Preclinical drug discovery generates massive datasets scattered across disconnected systems, forcing researchers to manually comb through structured metadata and decades of PDF reports. Traditional keyword searches fail to handle the complexity of scientific queries, creating bottlenecks in pharmaceutical development. Bayer tackled this challenge by building PRINCE, an agentic AI platform that evolved from basic search to conversational question-answering.

PRINCE leverages Retrieval-Augmented Generation to process both structured study data and unstructured documents like regulatory submissions. The system progressed through three phases: Search consolidated metadata from multiple silos, Ask enabled natural language queries against PDF content, and Do introduced multi-agent workflows for complex tasks like drafting regulatory documents. Built on Lang Graph orchestration with FastAPI serving, it integrates OpenSearch for vector representations, Athena for curated data, and PostgreSQL for state persistence.

The architecture reflects deliberate engineering principles the team now recognizes as context engineering and harness engineering. Context engineering controlled what information each model received and how it flowed between research, reflection, and writing stages. Harness engineering managed orchestration, tool boundaries, validation loops, and human review checkpoints. These design choices created pause points ensuring data completeness before generating responses.

By combining generative AI with rigorous validation, PRINCE turns fragmented preclinical data into an intuitive research assistant. The platform reduces manual analysis time and supports faster, evidence-based decisions in drug development pipelines. Researchers can now query historical study reports directly rather than reconstructing context from scattered sources.

The system demonstrates how pharmaceutical companies can operationalize agentic AI while maintaining scientific rigor through structured workflows and human oversight.