If you read yesterday's post,
you already know the number.
377 health systems.
49 AI companies.
Zero logging standards.
This issue is about what lives
inside that gap.
Not as an abstract governance problem.
As a liability compounding right now,
quietly, inside every health system
that deployed AI before it built the room.
Three things happened recently
that belong in the same sentence.
Most people have not put them there yet.
I. The Shopping Cart Problem
Every Epic deployment has a moment
leaders remember differently
depending on their role.
For the CDO it was the go-live.
For the CMO it was the first
physician complaint.
For legal it was the first time
someone asked what the AI
actually recommended
and nobody could produce
a clean answer.
Epic's strength was always the same thing.
It built the infrastructure
before the capability.
The chart existed before the AI filled it.
The order set existed before
the algorithm suggested from it.
The workflow existed before
the recommendation landed in it.
That is a design philosophy.
It is why Epic is trusted
in a way standalone vendors are not.
But the shopping cart makes
a structural limit visible.
Emmy, Epic's AI, runs in two modes
most leaders have not formally separated.
The first is informational.
Emmy surfaces a risk score,
flags a pattern, writes a summary.
The clinician sees it, decides,
documents the decision.
The AI informed. It did not act.
The second is adjudicative.
Emmy queues an order.
Generates a prior authorization.
Fills a field that will carry
a clinician's signature.
The AI did not inform the decision.
It made a move inside the workflow
that a clinician must accept or override.
Most health systems deploying Epic AI
have not formally distinguished
between these two modes.
The shopping cart is the adjudicative
loop made visible.
Emmy queues the order.
The clinician clicks accept.
The order fires.
What gets documented is the outcome.
What never gets documented is whether
the clinician reviewed the reasoning,
modified it, or cleared the queue
because the day had 40 more patients in it.
That is not a technology failure.
That is a Handoff failure.
The informational loop is incomplete.
The adjudicative loop
was never formally built.
II. The Loop That Is Not Complete
The NEJM just published one of the first
large randomized trials of commercial
AI medical scribes in real practice.
Hundreds of clinicians.
Tens of thousands of encounters.
What worked: documentation time
dropped about 10%. Measurable
reductions in burnout and cognitive load.
Time back.
What did not: clinicians found
occasional clinically relevant inaccuracies.
Missing details. Pronoun mix-ups
that changed who did or experienced what.
One mild patient safety event was recorded.
Read that again.
One mild safety event
across tens of thousands of encounters.
That is not a headline number.
That is a design specification.
Because the question the trial
does not answer, the one no one
is asking in the coverage,
is what happened in the space
between what the scribe generated
and what the clinician signed.
Was there a formal review step.
An adjudication surface.
Any infrastructure between AI output
and clinical signature designed to catch
the pronoun that changed the patient's story.
There was not. There never has been.
The scribe produces. The clinician signs.
What happens in between is still
improvised at every health system
deploying this today.
That improvised space is the Handoff.
And it is where the liability lives.
One safety event in tens of thousands
sounds manageable until you multiply it
by the encounters your system runs a year.
Then multiply that by the number
of AI tools generating outputs
that carry a clinician's signature.
The loop is not complete.
Every day it runs incomplete
is another day the liability compounds.
III. The Prediction Engine
With No Adjudication Layer
This week Epic made a foundation model
called Comet available for testing
in its Cosmos AI Lab.
Trained on de-identified records
from 118 million patients.
Roughly 115 billion medical events.
Competitive on 78 clinical tasks,
including diagnosis prediction and
outcome forecasting, without fine-tuning.
The most powerful clinical prediction
engine ever built directly into an EHR.
It will generate outputs for every
Epic system that activates it.
And this week, separately, a coalition
of 35 institutions, including Harvard,
Google DeepMind, Stanford, Microsoft,
Epic itself, Kaiser, and the Gates
Foundation, published a proposal
for Medlog.
A global standard for logging every
instance of AI use in clinical care.
Nine elements per event.
Model identity. Inputs. Outputs.
Outcomes. Feedback.
They published it because
the infrastructure does not exist.
Epic built the most powerful clinical
prediction engine in history.
Nobody built the room where a clinician
can stand behind what it produced.
When Comet forecasts a diagnosis
and a clinician disagrees, what is
the infrastructure for that disagreement.
When the prediction is wrong and a patient
is harmed, what does the log show.
Right now the answer at every system
activating Comet is the same answer
it was before Comet existed.
Nothing. Because nobody built the log.
The Question Every Leader Needs
To Answer Before The Quarter Ends
The shopping cart queues the order.
The scribe generates the note.
The foundation model forecasts
the diagnosis.
In every one of those workflows,
a clinician puts their name on an output
they did not fully produce.
That is not a criticism of the technology.
That is how clinical AI works in 2026
across every system deploying it.
The liability does not live in the AI.
It lives in the gap between what the AI
produced and what the institution
can defend.
If the answer to the signature question
is that the signature is the documentation,
you have an exposure compounding with
every deployment you have not audited.
That gap has a name.
It is the Handoff.
And it is the one infrastructure problem
in healthcare AI nobody has solved yet.
You already know which of your deployments
runs this way. The output that becomes
a signature with nothing designed
in between.
The only question is whether you audit it
before the log you never built
is the one a regulator asks for.
MedicoVigilance™ is published every two weeks by Mo Johnson, MD MBA, founder of GPe Research. Each issue teaches one piece of the clinical AI accountability discipline your institution needs before the next deployment decision.
Forward this to a colleague whose institution is running on output no one has been named to own.