Fanatics Betting and Gaming has built a multi-agent customer-support system on Amazon Web Services to answer sports-betting questions at scale, including account problems, deposit limits, state-specific rules and responsible-gaming resources. The system was designed for a business where support traffic can spike during live sports, when requests can climb to more than 40 every two minutes.
According to AWS, the company needed something more flexible than decision-tree chatbots because betting rules vary by jurisdiction, so a customer in one state may need a different answer from a customer in another. Fanatics also said its older support model required more human touches as it grew, which pushed operating costs up in step with the customer base. Ian Botts, the company’s chief technology officer, said the aim was faster and more accurate answers without compromising compliance, and something that would get better over time rather than simply bigger.
The architecture sends each message from the mobile app through Salesforce Einstein and then to a Spring AI service on Amazon EKS. From there, it passes through Amazon Bedrock Guardrails and a responsible-gaming classifier before reaching a supervisor agent that runs Anthropic Claude Sonnet on Bedrock and orchestrates a retrieval-augmented generation pipeline, account and transaction Model Context Protocol servers, and a transfer-to-agent tool.
The responsible-gaming layer uses Amazon Nova 2 Lite and evaluates both the current message and the conversation history, so it can detect escalation patterns instead of relying on single-message keyword matching. High-severity flags trigger an immediate handoff to a human agent with the full conversation context, while lower-severity flags are logged for compliance review and the exchange continues. Fanatics said responsible gaming is both a regulatory requirement and a core value, and Trevor Gurgick, head of applied AI, said a withdrawal question can itself be a responsible-gaming moment.
Fanatics built its own retrieval system rather than relying on a managed knowledge base, so it could control ingestion, chunking and retrieval. The support corpus includes state-specific payment method guides, FAQ articles, responsible-gaming material and account-management guides, with documents split into fixed token chunks, embedded with Amazon Titan V2 and stored in MongoDB Atlas. For jurisdiction-specific queries, the system combines state-specific and general document searches before generating a response, and the knowledge base is being expanded with hundreds of new documents each month as gaps are identified through conversation analysis.
The company said the design is modular. New tools can be added without rewriting the core system, the MCP server and Spring AI service scale independently on EKS, and autoscaling adds capacity automatically during traffic spikes. Fanatics also uses a round-robin strategy across model regions for the supervisor agent to avoid throughput limits during peak events.
In its first two months of deployment, Fanatics said containment improved by about 56 percent and resolution by about 53 percent, while the system resolved thousands of cases autonomously. Customer satisfaction is trending upward, and the company said conversation quality has improved to the point that customers often do not realise they are speaking to AI.
Ada CX says responsible gaming increasingly “lives or dies in execution,” while the New Jersey Division of Gaming Enforcement defines it as policies and practices meant to promote safe and enjoyable gambling experiences while minimizing harm, and says it works closely with licensed casinos and online operators.
