TLDR: Match Chat is a real-time, agent-driven AI assistant developed by IBM to enhance the tennis fan experience. Debuting at the 2025 Wimbledon and US Open, it provided nearly a million users with instant, accurate answers to match-related queries. The system integrates Generative AI (GenAI) with Generative Computing (GenComp) within an Agent-Oriented Architecture (AOA) to synthesize insights from live and static tennis data. It boasts 92.83% answer accuracy and an average response time of 6.25 seconds, achieved through innovations like GenAI shielding and a robust multi-agent framework, demonstrating a scalable and reliable approach for consumer-facing AI in dynamic environments.
Tennis enthusiasts often crave instant, accurate information during live matches, from player statistics to real-time predictions. A new system called Match Chat, developed by IBM, is designed to meet this demand, offering a real-time, agent-driven assistant that significantly enhances the fan experience. This innovative platform made its debut at the 2025 Wimbledon Championships and the 2025 US Open, serving nearly a million users with seamless access to a wealth of tennis data through natural language queries.
Match Chat represents a powerful fusion of Generative Artificial Intelligence (GenAI) and Generative Computing (GenComp). GenAI allows the system to synthesize key insights and generate human-like responses, while GenComp provides a structured, logic-based framework that guides and constrains these generative outputs. This hybrid approach ensures that the information provided is not only creative and insightful but also robust, interpretable, and factually grounded. The system’s architecture is built on an Agent-Oriented Architecture (AOA), which uses rule engines, predictive models, and specialized agents to process and optimize user queries before they reach the GenAI components.
How Match Chat Works
At its core, Match Chat processes a vast array of real-time and static data. This includes over 300 statistical measures per match, such as aces, serve percentages, and unforced errors, updated continuously within seconds. Beyond basic metrics, it captures detailed information on serve, return, rally patterns, and even player movement. Player data, including historical head-to-head records, career statistics, biographical details, and curated GenAI summaries, provides rich context. The system also incorporates two types of predictive models: a pre-match likelihood to win and a live (in-play) likelihood to win, which dynamically updates based on the evolving match state, momentum, and other factors.
For queries that fall outside the immediate scope of match or player data, such as questions about food vendors or parking, Match Chat intelligently redirects to a supplemental knowledge base. This ensures a comprehensive information service for fans, whether they are on-site or engaging remotely.
A Seamless User Experience
The design of Match Chat prioritizes an intuitive and frictionless user experience. Fans can access the assistant from live match score pages or other assistant systems, either by selecting predefined questions or submitting free-form queries. The interface presents high-level tennis categories like ‘Player Career Stats’ or ‘Likelihood to Win,’ allowing users to drill down or ask open-ended questions. This tiered interaction model makes the system approachable for all users, regardless of their technical familiarity. The system is designed to mask its underlying architectural complexity, providing advanced AI capabilities without any perceived delay or difficulty for the user.
Also Read:
- ACE Framework: Automating API Tooling for Large Language Models
- Empowering Students with a Local GPU-Accelerated AI Assistant
Ensuring Speed and Accuracy
Match Chat achieved an impressive answer accuracy of 92.83% with an average response time of 6.25 seconds, even under heavy loads of up to 120 requests per second. A key innovation enabling this performance is ‘GenAI Shielding.’ This strategy offloads over 50% of user traffic from direct GenAI inference by using predefined data synthesis patterns and static knowledge stores for common queries. When GenAI is necessary, the system employs a network of specialized agents. If a response isn’t returned within a user-defined 6-second timeout, a fallback mechanism activates a more robust data synthesizer, ensuring continuous responsiveness. Furthermore, a Hate, Abuse, and Profanity (HAP) pipeline is integrated to ensure content safety and appropriateness, maintaining a respectful and secure environment for all users.
The successful deployment of Match Chat at two major Grand Slam tournaments, supporting nearly a million unique users with 100% uptime, underscores its scalability and reliability. This work provides a practical blueprint for developing high-performance, consumer-facing AI systems that excel in speed, precision, and usability within dynamic, real-time environments. To learn more about the technical details, you can read the full research paper here.
