Bayer
Bayer's preclinical data was fragmented across systems: structured study metadata sat alongside decades of PDF safety-study reports, while historical migrations left some metadata missing or inaccurate. Keyword search could not reliably answer questions spanning studies and documents, so researchers spent substantial time finding and checking the approved reports.
Bayer and Thoughtworks evolved PRINCE from a structured-data search platform into a research assistant. The team integrated study metadata and reports, used RAG for PDF evidence and Text-to-SQL for structured data, and used LangGraph to coordinate intent clarification, planning, research, evidence sufficiency checks, and answer writing. Answers link back to source documents. Researchers review prompts before regulatory-related drafting and remain responsible for the final text.
Researchers can select or adjust data sources and verify answers against linked report pages and original passages. For regulatory documents, a user reviews and edits the drafting prompt, then reviews, validates, and finalizes the output. Domain experts curate evaluation questions and reference answers; the team uses user feedback and production traffic to monitor quality.
- Start with researchers' cross-system workflow for finding safety-study evidence and build a searchable data entry point
- Connect PDF originals with incomplete historical metadata while retaining approved reports as verifiable sources
- Route unstructured retrieval and structured SQL queries by question type, with bounded tool permissions
- Use expert-curated question sets and real user feedback to evaluate retrieval, answers, and complex-query response time
- Add checkpoints, node retries, model fallback, and production tracing to multi-step workflows
- Preserve scientists' review, validation, and final accountability before and after drafting
- When historical metadata is unreliable, treat approved original documents as traceable sources of truth
- For structured tables and unstructured reports, design separate SQL and retrieval paths before synthesizing evidence
- Multi-agent workflows need recoverable state, explicit tool boundaries, and stage-specific evaluations
- High-stakes research and regulatory documents need granular citations and expert sign-off
A paper coauthored by Bayer and Thoughtworks reports that PRINCE integrates more than 18,000 internal studies. Among 15 to 20 frequent users surveyed, 75% reported significantly less time spent searching for information. After the multi-agent system was introduced, average response times for complex queries improved by 30%. The system's ability to fully meet user needs scored 3.1/5 on average; no drug-development timeline or financial return was reported.
- martinfowler.com hosts the article, but its author is Thoughtworks consultant Sarang Sanjay Kulkarni, not Martin Fowler.
- Bayer and Thoughtworks project participants coauthored the paper; peer review does not make its outcome claims independently audited.
- The 75% figure comes from feedback by 15 to 20 frequent users. The 30% figure concerns average response time for complex queries; the paper does not publish a full baseline or test sample.
- The paper also reports a 3.1/5 user-needs score. Neither a shorter drug-development cycle nor automated regulatory decision-making is established.
- Domain-specific Researcher sub-agents and automatic metadata repair are described as future work, not deployed capabilities.