Bayer’s PRINCE AI Can Read Decades of Drug Safety Reports in Seconds—Here’s How

Preclinical drug discovery is notoriously data-intensive, with researchers struggling to navigate vast volumes of safety study reports using rigid Boolean keyword search. To address this, Bayer AG partnered with Thoughtworks to build PRINCE (Preclinical Information Center), a cloud-hosted platform that evolved from basic search into an intelligent, production-ready agentic AI system. The system leverages Agentic Retrieval-Augmented Generation (Agentic RAG) and Text-to-SQL to integrate decades of proprietary data, allowing researchers to ask complex natural language questions and receive accurate, context-rich answers. PRINCE even helps draft regulatory documents, dramatically reducing the time spent on manual data retrieval and analysis.

The engineering behind PRINCE is framed through two key lenses: context engineering and harness engineering. Context engineering dictates what information each specialized agent (researcher, reflector, writer) receives and how context flows between steps like research, reflection, and writing. Harness engineering builds robust scaffolding around the models—including orchestration, tool boundaries, state persistence, retries, fallbacks, validation loops, observability, and human-in-the-loop integration. This approach ensures reliability, transparency, and explainability, which are critical for regulated pharmaceutical environments. The system is now being iteratively improved with named entity recognition and annotation to further enhance data quality. PRINCE demonstrates that agentic AI can move from experiments to production, delivering tangible efficiency gains while maintaining governance and compliance.