A story in six scenes
The scribe was
listening.
How AI walks into a medical practice without an invitation — and what it takes to walk it back. A composite story; every pattern in it is real.
9:04 AM
A Tuesday at a cardiology practice. The first patient of the day describes her symptoms in her own words — her history, her habits, her fears.
Everything is normal.
Scene two
The scribe is listening.
Six weeks ago, a doctor tried an AI scribe she saw in an ad. It writes her notes now. Nobody signed anything. It hears everything:
“chest tightness since March”
“my mother died of a heart attack at 61”
“I've been drinking more than I should”
“date of birth: ————”
“I haven't told my husband yet”
Every word is now data. Protected health information, by law — in motion, by design.
Scene three
Where does it go?
The exam room
where the words were spoken
The scribe vendor's cloud
audio + transcript, retained
A foundation-model provider
the subprocessor behind the vendor
Human QA reviewers
sampled transcripts, somewhere
Three organizations the practice has never audited now hold fragments of that conversation. There is no BAA that names them.
Scene four · The questions that decide everything
Is there a BAA that covers the prompt logs?
How long is the audio retained?
Is patient data training someone's model?
Can the audit log actually be exported?
Did anyone ask the patient?
The practice’s answer to all five: “we don’t know.”
Scene five · Some months later
U.S. Department of Health and Human Services
Office for Civil Rights
Re: Complaint No. ——— · Data Request
“…provide a copy of your most recent risk analysis, including documentation of risks associated with all systems that create, receive, maintain, or transmit ePHI…”
The first document OCR asks for is the risk analysis. The scribe isn’t in it — nobody knew it was there.
Scene six · The other path
None of this was inevitable.
The same clinic, run differently — not by banning the tools, but by being able to answer the five questions:
01
Find every tool
Shadow-AI discovery across browser, SaaS, and the EHR — sanctioned and not.
02
Map every flow
What goes in, where it's retained, who the subprocessors are.
03
Decide, in writing
BAAs that cover reality, an approved-tool catalog, a consent script the front desk can say.
04
Keep it current
A quarterly cadence — because the tools change faster than policies do.
The doctor isn’t the villain here.
She reached for a tool that gave her evenings back. The failure was structural: nobody was positioned to see the tool arrive, ask the five questions, and say yes, safely.
That position is buildable. In weeks, not quarters.
This is a composite story. No real patient, practice, or vendor is depicted — but every pattern in it (unsanctioned scribes, silent subprocessors, BAA gaps, the risk-analysis data request) comes from the enforcement record and from real assessments.