How Fanatics Betting and Gaming built a multi-agent customer support system
Amazon Web Services Parker Bradshaw
Fanatics Betting and Gaming built a multi-agent AI support system on AWS. It handles state-by-state betting rules, and sends risky chats to humans fast.
Based on reporting by Amazon Web Services, Parker Bradshaw — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Fanatics Betting and Gaming has built a support system that looks a lot less like a chatbot and a lot more like a small team of specialists. That’s the point. Sports betting support is messy in a way most customer service systems never have to deal with: live events, sudden spikes, and rules that change depending on which state a customer is in.
The company says demand can jump to more than 40 inquiries every two minutes during major sporting events, including the NFL playoffs and the Super Bowl. A customer in Indiana does not necessarily get the same answer as one in New Jersey, and the system has to keep up without turning every tricky question into a wait for a human queue.
So FBG built an orchestration-based setup on AWS instead of one giant chatbot. A main supervisor agent receives the message, then calls out to specialized tools and sub-agents before sending back one response. The stack runs on Amazon EKS, with Spring AI services and MCP servers deployed as Kubernetes services so tools can scale and update separately.
The company also leaned hard into model choice. Amazon Bedrock gives it access to multiple foundation models through one API, while Bedrock Guardrails helps block prompt injection and keep conversations inside the lines. For responsible gaming classification, FBG uses Amazon Nova 2 Lite, which reviews each message together with the conversation history. If the risk looks high, the system hands the customer to a human immediately, along with the full context.
The retrieval layer is custom too. Instead of a managed knowledge base, FBG built its own RAG pipeline with documents from state payment guides, FAQs, responsible gaming resources, and account docs. Chunks are embedded with Amazon Titan V2 and stored in MongoDB Atlas, then searched with both state-specific and general queries when needed. And because the knowledge base keeps growing, with hundreds of new documents added monthly, that control matters more than another shiny demo feature.
My take — AI-written commentary, not fact-checked reporting
This is the sensible AI move almost nobody wants to make: stop worshipping one giant model and split the job into smaller parts. Regulated support with state-by-state rules and human escalation is exactly where modular systems beat chatbot theater. The real innovation here is not the model; it’s admitting the model should not run the whole shop.
Read more about this at: Amazon Web Services