Confidentiality, governance, and value: how to move forward responsibly
AI agents are still at an early stage of development, and improvement is expected. In practice, current limitations are not only technical, they also reflect limited organisational experience in applying these tools to established processes.
In clinical endpoint adjudication (CEA) and drug development more broadly, the challenge is further compounded by patient confidentiality and corporate intellectual property constraints.
This final article discusses practical considerations for deploying AI agents in CEA and suggests realistic ways to move forward while preserving oversight and accountability.
Confidentiality and Controlled Use of Data
One of the major hurdles when applying AI-based solutions in corporate settings is data confidentiality. Many high-performing large language models (LLMs) have been developed in the public-domain and, when in use, are incorporating any additional information to their learning base. With such models, companies may fear confidential data could be exposed or retained inappropriately. This concern is especially relevant in the clinical setting when patient data or sensitive research information is involved.
At the same time, internally developed models are emerging, although their performance may not match the most advanced public models. Programming AI agents for clinical research use requires access to several distinct categories of information: specialised medical knowledge (public), patient information (private), and research information (confidential). A practical direction is therefore to develop internal capabilities that allow AI agents to operate within controlled environments and governance frameworks.
This will require time and effort. However, once developed, such capabilities can be reused across clinical trials within the same therapeutic domain. Because clinical development programmes typically include multiple studies, potential synergies can be significant.
Investment, Return, and the “Learning Gap”
Deploying new technologies has a cost, and in the case of AI the cost can be high. Organisations should therefore expect a clear return on investment. A recent MIT study1 has shown that despite large investments in generative AI, many organisations report little or no measurable return from their projects. The study identifies the key barrier as the “learning gap” : most corporate systems do not retain feedback, do not accumulate operational knowledge, and do not improve over time.
This observation connects directly to the core distinction made earlier in this series: “chatting with a model” is not the same as using agentic AI. Unlike traditional automation tools, AI agents can be designed with memory and feedback loops so that performance improves with experience. That improvement, however, does not happen automatically. It depends on teaching, learning, and, most importantly, through knowledge of processes, stakeholders, inputs, and goals.
In endpoint adjudication, this means that successful deployment requires more than technology selection. It requires clarity on roles, governance, decision boundaries, and on how feedback is captured and used to improve performance over time.
A Caution on Full Autonomy
It is also worth considering the longer-term trajectory of automation. Manufacturing has seen the emergence of “dark factories” or “lights-out factories”, highly automated production environments with minimal human presence. The question naturally follows: could medicine and clinical research move toward similar levels of machine autonomy, such as “dark operating rooms” or “dark clinical research hubs”?
This article does not argue whether such outcomes are desirable or imminent. Rather, the comparison highlights the importance of making deliberate choices about where autonomy is appropriate and where human oversight must remain central. In clinical endpoint adjudication, accountability for decisions, patient protection, and data stewardship remain non-negotiable.
Conclusion: A Practical Path Forward for CEA
Across this series, the recurring theme has been consistent: the value of AI agents in endpoint adjudication depends on process clarity, appropriate role alignment, and disciplined governance, not on technological novelty.
A practical way forward is to prioritise controlled deployments that respect confidentiality constraints, focus first on well-defined tasks, and build learning over time through structured feedback and oversight. This approach supports measurable operational gains while maintaining sponsor accountability and clinical integrity.
AI will continue to evolve. The most effective response is not to pursue maximal autonomy, but to apply AI agents where they can strengthen existing adjudication processes and reduce operational burden, without compromising responsibility, transparency, or control.
1MIT “The GenAI Divide state of AI in business 2025"
By Dimitri Stamatiadis, PhD, MBA
A consultant with extensive experience in clinical research in Europe and the USA, Dimitri has published numerous articles on drug development and the use of enabling technologies in the Pharma industry.
