BMS grew from fragmented local analytics to a platform supporting thousands of researchers and over 80 peer-reviewed publications. They achieved this by making reproducible, reusable workflows the default across the organization, which proves that true scale is about multiplying decisions rather than just adding users or compute.
Workflows that can be rerun years later with identical results protect institutional knowledge and satisfy the auditability requirements of regulated pharmaceutical environments.
BMS measures platform adoption by how many teams borrow and build on each other's patterns. Raw growth in headcount or compute spend without reuse is expensive infrastructure, not scale.
Data science assets frequently get stuck between exploration and deployment. BMS established a clear process to move validated workflows into production.
Locking down tool environments cuts off the initial pipeline of experimentation where early clinical and manufacturing breakthroughs happen. Broad, low-friction access to data science infrastructure is where enterprise AI capability starts.
Connecting tools like large language models to platform APIs allows a researcher to move from a raw data file to an interactive analytics app via a single prompt.
Bristol Myers Squibb started building its enterprise data science capability in 2015 from a fragmented environment. Compute access was siloed, data scientists were limited by local hard drives, version control barely existed, and spinning up a basic application could take eight months. There was no shared infrastructure to move models from exploration into production.
BMS removed that friction by standardizing on Domino. Over the last decade, that single foundation has expanded to support 5,000 users across nine business units, driving 30,000 projects. Jacob Albrecht, Director of the Intelligence Systems Lab, walks through the phases of this scale:
The practical impact of this architecture is clear. Jacob shares how robotic screening workflows combined with machine learning compressed months of manual chemistry experiments down to a single week. He also highlights a dataset built inside BMS for optical chemical structure recognition that became an external benchmark used by OpenAI to test its next-generation models.
