I couldn’t agree more with this distinction between “Human in the Loop” and “Human at the Helm” (MERGETIC, 10/23/2026). The latter captures what thoughtful AI governance should look like: clinical expertise, scientific judgment, and human accountability built into the system—not added as safeguards at the end.

This principle is not new. Across more than 12 years in clinical data management, oncology real-world data, and clinical research operations, I’ve helped design and lead human-centered workflows, including AI-assisted clinical abstraction. The key lesson: clinical and research expertise must shape the system before the first record is processed.

The strongest approaches I helped develop brought human abstractors—including clinician abstractors—experienced clinical data managers, and clinical research professionals into design alongside technical experts. Particularly for abstractors working directly with source records, understanding the research questions, clinical context, data limitations, and intended use of the evidence is indispensable. Their expertise shaped abstraction logic, prompts and extraction specifications, decision rules, validation, quality controls, exception handling, and release criteria. They werent merely checking outputs; they helped define what the system needed to recognize and distinguish. Why does that matter? High sensitivity, specificity, or F1 scores alone do not establish scientific validity or fitness for purpose. A model evaluated against an incomplete or incorrectly labeled reference standard can score well while missing clinically consequential distinctions: historical versus active disease, suspected versus confirmed diagnoses, or treatment intent versus actual administration. Technical performance cannot rescue flawed clinical definitions or an unrepresentative validation set.

I become concerned when complete removal of humans is presented as the ultimate measure of AI maturity. It risks treating expertise as an obstacle to efficiency rather than a design input that enables responsible automation. The better question is where human judgment adds the most value and prevents consequential errors—not how quickly it can be eliminated.

Nor must this slow innovation. Embedding clinical and research expertise upfront can reduce rework, repetitive manual review, and costly corrections. Human governance does not mean a person must manually approve every output; it means qualified people establish the rules, thresholds, escalation paths, and release conditions appropriate to the use case.

In clinical research, including studies using real-world data (RWD), the goal is not simply accurate extraction or impressive model metrics. It is traceable, reproducible, clinically meaningful, fit-for- purpose data that can support defensible scientific and regulatory decisions.

The most mature AI systems are not those with the fewest humans, but those in which human expertise governs the decisions that matter most. Inspiration and reference: This reflection was prompted by MERGETIC’s October 23, 2026 LinkedIn post discussing “Human in the Loop” and “Human at the Helm.” The clinical research perspectives and examples are my own. MERGETIC. (2026, October 23). [LinkedIn post comparing human-in-the-loop and human-at-the-helm approaches]. LinkedIn. https://lnkd.in/p/ej8ZVTrb