Human in the Loop Is Not Sufficient Without Decision Authority
Adeola Fujah
Corevident
AI

can accelerate clinical work faster than organizations can redesign the decisions around it. A human in the loop is not sufficient unless that person has clear authority to act, escalate, or stop the workflow.

At the DIA Global Annual Meeting 2026, an audience member from a smaller biopharmaceutical company walked to the open microphone at a session and described an artificial intelligence (AI) deployment that compressed clinical study startup from weeks to days. The protocol design team welcomed the speed. Their statistician was not amused. The statistician had planned to the old timeline and was not ready to begin. AI had accelerated the workflow, but the operating model around it had not caught up.

This same pattern shows up wherever an automated system produces a deliverable faster than the receiving team was organized to handle it. Authority questions appear when AI flags:

  • a monitoring anomaly
  • a protocol deviation
  • a safety signal
  • a data discrepancy.

These events can be flagged almost instantaneously, but the person looking at the flag often does not know what they are authorized to do next. Speed without authority creates risk.

Regulators are converging on the same expectation: the US Food and Drug Administration (FDA) and the European Medicines Agency (EMA) jointly published Guiding Principles of Good AI Practice in Drug Development in January 2026, emphasizing human responsibility and oversight of AI. Industry’s shorthand for this expectation is to “keep a human in the loop.” But naming a human is not the same as designing what that person is authorized to decide and is ready to act upon. Unclear decision authority is not new, nor is it unique to AI. Basic automation forced the same questions in regulated industries years ago; AI simply compresses the time organizations once had to work around it.

“Authority” in this context means being explicitly given the right to decide, escalate, or stop, and knowing the rationale for each. Authority is one component of the operating model (alongside roles, timelines, and capability), but it is the component that organizations most consistently fail to design, and the component this article examines.

The Authority Gap Was Already the Hard Part

The evidence comes not from a clinical study but from regulated financial services. The banking case study: automating the manual re-entry of data from one system to another. Clinical teams are familiar with this type of work: Moving the same study information between data capture, trial management, and electronic trial master file (eTMF) systems, all with inconsistent versions of current state.

In the banking example:

  • Before: 2,500 transactions per month moved manually from a legacy system into a service management platform, requiring roughly five minutes of manual effort per transaction.
  • Intervention: Automated the repetitive, rules-based transfer, and deliberately redesigned the human role to actively support the automation as needed, not merely leave it unchanged.
  • Redesigned role: The operations specialist monitored exceptions and was authorized to resolve defined cases, escalate others, or stop the process when the routine path no longer applied.
  • Outcome: About 208 hours of routine manual transfer work removed each month, roughly 1.2 full-time equivalents of capacity returned annually.

In this example, the flagged exception that the bank operations specialist was authorized to handle can be compared to a follow-up queue in a trial that lists “CRA” as the final owner of every flag, without specifying whether the remote or onsite clinical research associate (“CRA”) owns the resolution. The system simply found the problem, but not the decision owner who can solve it.

The objective of human oversight as a governance mechanism is not to preserve manual work or require a person to click approve or reject. It is to place the appropriate human experts where judgment, authority, accountability, and contextual understanding are genuinely required. There was no AI in this banking system, yet deciding what the operations specialist was authorized to do when the automation flagged an exception was still the hard part.

While this banking case does not fully illustrate a clinical operating model, it highlights a decision pattern that clinical organizations must solve under more complex conditions. The price of weak authority design is categorically higher in a trial. A mishandled transaction exception at a bank is money, remediation, maybe a reporting obligation. A mishandled safety signal is potentially a patient death.

What Transfers, and What Does Not

The governance approach in this banking example cleanly transfers to clinical research, but the operating environment does not. In the banking case, the automation flagged a transaction it could not complete, and the operations specialist monitoring it had to be authorized to resolve, escalate, or stop it. If an eTMF dashboard flags hundreds of “missing” documents in a study, someone must be authorized to determine which alarms are real. In many organizations, that determination is re-explained in monthly communications and meetings, but no one owns the authority to ensure that the outcome and its accompanying evidence are captured for posterity.

A flagged exception in a study may also move across sponsor, contract research organization (CRO), site, investigator, and vendor boundaries, making authority design far more complex than in a single-organization setting. A clinical trial distributes responsibilities across many parties while remaining subject to external regulatory inspection, so, even though both industries are highly regulated, the clinical trial industry cannot simply import the banking playbook. It must build its authority construct natively.

Authority moments are prevalent across clinical operations: A central monitoring anomaly whose significance must be determined; a detected protocol deviation that a clinical trial manager (CTM) must assess and close; a safety signal to act on, escalate, or stand down; or a data discrepancy to resolve or query.

Audit trails may show what changed, when, and by whom. Evidence chains connect these changes to the CTM’s authority, the information considered, the rationale for the decision, and any required approval or escalation. When evidence chain elements do not travel with the decision, the organization must reconstruct them later. Reconstruction is typically more time-consuming and expensive than capturing evidence in real time. Parallel trackers are often needed when systems record an event but do not preserve the decision context, ownership, and evidence required to manage or explain it later. Introducing new tools before designing decision authority creates risk, leaving teams to maintain manual trackers that reconstruct state, ownership, and evidence after the fact.

Every authority decision in a trial is potentially inspectable. The evidence of who was authorized, who decided, and on what basis is not just internal hygiene but material that may have to be produced years later to an external party (a regulator) with enforcement power.

Pilot Paralysis and the Human Angle

The prevailing concern in public forums where clinical research professionals engage in open discourse is not fear of technology. Practitioners describe being asked to give free domain expertise to tool builders who do not even know what GCP (Good Clinical Practice) stands for. Site staff describe having to learn and juggle more than a dozen “easy” systems, each one a sponsor’s customized enterprise solution. A common refrain is that the people who understand the work must be in the room when the tools that will reshape that work are designed.

It stands to reason that AI pilots can stall when the people expected to operationalize them cannot see a clear, supported role for themselves in the new workflow. Practitioners are asking whether they are being “set up” to be accountable for something they are not authorized to control. “Pilot paralysis” is often an operating model and change management failure, not a technology failure. One example is AI-supported signal detection. It can generate follow-on investigative work, particularly when teams have not defined who reviews the signal, what they are authorized to decide, and when escalation is required. Teams see value in automation only when response roles and follow-up ownership are clearly designed upfront.

Two Halves of One Equation

Authority is one half of the equation. But to truly empower teams, there must also be capability. Practitioners (whether in industry or at clinical sites) need access to approved tools and safe environments in which to experiment, explore role-specific learning scenarios connected to real workflows, and challenge AI outputs with confidence. Structured experiments, including hackathons using approved models, can bring domain, technical, quality, and operational expertise together. As AI produces more output, human expertise becomes even more important because practitioners must judge context, consequences, and appropriate action.

Designed authority is not bureaucracy, but it is what empowers teams to move fast, because they are equipped with answers before the government inspector asks any questions.

Build the Transition Gate

To break free of “pilot paralysis” and effectively scale AI pilots into controlled use, organizations must establish a formal transition gate that requires the team to ask and answer:

  1. Whose work and timelines change, and are they equipped and ready?
  2. What may AI produce or recommend?
  3. Who may accept, challenge, override, resolve, or escalate, to whom, and within what timeframe?
  4. What evidence must accompany each decision?
  5. Who can stop the workflow, and who is authorized to restart it?

Conduct this design work WITH the practitioners who will hold the authority, not TO them. For example: If the study startup acceleration had been collaboratively designed using these transition gate questions with the statistician in the room, their timeline concerns would have surfaced early and the team could have adjusted the operating model around them before the pilot stalled.

In clinical research, AI governance is not achieved by simply adding a human approval step after the system produces an answer. It is achieved when the appropriate qualified experts have the authority, capacity, capability, and evidence to act when judgment is required. That is how organizations turn AI acceleration into controlled progress.

Declaration of AI Use: During the preparation of this article, the author used Claude (Anthropic) to help structure the outline, suggest edits, and refine successive drafts. All content, evidence, and final text and edits are the author’s own, and the author takes full responsibility for the content of this publication.