The Supervision Trap
The 4:15 PM Ghost
It’s 4:15 PM on a Tuesday. You are three patients behind, and your Inbox is humming. It has been a busy digital assistant today—scanning labs, drafting portal messages, and suggesting titration for your chronic care patients.
You see a pre-written message to a 72-year-old with heart failure: "Your labs look great! We are increasing your lisinopril to 20mg as planned." With a waiting room full of people, you click "Batch Send" to clear the queue.
Two days later, that patient is in the ER with acute kidney injury. The Agentic AI model saw a "normal" potassium but missed the subtle 0.3 creatinine "creep" over the last month—a nuance you would have caught in five seconds if you weren't "supervising" fifty automated decisions in the frantic margins of a ten-minute physical exam.
The Historical Parallel: A Lesson Unlearned
We have been here before. Twenty years ago, the "Mid-Level" revolution promised to extend the physician’s reach. But for many Primary Care Physicians, it became a game of Legal Shielding.
We were never taught how to supervise. There was no "Residency in Oversight." We were simply told that our signature at the bottom of a chart—even one we barely glanced at—meant the care was safe. We were expected to be in two places at once: maintaining our own clinical excellence while acting as a "magical" legal backstop for decisions we didn't personally make.
The New Challenge: The Agentic AI Model
The industry expects the same magical supervision for the Agentic AI model. They promise that AI will "reduce the burden," but in reality, it just shifts the burden.
We are moving from doing the work to auditing the work. And as any clinician knows, auditing a "black box" is often more cognitively taxing than just seeing the patient yourself. It requires you to reverse-engineer a machine’s logic while the clock is ticking.
The Proposal: From Ghost to Pilot
If we repeat the mistakes of the past—where supervision was an afterthought—we are creating a more efficient way to fail our patients. We need a new model of Algorithmic Stewardship:
Forensic Training: "Informatics" can no longer be an elective. Physicians need formal training in identifying where AI trips on nuance—like the difference between a "stable" lab and a "trending" catastrophe.
Audit-Protected Time: Supervision is work. If a physician is responsible for the output of an Agentic AI model, that time must be carved out of their clinical quota. You cannot safely audit an algorithm in the "white space" between appointments.
Tiered Autonomy: We must define the "Red Lines." Administrative tasks? Let the agent run. Clinical adjustments in a multi-morbid elderly patient? The AI stays in "Read-Only" mode.
Closing Thoughts
We are at a crossroads. We can continue to be the "ghosts" in a machine we don't control, or we can demand the training and the time to be true Chief Pilots.
I leave you with these three questions:
If we are held legally responsible for an Agentic AI model’s output, at what point does "supervision" become an impossible cognitive burden that outweighs the time the AI actually saves us?
If we move from a profession of doers to a profession of auditors, will the next generation of physicians ever develop the "clinical gut feeling" necessary to know when the machine is wrong?
Where do we, as a collective profession, draw the absolute Red Line for autonomy—tasks that an AI agent should never be allowed to execute, regardless of how high its confidence score might be?