How Fanatics Betting and Gaming built a multi-agent customer support system
Fanatics Betting and Gaming (FBG) built a multi-agent AI customer support system on AWS to handle the unique complexities of sports betting, where rules vary by jurisdiction and support volume surges during live events The system uses an orchestrator pattern with a Supervisor Agent coordinating specialized sub-agents and tools, deployed on Amazon EKS with Amazon Bedrock for model-agnostic LLM access Key architectural components include Amazon Bedrock Guardrails for prompt injection detection, a
Analysis
TL;DR
- Fanatics Betting and Gaming (FBG) built a multi-agent AI customer support system on AWS to handle the unique complexities of sports betting, where rules vary by jurisdiction and support volume surges during live events
- The system uses an orchestrator pattern with a Supervisor Agent coordinating specialized sub-agents and tools, deployed on Amazon EKS with Amazon Bedrock for model-agnostic LLM access
- Key architectural components include Amazon Bedrock Guardrails for prompt injection detection, a Responsible Gaming classifier powered by Amazon Nova 2 Lite, a RAG pipeline, and Model Context Protocol (MCP) servers for account/transaction data
- The modular design allows independent scaling and improvement of each agent, enabling rapid addition of new capabilities without rewriting the core system
- The system handles surges of over 40 inquiries every two minutes during major sporting events while maintaining response quality and compliance with responsible gaming requirements
Why It Matters
This case study demonstrates a production-grade multi-agent architecture for a high-stakes, compliance-heavy industry where accuracy and speed are critical. It provides a replicable blueprint for organizations dealing with jurisdiction-specific rules, volatile traffic patterns, and the need for real-time responsible-safety classification—challenges that extend well beyond sports betting into fintech, healthcare, and other regulated domains.
Technical Details
- Architecture Pattern: Orchestrator-based multi-agent system where a Supervisor Agent receives customer messages, delegates to specialized tools and sub-agents, and returns unified responses. Built on Amazon EKS for containerized, independent deployment and scaling of each agent.
- LLM Infrastructure: Amazon Bedrock provides model-agnostic access to multiple foundation models through a single API, enabling task-specific model selection and seamless model swapping. Amazon Nova 2 Lite powers the Responsible Gaming classification agent.
- Security & Compliance Pipeline: Requests flow through Amazon Bedrock Guardrails for prompt injection detection, then through a Responsible Gaming classifier using a compliance-approved classification framework. High-severity classifications trigger escalation logic.
- Tooling & Integration: The system integrates a RAG (Retrieval Augmented Generation) pipeline for knowledge retrieval, MCP (Model Context Protocol) servers for account and transaction data access, and a transfer-to-agent tool for human escalation. Salesforce Einstein serves as the chat interface layer, with Spring AI handling request validation and routing on EKS.
- Scale Requirements: Designed to handle support volume spikes exceeding 40 inquiries per two minutes during events like the Super Bowl and NFL playoffs, with state-specific rule variations across multiple U.S. jurisdictions.
Industry Insight
- The orchestrator-plus-specialized-agents pattern is proving to be a scalable alternative to monolithic chatbots, particularly for domains requiring jurisdiction-aware responses and real-time compliance checks—organizations should evaluate this architecture when their support complexity exceeds what a single model or decision tree can handle.
- The integration of responsible AI guardrails (Bedrock Guardrails) and domain-specific classifiers (Responsible Gaming) directly into the request pipeline demonstrates that compliance cannot be an afterthought in regulated industries; it must be architecturally embedded at every layer of the AI inference path.
- The use of MCP servers for structured data access (account/transaction tools) alongside RAG for unstructured knowledge highlights a emerging best practice: separating structured operational data from unstructured domain knowledge, each served through purpose-built retrieval mechanisms rather than relying on a single knowledge base.
Disclaimer: The above content is generated by AI and is for reference only.