I like AI. I use it. I trust it more than I probably should.
Which is why this story still bugs me.
Earlier this year, I followed my primary care doctor from Mass General Brigham to Beth Israel. She’s great. I’ve seen her for more than 25 years. When she moved, I moved. Simple.
A few months later, I booked a routine annual physical. Nothing special. Same checkup, different logo on the building.
During the visit, she asked if she could transfer my records from the old Epic system to the new one. Sure. Then she asked if she could use AI to take notes during the appointment.
Also sure. That sounded efficient. Modern. Sensible.
As she reviewed my history, she talked through past issues out loud – things like a deviated septum surgery from a decade ago, Bell’s palsy, old test results. I nodded along. No new complaints. No active issues. Just context while data moved from one system to another.
Physical done. Labs drawn. I left.
Weeks later, I got a bill.
The preventive physical was covered. The labs were covered. But there was a separate line item: “Office Visit” – $337. My responsibility.
Here’s what happened: the AI-generated notes captured a detailed medical discussion. Those notes were later used to code the visit. The system interpreted historical context as an active evaluation. No human stepped in to say, “This was just a physical.”
So I entered the maze.
Insurance said, “We just process what we’re sent.”
Billing said, “We don’t decide coding.”
The doctor’s office said, “Different department.”
Four months later – after a formal grievance to Blue Cross – the charge was waived. No apology. No explanation. Just… fixed.
This is the quiet risk with AI in complex systems. It doesn’t fail loudly. It fails procedurally. Everything looks correct. The notes are accurate. The workflow is followed. And the outcome is still wrong.
Healthcare is already hard to navigate. Adding automation without safeguards doesn’t simplify it – it just hides the sharp edges better.
What I’d Do Differently Next Time
If I could replay the visit, I wouldn’t say no to the AI. I’d say “yes, with guardrails.”
I’d be explicit about intent: “This is a preventive physical. We’re reviewing history for transfer, not evaluating active issues.”
I’d also ask one slightly uncomfortable question at the end: “Just to confirm – this is being coded as preventive only, right?”
That’s not something patients should have to do. But right now, it’s the only way to keep context from getting lost between systems, software, and departments that don’t really talk to each other.
AI is powerful. But context still matters. And when no one owns the context, the patient does – usually with their wallet.
Until the systems catch up, the safest move is still a human pause and a clarifying sentence – spoken out loud, not buried in notes.
