AI Adoption Challenges 

Data Quality & Data Availability  

By a long shot, the group voted data quality and availability as the biggest impediment to AI in biopharma.  In large and complex pharma organizations, decentralized and siloed data capture and management leads to incomparable results and loss of information. Even something as simple as different naming conventions for a given assay can create chaos when it comes to obtaining data clean enough for AI applications.  

Companies ranging from start-ups to pharma giants like Sanofi recognize that the first step in this journey is building the procedures and infrastructure required for a high-quality, balanced, and accessible data corpus. Therefore, they have invested in enterprise software platforms that centralize, structure, contextualize, and harmonize their data.  

“The promise of AI in drug discovery hinges on feeding it the right data. With Genedata, we can generate structured, high-quality datasets that drive innovation and enable scalable automation for future AI and ML use cases.”  

- Patrick MM Shelton, Ph.D. Principal Scientist, Pfizer 

Watch Sanofi webinar 

Integration with Existing Data Systems and Infrastructure 

What's the point of a fancy tool that lives in a vacuum? For AI to deliver ROI, it must be integrated both technically and operationally into an organization’s existing systems and workflows. At the technical level, this can mean using data federation and integration to connect data sources, creating a development environment where organizations can plug in and manage their own machine learning models. It also means enabling AI agents and LLMs to access data through model context protocol (MCP) servers. 

Security and Privacy 

Data security is critical to protecting valuable IP, such as target or candidate information, especially in an industry like biopharma, where new product development requires years-long investments. Moreover, biopharma handles patient data, where privacy is critical.  

A big challenge when implementing AI is the need for computing resources, which means using external services such as AWS—and thus giving up some of the control that comes with an on-premises deployment. For companies interested in using LLM-based AI agents, access to foundation models can pose another big security risk. LLMs must be prevented from accessing data outside the relevant context, because doing so might potentially bypass governance and audit logs. LLMs must also be prevented from changing data without review.  

This makes it even more important to work with software and solution partners qualified in regulatory compliance and data protection (for example, with SOC2 and ISAE3000 accreditation). They should offer configurable authentication options or APIs so IT teams can apply their own strategies and restrictions for ensuring only authorized access to their systems. At the most advanced, they may even have experience implementing AI or LLM agents and other AI technologies in a GxP environment: for example, as we have at Merck, setting up technical guardrails such that the LLM sees only the key metadata it needs to execute its tasks, without sending it any sensitive data. Resistance to Change 

The final hurdle to adopting AI isn’t technical at all, but entirely human. Fear of being replaced by AI, difficulty learning to use AI tools effectively, and skepticism about overhyped AI technology among the workforce can impede adoption. Here, it's best to integrate AI into daily workflows where it has a clear, sensible use case and delivers tangible value to users. For example, we have used AI to automate the most tedious image analysis and data labeling tasks, freeing scientists to do more rewarding work.  Human-in-the-loop approaches and interpretability techniques can also help ensure trust in AI outputs. Finally, training and education on the appropriate and effective use of AI can increase awareness and maximize adoption of newly introduced AI workflows and tools.