Why AI Pilots Don’t Make It to Production: What We Learned at SCDM 2026

That creates a fundamental problem:

AI cannot overcome a fragmented data foundation simply because the model itself is sophisticated.

Before AI can reliably analyze clinical data, organizations need a way to collect information across sources, transform and harmonize it, preserve context and provenance, and make it available in a consistent form.

This echoed a question we have been asking at Adaptive for some time:

Is your data AI-ready?

AI readiness is not just about having enough data.

It means having data that is connected, harmonized, traceable, accessible, and trustworthy enough to support the intended use case.

Without that foundation, organizations risk introducing AI on top of the same fragmentation they were trying to solve.

2. AI can be brilliant and brittle at the same time

One of the most memorable ideas discussed at SCDM was the “jagged frontier” of AI.

Modern AI can perform extraordinarily well on certain tasks and unexpectedly fail on others that appear straightforward.

That distinction becomes especially important in regulated environments.

A system that produces a convincing answer is not necessarily producing a correct answer. And when an AI system encounters incomplete, misleading, or unexpected information, its response may still sound authoritative.

For clinical data management, that changes the question from:

“Can AI perform this task?”
to:
“Can we trust, verify, trace, and govern how AI performs this task?”

The healthcare industry is increasingly developing frameworks around exactly this challenge. Responsible AI initiatives are emphasizing areas such as lifecycle governance, risk assessment, responsible data management, organizational accountability, education, and continuous evaluation.

For clinical research organizations, trustworthy AI therefore cannot be separated from the infrastructure surrounding it.

The model matters. But so do the controls that determine what information it receives, how its output is evaluated, and when a person needs to intervene.

Clinical data management involves judgment, interpretation, review, and accountability. Those responsibilities cannot simply disappear because an algorithm can complete part of the workflow.

Instead, AI can take on appropriate repeatable work while experienced professionals remain responsible for directing the process, evaluating exceptions, and standing behind the outcome.

This is particularly relevant as organizations explore Digital Workers and more autonomous AI agents.

The opportunity is not necessarily to remove clinical data managers from the workflow. It is to reconsider which parts of the workflow genuinely require their expertise.

If AI can handle repetitive review, information gathering, comparison, or preparation tasks, CDMs can spend more time on the decisions where human judgment creates the greatest value.

4. A successful proof of concept is not the same as successful AI adoption

Adaptive Clinical Systems chaired a panel at SCDM focused specifically on this gap:

Why AI Pilots Don’t Make It to Production — and What Actually Works in Clinical Data Management.

The discussion brought together perspectives from Adaptive Clinical Systems, Merck, Syneos Health, and Axcellion.ai around what happens after an AI proof of concept demonstrates technical potential.

This is where many initiatives become stuck.

A sophisticated agent may work. A demo may impress stakeholders. A proof of concept may produce the expected result.

But none of those automatically mean the organization is ready to deploy that technology across real workflows.

The panel highlighted four distinct considerations:

Good science.
AI initiatives need a clearly defined problem, appropriate methodology, and evidence that the approach actually works.

Robust architecture.
An AI solution needs reliable access to the right data and systems, supported by infrastructure designed for production rather than a standalone demonstration.

Balanced ROI.
Organizations need to determine whether the value created justifies the implementation, validation, governance, and operational effort required.

Organizational readiness.
The people expected to use, supervise, and trust the technology need to understand how it changes their work and responsibilities.

During the panel discussion, that final consideration emerged as especially important.

They also need the opportunity to build confidence in the technology through real-world use.

Trust is not created by telling users that an algorithm is accurate. It develops when people can understand its role, see how decisions are supported, recognize its limitations, and experience consistent performance within their actual workflows.

This may explain why technically successful pilots can still struggle to scale: production readiness belongs to the organization as much as it belongs to the technology.

AI readiness starts with data readiness

SCDM 2026 reinforced that the industry is moving beyond asking whether AI belongs in clinical data management.

The more important challenge now is creating the conditions in which it can be used reliably and responsibly at scale.

For Adaptive Clinical Systems, one of those conditions starts upstream: the data foundation. 

Clinical AI needs to work across information originating from different systems, technologies, and formats. A source-independent interoperability layer can connect, transform, and harmonize that information before it reaches downstream applications, while preserving the context and traceability required in clinical research.

This does not solve every challenge involved in bringing AI into production. Good science, appropriate governance, viable economics, and organizational readiness remain essential.

But without reliable data flowing between the systems where clinical work happens, even the strongest AI strategy starts with a disadvantage.

That is why we believe the question “Is your data AI-ready?” should come before “Which AI should we use?”

Because before clinical data can power the next generation of AI and Digital Workers, it needs to be ready for them.