42Q's journey into implementing GenAI capabilities in their Manufacturing Execution System (MES) represents a thoughtful and phased approach to bringing LLMs into a production environment. This case study showcases both the potential and challenges of integrating AI assistants into complex enterprise systems. **System Context and Initial Challenge** 42Q operates a cloud-based MES system that handles manufacturing operations with numerous features and modules. The system's complexity created challenges for users in terms of understanding functionality, keeping up with new features, and dealing with different terminology across organizations. The traditional solution of consulting MES administrators was becoming a bottleneck due to their limited availability. **Phase 1: Interactive Helper Chatbot** The initial implementation focused on creating an AI assistant (named Arthur, inspired by "The Hitchhiker's Guide to the Galaxy") that could understand and explain the system. The team: * Trained the system on comprehensive documentation * Transcribed and incorporated training videos * Embedded the chatbot directly into the 42Q portal * Implemented multilingual support * Maintained conversation context * Provided source references for all responses The system architecture leverages AWS services extensively, with documentation and transcribed content stored in S3 buckets, implementing a Retrieval Augmented Generation (RAG) approach. They selected Anthropic's Claude through AWS Bedrock as their primary LLM after testing various options. **Phase 2: Data Integration** The second phase expanded Arthur's capabilities to include real-time data access. This involved: * Connecting to the MES database through APIs * Implementing live query capabilities * Supporting natural language queries about production data * Enabling data visualization and formatting options * Maintaining context across both documentation and live data queries An interesting observation was how the system began making unexpected but valuable connections between training material examples and live data, providing richer context than initially anticipated. **Technical Implementation Details** The solution architecture includes: * AWS Bedrock for LLM integration * Lambda functions for API calls * API Gateway for access control * S3 storage for documentation and training materials * RAG implementation for context enhancement * Guardrails for responsible AI usage **Security and Data Privacy Considerations** The team placed significant emphasis on data security, ensuring: * All data remains within the AWS account * Authentication through existing portal login * Implementation of appropriate access controls * No external training on customer data **Current Challenges and Future Considerations** The team is currently grappling with the extent of automation to allow. Key questions include: * Whether to allow the AI to make operational changes * How to implement appropriate guardrails * Balancing automation with safety * Manufacturing industry readiness for AI-driven decisions **Results and Benefits** The implementation has shown several positive outcomes: * Faster response times compared to manual documentation lookup * Comprehensive system understanding through combined documentation and practical examples * Increased usage by internal teams * Adoption for training purposes * Enhanced support for night shift operations * Improved accessibility of system knowledge **Production Deployment Considerations** The team's approach to production deployment shows careful consideration of enterprise requirements: * Single chatbot interface for all resources * Integration with existing authentication systems * Implementation of guardrails * Maintenance of data privacy * Scalable architecture using AWS services A particularly noteworthy aspect of this implementation is the careful, phased approach to adding capabilities. Rather than immediately implementing all possible features, the team chose to start with documentation understanding, then added data access, and is now carefully considering the implications of allowing the system to make operational changes. The case study also demonstrates the importance of domain knowledge in LLM implementations. The system's ability to combine documentation understanding with real-world manufacturing data and training examples shows how LLMs can provide value beyond simple question-answering when properly integrated with domain-specific content and data. The team's current debate about allowing the AI to make operational changes reflects the broader industry discussion about the role of AI in critical systems. Their cautious approach to this next phase, considering both technical capabilities and organizational readiness, demonstrates a mature understanding of the challenges in implementing AI in production environments.