Transforming GSK's supply chain insights across 37 sites
GSK is a global pharmaceutical company producing vaccines, specialty medicines and general medicines across 37 manufacturing sites.
Its supply chain spans demand and planning, manufacturing, logistics and quality, and each persona asks different questions, from whether a plant is producing efficiently to whether a deviation will cause a stockout downstream. Each site also operates a little differently. GSK wanted to change how people interact with the supply chain, but its data spans more than 37,000 tables. The challenge was giving agents enough context to answer accurately at that scale.
Developing an AI-powered supply chain assistant with Genie Agents
GSK built a single interface where supply chain teams receive personalized next-best-action reports by persona and site, plus conversational agents to dig deeper.
Underneath, Genie Agents serve as the semantic layer for structured data on a common data model, alongside tools for anomaly detection and report generation and connectors to key supply chain systems. Using Genie, GSK moved its supply chain AI assistant into production within three months. The assistant now automatically answers questions and generates insights and root cause analyses, and GSK is actively working on agents that act on recommendations.
Implementing Genie Agents for enhanced adoption
To keep quality high, the team codified Genie Agents setup into a checklist every space in the global supply chain follows: one space per business domain built around real user questions, business logic pushed into the data layer, parameterized SQL examples instead of long prose instructions and benchmarks against 20 to 50 gold standard SQL queries. Genie’s streamlined interface makes these practices easier to follow for business users and subject matter experts, and a governed CI/CD pipeline lets SMEs train Genie Agents themselves. User feedback and training helped build trust, and users who were once hesitant now want Genie-powered agents across the whole supply chain.
Scaling Genie Agents with observability and caching
As usage grew, some answers took more than a minute. The team instrumented each step inside Genie, from SQL generation to warehouse execution and retries, and built dashboards from system audit tables to track latency and queries per minute. They then put a tiered system in front of Genie: templated SQL tools for the most common questions, a semantic cache that reuses or adapts SQL from similar past questions and Genie itself for everything else. In benchmarks, latency dropped from 34 to 22 seconds, about a third, and queries reaching Genie fell by 64%, with a 3% to 4% drop in correctness the team considers a fair trade. Genie Agents are now versioned, reviewed, tested and promoted through Git like any other software.