, increasing containment and resolution rates by over 50%, and handling high traffic volumes without degrading performance.

Fanatics continuously evaluates and refines the system through conversation reviews, prompt engineering, and real-time monitoring, ensuring ongoing improvements and compliance. For organizations aiming to develop similar support systems, starting with a narrow scope, deploying on Amazon EKS or Bedrock AgentCore, and investing early in evaluation infrastructure are recommended steps.