Atlanta Perinatal Associates · Clinical AI Series
From Dictation to Ambient Intelligence
Dragon NaturallySpeaking trained you to talk like a machine. Ambient AI learned to listen like a clinician. Here is what changed after 2022 and why it matters for your patients.
CO
Chukwuma Onyeije, MD, FACOG
Maternal-Fetal Medicine · Atlanta Perinatal Associates
01 / 09
Pre-2022 Era
The Dragon Years
An acoustic model that heard your phonemes but did not understand your meaning. You adapted to it. It did not adapt to you.
What it cost you
🎤
You changed how you speak
Slower. Flatter. Announcing every comma and period. Non-American accents required months of retraining.
🖥️
You turned away from the patient
The screen demanded your attention. Documentation was a competing task, not a byproduct of the encounter.
🧠
Every word required a command
Dragon transcribed. It did not infer structure, populate sections, or follow clinical logic.
02 / 09
The Inflection Point
2022 Changed the Architecture of the Problem
Large language models did not just improve transcription accuracy. They added comprehension. The system stopped hearing words and started understanding clinical context.
🎙️
Pre-2022Acoustic models. Phoneme matching. You speak, it transcribes.
→
🤖
2022–2023Transformer-based ASR. LLM post-processing. Structure inferred.
→
🩺
NowAmbient scribes. Passive listening. Note is a byproduct of care.
03 / 09
How It Works
Ambient Scribes: The Clinical Workflow
01
You enter the roomThe ambient system activates. No microphone to position, no commands to speak.
02
You talk to the patientThe AI listens to the encounter in context. It understands who is speaking and what role each utterance plays in the clinical narrative.
03
A structured draft appearsAssessment, plan, HPI, ROS — populated according to clinical logic, not dictation sequence. You review and sign.
Platforms
🩺 Heidi AI
🔵 Nuance DAX
📋 Abridge
🎧 Suki
📝 Nabla
🔬 DeepScribe
04 / 09
Why It Matters Clinically
Four Benefits That Change Patient Care
👁️ Gaze Direction
When the scribe listens passively, the physician faces the patient. Full eye contact. No keyboard between a family and a serious diagnosis.
📊 Note Accuracy
Reconstructing a conversation an hour later compresses and distorts. Ambient capture during the encounter produces a more faithful record.
⏱️ After-Hours Burden
Notes completed at 9 PM on a Saturday represent a different kind of documentation failure. Ambient systems close that gap at the point of care.
🧠 Cognitive Load
Dragon moved load from fingers to throat. Ambient moves it out of the encounter. Clinical reasoning is no longer competing with documentation management.
05 / 09
You Already Have Access
Ambient Transcription Is Now Mainstream
Enterprise platforms are not the only entry point. These tools are already on your devices today.
For physician-developers: the transcription layer is commoditized. Apple, Google, and Microsoft solved it. The clinical intelligence layer — SOAP/APSO structuring, ICD-10 inference, follow-up flagging — is where the build opportunity lives.
06 / 09
Honest Assessment
What Ambient Scribes Do Not Yet Solve
These are drafting systems. The physician still reviews, edits, and signs. That responsibility has not changed. It should not change.
⚠️
Hallucination RiskAmbient systems occasionally generate clinical content that was not spoken. Every note requires genuine review, not rubber-stamping.
👥
Multi-Speaker EnvironmentsBusy rooms with nurses, students, and family members can confuse speaker attribution and clinical intent.
🔬
Subspecialty VocabularyMFM, neonatology, and other subspecialties use technical language that some general-purpose ambient systems handle imprecisely.
🔒
HIPAA ComplianceVerify your platform's BAA, data retention policy, and de-identification approach before clinical deployment.
07 / 09
Practical Path Forward
How to Start Experimenting Today
STEP 1
Start with Google Docs Voice Typing or Word Dictate — use it to draft a non-patient note or a clinical letter. Get a feel for ambient transcription quality without any HIPAA exposure.
STEP 2
Explore Heidi AI or Nabla free tiers — both offer trial access. Run a simulated encounter. Evaluate structured output quality against your documentation standard.
STEP 3
Evaluate enterprise options with your EHR vendor — Nuance DAX integrates with Epic and Dragon Medical One. Ask about BAA terms and pilot access before committing.
STEP 4
Read, review, sign — treat every AI-generated draft as a first draft requiring your clinical judgment. The note carries your signature and your license.
08 / 09
The Bottom Line
Dragon taught us to speak to machines. The machines finally learned to listen.
Ambient AI is the most clinically meaningful documentation development of the past decade. Not because it saves time, though it does. Because it keeps the physician present with the patient where the care actually happens.
See more at https://www.doctorswhocode.blog/blog/dictation-before-after-llm/
Chukwuma Onyeije, MD, FACOG · Atlanta Perinatal Associates · DoctorsWhoCode.blog
09 / 09