Key points

  • Generative AI is in the trough of disillusionment; agentic AI is approaching the same peak of inflated expectations that preceded it
  • AI readiness is a measurement of data discipline, not budget size
  • Good AI needs bad data: models trained only on successful experiments carry an inherent success bias
  • AI in the lab should be agentic but not autonomous; the scientist decides, the AI retrieves and builds

The hardest work happens before a single model is deployed

At a recent joint Zifo and Sapio Sciences event in Hamburg, R&D leaders and data scientists gathered to tackle a question at the heart of AI readiness in biopharma: why isn’t this scaling?

The technology exists. The investment is there. Yet across the biotech and pharmaceutical CDMO sector, the gap between proof of concept and operational reality remains stubbornly wide. The answers shared in Hamburg were more grounded than the question usually gets.

The cycle of inflated expectations

Generative AI has officially entered what Gartner calls the trough of disillusionment, the phase where early promise collides with operational reality, and organizations start questioning what they actually bought. Agentic AI is now climbing the opposite end of that curve, attracting inflated expectations that tend to precede a hard correction.

Adam Paton, Head of Strategic Accounts at Zifo, pointed to Google Glass as the cautionary tale. In 2014, it was supposed to revolutionize the lab, allowing scientists to run discovery from across the globe. It failed because the infrastructure was not ready for the hardware. The way organizations actually work and handle data had not caught up with the technology. The technology moved faster than the foundations required to support it. That pattern is repeating today.

The data potential myth

Scientific data environments are complex by default. Instrument outputs, LIMS records, legacy PDFs, and VBA-heavy Excel files accumulate across systems that were never meant to speak to one another. Add inconsistent naming conventions and absent metadata, and the core problem comes into focus. Dr. Marko Gentzsch, Team Lead of the Digital Office at Richter Biologics, was blunt: most organizations aren’t sitting on a goldmine of untapped AI potential. They are sitting on data that AI cannot reliably use.

In practice, deploying isolated AI tools at a departmental level often makes this worse, creating a proliferation of ungoverned, disconnected solutions that are harder to consolidate than the fragmented systems they were meant to replace.

What genuine progress looks like

The organizations making headway are those that stop looking for a tool and start building the data foundation for AI, beginning with their system of record. Dr. Prashant Vaidyanathan, Director of Data Science and Software Development at OXB, described a multi-year trajectory from a fragmented ecosystem to a unified platform. By building a custom ontology first, ensuring every sample and measurement meant the same thing across every department, OXB turned their data into a digital feedback loop. Causal inference models and internal LLMs are operational realities at OXB today, but only because they did the unsexy work of standardization first.

This shift is also visible in how platforms themselves are evolving. Kelly Maddison, Solution Engineer at Sapio Sciences, described how the latest generation of ELN technology embeds AI directly into the scientific workflow rather than layering it on top, enabling scientists to query data, generate experiment templates, and run analyses through natural language without leaving the platform.

To get good AI, you have to give it your worst data

Perhaps the most counterintuitive lesson from the event involves how these systems are trained. The instinct is to clean data before feeding it to AI, showing the model only successful experiments. Yuri de Lugt, Global Director of Field Marketing at Sapio Sciences, argued that this creates an inherent success bias. The biggest breakthroughs in science rarely come from repeating what worked; they come from finally understanding why something didn’t. Building an AI-ready foundation means capturing and structuring failed experiments with the same rigor as the successes. Good AI needs bad data.

Agentic but not autonomous

Underpinning all of this is a design philosophy that ran as a consistent thread through the day: the scientist stays in the loop. As the industry moves toward agentic workflows, where AI retrieves data, builds experiment backbones, and suggests next steps, the consensus in Hamburg was clear. 

AI in the lab should be agentic but not autonomous. The AI retrieves and builds, but the expert decides. Crucially, this isn’t just a safeguard for regulated GxP environments. It’s a prerequisite for the organizational trust that makes adoption stick.

The AI readiness gap in biopharma is closeable. But it closes from the foundation up, not from the tool down.

Adam Paton, Dr. Marko Gentzsch, Dr. Prashant Vaidyanathan, and Kelly Maddison

Adam Paton, Dr. Marko Gentzsch, Dr. Prashant Vaidyanathan, Yuri de Lugt and Kelly Maddison presented at Practical AI for Science Leaders, a joint Zifo and Sapio Sciences event held in Hamburg in April 2026.