AI in Pharmaceuticals

GSK pharmacovigilance scientists built MLIT, an XGBoost-based tool that predicts how safety reviewers will triage statistical safety alerts and explains its predictions. In two four-month prospective validations covering more than 2,300 alerts, 92% of 881 vaccine alerts were closed in line with its top prediction, and reviewers self-estimated a 24% cut in triage time. It is not yet in routine use.

Novartis consolidated its market research into Sherlock, a Market Logic Software platform used by over 7,500 employees, and in July 2024 added the DeepSights generative AI tool, which answers questions from the existing research archive using retrieval-augmented generation with citations. Novartis reports a 56% reduction in primary market research spend, a $30 million annual reduction in research spend and two hours saved per major research query.

In an ISPE podcast write-up, Sanofi's Miguelina Matthews describes Plai, an AI-powered quality intelligence tool now used across multiple manufacturing sites globally. It started as a pilot automating deviation and complaint trending, runs under Sanofi's RAISE governance framework, and cuts investigations that once took hours or days.

Merck built an internal generative AI platform that combines table pre-processing with LLM authoring to draft clinical study reports, cutting first drafts from two to three weeks to three to four days. Across multiple studies, a fully human-reviewed first draft fell from 180 to 80 hours on average and draft errors dropped 50%; the first live CSRs were submitted in 2025.

Bayer fine-tuned a Microsoft Phi small language model on its proprietary crop protection label data, regulatory rules and expert-authored Q&A to build E.L.Y. Crop Protection (Mini), hosted in Azure AI Foundry and delivered through its E.L.Y. Copilot to advisors and selected retail partners. Bayer estimates 5-10% productivity gains among early users, with complex questions answered in under 30 seconds instead of days.

Bayer's PRINCE (Preclinical Information Center) is an agentic RAG system built on LangGraph that lets preclinical researchers search and ask questions across structured study metadata and unstructured PDF study reports. A five-stage agent workflow (Clarify, Think and Plan, Researcher, Reflection, Writer) gives cited answers, with RAGAS evaluation run daily on live traffic.

Bayer China built Dr. Bei, a voice role-play training bot for pharmaceutical medical representatives on the Claude model via Amazon Bedrock, with AWS Lambda and API Gateway. It went from proof of concept to launch in 72 days, grew from 450 users at launch to over 1,000 within six months, and saved 6,000-7,000 training hours, raising role-play coaching efficiency by over 50%.

In October 2025 Eli Lilly and NVIDIA announced a strategic collaboration to deploy the world's first NVIDIA DGX SuperPOD with DGX B300 systems — an AI factory of 1,016 NVIDIA Blackwell Ultra GPUs delivering over 9,000 petaflops of AI performance, equivalent to ~7 million Cray supercomputers per GPU. The factory trains large biomedical foundation and frontier models for drug discovery, runs NVIDIA BioNeMo for molecular and antibody design, Omniverse digital twins of manufacturing lines, Isaac robotics in factories, NeMo-based AI agents in labs, and MONAI for imaging. Selected models are exposed externally via Lilly TuneLab, a federated-learning platform providing biotechs access to over USD 1 billion of Lilly's proprietary drug-discovery data, now also offering NVIDIA Clara open foundation models. The deployment is anchored by Lilly's USD 50B U.S. manufacturing and R&D expansion including the $4.5B Medicine Foundry in Indiana.