Bayer has launched PRINCE, an agentic artificial intelligence platform designed to streamline preclinical drug discovery by transforming how researchers access and analyze vast, siloed data repositories. Initially deployed to end users in early 2024 with advanced agentic capabilities integrated later that year, the system addresses longstanding inefficiencies in pharmaceutical R&D where traditional keyword-based searches fail to navigate complex scientific queries. PRINCE represents an evolutionary platform that progressed from a structured metadata search tool to a natural language question-answering system, and finally to an active research assistant capable of executing multi-step workflows. At its core, the architecture leverages LangGraph for orchestration, coordinating a specialized multi-agent system. This workflow begins with intent clarification, followed by a dedicated reasoning module that evaluates strategy and tool selection. The primary Researcher agent then employs a hybrid retrieval approach, combining Retrieval-Augmented Generation for unstructured PDF study reports with Text-to-SQL queries for precise structured metadata in Amazon Athena. A Reflection agent subsequently validates data sufficiency, prompting targeted follow-ups if information is incomplete, before a Writer agent synthesizes findings into final responses with granular, verifiable citations. Engineering reliability remains the platform’s defining characteristic. Bayer explicitly applies context engineering to ensure each agent receives only the necessary information for its specific stage, preventing context pollution and improving traceability. Harness engineering governs the broader workflow, enforcing strict boundaries, state persistence in PostgreSQL, automated retries, and multi-provider language model fallbacks to maintain continuity during service disruptions. The system continuously monitors performance via Langfuse, utilizing both curated expert datasets and daily live-traffic evaluations to measure faithfulness, answer relevance, and semantic accuracy. To maintain data integrity, PRINCE incorporates a named entity recognition pipeline that auto-annotates historical study reports, enriching structured metadata with high-confidence extracts while flagging lower-confidence entries for scientific review. User trust is further reinforced through transparent intermediate processing steps, hover-to-verify citations, and mandatory human review for regulatory drafting outputs. The deployment of PRINCE demonstrates that production-grade agentic AI in regulated industries requires more than advanced language models; it demands rigorous workflow control, transparent validation, and resilient error handling. By converting decades of fragmented preclinical data into an interactive, conversational research environment, the platform enables scientists to accelerate data-driven decisions, reduce redundant experiments, and ultimately shorten the timeline for developing safer therapies. Bayer’s iterative development approach, prioritizing early user feedback and continuous architectural refinement, establishes a scalable blueprint for enterprise AI systems where accuracy, auditability, and operational continuity are non-negotiable.