Architecture Maps
❦ ❧ ❦

Augmedix

Ambient AI scribe on Google Cloud — Commure’s ambient documentation engine, illuminated as a parchment manuscript.

Public Sources Only Acquired by Commure · Oct 2024 · $139M 1,789+ Clinicians (Q4 2023) 70K+ Notes / Week Updated May 2026
Chapter I

Overview

Augmedix is the ambient AI documentation engine inside Commure. Founded in 2012 at Stanford on the premise of streaming a clinician’s first-person view to remote scribes in Dhaka, it evolved through Google Glass, hybrid AI-assisted scribing, and finally a fully automated multi-LLM stack on Google Cloud. Following the October 2024 take-private, Augmedix is rebranded under the “Commure Ambient” banner and serves as the documentation front end of Commure’s end-to-end clinical platform.

1,789+
Clinicians (Q4 2023)
$44.9M
2023 Revenue · +45% YoY
70K+
Notes / Week Corpus
800+
Bangladesh Staff (~500 Scribes)
$139M
Commure Acquisition (Oct 2024)
50+
Specialties Supported

What it is, in one paragraph

An iOS app on the clinician’s phone (or a Bluetooth lab-coat mic) captures the encounter audio. Audio streams to Google Cloud, where Speech-to-Text produces a diarized transcript, then a multi-pass NLP/LLM pipeline on Vertex AI — using MedLM, MedGemma, Gemini, and Augmedix-fine-tuned models — drafts a clinical note across HPI, PE, Assessment & Plan. Depending on product tier, a Medical Documentation Specialist in Bangladesh or India polishes the draft inside NoteBuilder before it is written back to the EHR (Epic, Oracle Health, athenahealth, MEDITECH) as a FHIR DocumentReference. The clinician reviews and signs.

The trajectory in one paragraph

Augmedix’s architectural story is a gradient: from 100% human (Glass → scribe) in 2014, to AI-assisted scribes (NoteBuilder, 2021), to fully autonomous AI drafting (Go, September 2023), to ED-specialized AI (Go ED, April 2024), and finally to a deeply integrated layer in Commure’s revenue-cycle platform — where ambient notes feed autonomous coding, CDI, and claims submission. Across that arc the GCP substrate has remained constant: STT, Vertex AI, Pub/Sub, Spanner, BigQuery, FHIR API.

“Documentation is a wound on the practising physician.” — the founding insight: physicians were spending roughly 35% of their day charting. Augmedix grew from that observation by Pelu Tran during his Stanford clinical rotations and Ian Shakil’s wager that Google Glass + remote scribes in Dhaka could close the gap.
Chapter II

The Six Products

Augmedix’s portfolio is a gradient along a single axis: human-scribe-heavy on one end, fully autonomous AI on the other. All six products share the same NoteBuilder/Vertex AI substrate; what differs is how much human labor sits in the loop and where in the encounter the work happens. Post-acquisition all six are marketed under “Commure Ambient.”

Diagram 1 — Product Gradient (Human → AI)
graph LR
    PREP["Augmedix Prep
(pre-visit chart
preparation, Sep 2022)"] LIVE["Augmedix Live
(real-time human MDS,
2014 origin)"] NOTES["Augmedix Notes /
NoteBuilder
(AI draft + MDS, 2021)"] ASSIST["Augmedix Assist
(AI + 1-hr MDS polish)"] GO["Augmedix Go
(fully automated AI,
Sep 2023)"] GOED["Augmedix Go ED
(ED-specialized AI,
Apr 2024)"] PREP --> LIVE LIVE --> NOTES NOTES --> ASSIST ASSIST --> GO GO --> GOED style PREP fill:#c9b87a,stroke:#704214,color:#2a1f1a style LIVE fill:#a8533a,stroke:#722f37,color:#f4e4c1 style NOTES fill:#8b6914,stroke:#4a3728,color:#f4e4c1 style ASSIST fill:#704214,stroke:#4a3728,color:#f4e4c1 style GO fill:#8b3a4a,stroke:#4a3728,color:#f4e4c1 style GOED fill:#722f37,stroke:#2a1f1a,color:#f4e4c1
Reading the gradient

Leftmost (Prep, Live) is the most human-labour-intensive: a Medical Documentation Specialist watches the encounter live or builds context ahead of it. Middle (Notes, Assist) is the hybrid tier: AI drafts, humans polish — the original NoteBuilder workflow. Rightmost (Go, Go ED) is fully autonomous: AI drafts the note within moments, the clinician reviews and signs, no human scribe in the loop.

Product Launched Human in Loop? Delivery Time Claimed Savings
Augmedix Prep Sep 2022 MDS QA on pre-built note structure Before visit Chart-prep burden reduced
Augmedix Live 2014 (Glass era) MDS watches encounter in real time Signature-ready immediately post-visit 2–3 hrs/day
Augmedix Notes / NoteBuilder Apr 2021 MDS finalizes AI-assisted draft Same day Per-note labour cut
Augmedix Assist ~2023 MDS edits AI draft + coding Within 1 hour 1.5–2 hrs/day
Augmedix Go Sep 27, 2023 None — clinician reviews directly Within moments ~1 hr/day
Augmedix Go ED Apr 24, 2024 None — first fully automated ED ambient AI Continuous across multi-touch ED visits 99% patient consent in HCA pilot

Go · The flagship

iOS app, iPhone 12+, iOS 15+. Multi-recording stitching used by 50% of clinicians. CareCues real-time care-gap alerts. ICD-10 suggestions. Problems Page. 32-language support. visionOS 1.0+ supported.

Go ED · The hardest case

Purpose-built for emergency departments: multi-hour, non-sequential interactions with one patient, high ambient noise, documenting reasoning across re-evaluations. Piloted at four HCA hospitals.

Live · The original

Highest-touch tier. MDS observes encounter, two-way chat with clinician, signature-ready note immediately. Includes prior-auth assistance, referral letters, AVS drafting, CPT/E&M/HCC coding.

NoteBuilder · The brain

The proprietary internal tool that all human-in-loop tiers run on. ASR + NLP surface dynamically filtered structured suggestions by section (HPI, ROS, PE, A&P) and specialty. Medication datasets integrated.

Prep · The pre-visit layer

Pre-populates patient demographics, med changes, history, labs/imaging, immunizations, family/social. Physician walks into the room with a pre-built note structure. McKesson Ventures highlighted as strategic.

Structured Data · The exhaust

Beyond notes, Augmedix exports three artefacts per encounter: full transcript, clinical note (pre and post physician edit), and structured clinical data — with optional PHI de-identification on the export path.

Chapter III

Capture Pipeline — End-to-End

A single clinical encounter traces a consistent path from the clinician’s pocket microphone to the signed EHR note. The specific hop count varies by product tier (Go skips the human-scribe queue), but the skeleton is the same: capture → transcribe → reason → (optionally polish) → write back → sign.

Diagram 2 — The Encounter Lifecycle
sequenceDiagram
    participant C as Clinician
    participant APP as iOS App
+ Bluetooth Mic participant STT as Google STT
(diarization) participant VAI as Vertex AI Pipeline
(MedLM + MedGemma + Gemini) participant MDS as MDS Queue
(Live / Assist tiers) participant NB as NoteBuilder QA participant EHR as EHR Write-Back
(FHIR DocumentReference) participant SIGN as Clinician Sign-Off C->>APP: Speak with patient APP->>STT: Stream audio STT->>VAI: Diarized transcript VAI->>VAI: Multi-pass section drafting alt Live / Assist tier VAI->>MDS: AI draft to scribe queue MDS->>NB: Human polish in NoteBuilder NB->>EHR: Upload pending note else Go / Go ED tier VAI->>EHR: AI draft direct end EHR->>SIGN: Note in pending status SIGN->>EHR: Final signed note
Diagram 3 — Capture-to-Sign Linear Flow
graph LR
    CLIN["Clinician"]
    APP["iOS App
(Augmedix Go)"] MIC["Bluetooth Mic
(lab coat pocket)"] STT["Google Cloud
Speech-to-Text"] NLP["NLP Extraction
(specialty + section)"] LLM["Multi-LLM Pass
(MedLM / MedGemma /
Gemini + cross-check)"] MDS["MDS Queue
(Bangladesh / India,
Live + Assist only)"] NB["NoteBuilder QA
(structured suggestions,
med datasets)"] EHR["EHR Write-Back
(FHIR DocumentReference
or section-level)"] SIGN["Clinician
Reviews & Signs"] CLIN --> APP MIC --> APP APP --> STT STT --> NLP NLP --> LLM LLM --> MDS LLM --> EHR MDS --> NB NB --> EHR EHR --> SIGN style CLIN fill:#5a4633,stroke:#2a1f1a,color:#f4e4c1 style APP fill:#6b8e6f,stroke:#2d4a3e,color:#2a1f1a style MIC fill:#6b8e6f,stroke:#2d4a3e,color:#2a1f1a style STT fill:#4a6b8a,stroke:#1e3a5f,color:#f4e4c1 style NLP fill:#8b3a4a,stroke:#722f37,color:#f4e4c1 style LLM fill:#8b3a4a,stroke:#722f37,color:#f4e4c1 style MDS fill:#a8533a,stroke:#722f37,color:#f4e4c1 style NB fill:#a8533a,stroke:#722f37,color:#f4e4c1 style EHR fill:#8b6914,stroke:#4a3728,color:#f4e4c1 style SIGN fill:#5a4633,stroke:#2a1f1a,color:#f4e4c1
The retention window

Per the Google Cloud Developer Community article (June 2024), Augmedix retains encounter audio and note data for one week after EMR transmission before permanent deletion. This is shorter than the “up to six years after account closure” window stated in the general privacy policy — the one-week figure is specifically for the audio + raw transcript artefacts under the BAA. The shorter retention is part of the BAA boundary that lets identifiable PHI flow through Vertex AI inference.

Connectivity is non-negotiable

The Go product schedule confirms that the app requires continuous internet connectivity during use; audio processing is server-side, not on-device. The Live schedule additionally requires WPA2 PSK (AES) or WPA2 PEAP wireless authentication on the clinical network with a minimum −65 dBm RSSI and QoS elevated to platinum/voice priority.

Chapter IV

The Google Cloud Substrate

Augmedix is one of the deepest single-vendor Google Cloud deployments in healthcare AI. The richest public technical source — the Google Cloud Developer Community article from June 2024 — enumerates the services by layer: ASR/AI, messaging, storage, compute, analytics. SADA serves as the integration partner; Augmedix maintains a Google Cloud Marketplace listing so health systems can procure through GCP billing.

Diagram 4 — GCP Services by Layer
graph TD
    subgraph CAPTURE["Capture Layer"]
        IOS["iOS App
(Augmedix Go)"] end subgraph ASRAI["ASR & AI Layer"] STT["Speech-to-Text
(diarization)"] VAI["Vertex AI"] MEDLM["MedLM
(Med-PaLM 2 lineage)"] MEDG["MedGemma"] GEM["Gemini
(multimodal)"] end subgraph MSG["Messaging Layer"] PUBSUB["Cloud Pub/Sub
(encounter topics)"] end subgraph COMPUTE["Compute Layer"] GKE["GKE
(encounter_service)"] CRUN["Cloud Run"] end subgraph STORAGE["Storage Layer"] SPAN["Cloud Spanner
(encounter tables)"] BQ["BigQuery
(analytics warehouse)"] HCFHIR["Healthcare FHIR API
(R4)"] HDE["Healthcare Data
Engine (HDEv2)"] end subgraph ANALYTICS["Analytics Layer"] LOOK["Looker
(dashboards)"] COLAB["Vertex AI Workbench /
Colab Enterprise"] end IOS --> STT STT --> VAI VAI --> MEDLM VAI --> MEDG VAI --> GEM VAI --> PUBSUB PUBSUB --> GKE GKE --> SPAN SPAN --> BQ HCFHIR --> HDE HDE --> BQ BQ --> LOOK BQ --> COLAB style IOS fill:#6b8e6f,stroke:#2d4a3e,color:#2a1f1a style STT fill:#4a6b8a,stroke:#1e3a5f,color:#f4e4c1 style VAI fill:#8b3a4a,stroke:#722f37,color:#f4e4c1 style MEDLM fill:#8b3a4a,stroke:#722f37,color:#f4e4c1 style MEDG fill:#8b3a4a,stroke:#722f37,color:#f4e4c1 style GEM fill:#8b3a4a,stroke:#722f37,color:#f4e4c1 style PUBSUB fill:#c9b87a,stroke:#704214,color:#2a1f1a style GKE fill:#704214,stroke:#4a3728,color:#f4e4c1 style CRUN fill:#704214,stroke:#4a3728,color:#f4e4c1 style SPAN fill:#4a6b8a,stroke:#1e3a5f,color:#f4e4c1 style BQ fill:#4a6b8a,stroke:#1e3a5f,color:#f4e4c1 style HCFHIR fill:#4a6b8a,stroke:#1e3a5f,color:#f4e4c1 style HDE fill:#4a6b8a,stroke:#1e3a5f,color:#f4e4c1 style LOOK fill:#8b6914,stroke:#4a3728,color:#f4e4c1 style COLAB fill:#8b6914,stroke:#4a3728,color:#f4e4c1
Service Layer Role
Speech-to-TextASRDiarized transcript from raw encounter audio
Vertex AIAI platformHosts MedLM, MedGemma, Gemini; fine-tuning runs for Augmedix-proprietary models
MedLM on Vertex AIFoundation modelsMedical foundation models built on Med-PaLM 2 lineage; early-tester since Dec 2023
MedGemmaFoundation modelsMedical-specific Gemini-derivative; confirmed in Google Dev article as active in the Clinical Notes Engine
GeminiFoundation modelsMultimodal LLM supporting advanced note processing
Cloud Pub/SubMessagingAsynchronous publish of structured encounter JSON to customer topics
GKEComputeHosts encounter_service container that subscribes to Pub/Sub and writes to Spanner
Cloud RunComputeAlternative serverless deployment of encounter_service
Cloud SpannerStorageRelational encounter table storing structured clinical data
BigQueryStorage / analyticsCombines Augmedix encounter data with longitudinal patient records for analytics
Healthcare FHIR APIStorageFHIR R4 patient record store; RESTful patient data access
Healthcare Data Engine (HDEv2)Storage / harmonizationAggregates and standardizes multi-source data to FHIR R4 (private preview as of June 2024)
LookerAnalyticsPopulation dashboards, care-gap alerts, clinical recommendations
Vertex AI Workbench / Colab EnterpriseAnalyticsUsed by health-system AI researchers for custom model dev on the encounter corpus

SADA · Integration Partner

SADA — a Google Cloud premier partner — was named (April 2024) as Augmedix’s scaling partner for ambient documentation and structured data products to health systems nationwide.

Google Cloud Marketplace

Augmedix maintains a Marketplace listing at console.cloud.google.com/marketplace/product/augmedix-public/augmedix, enabling health systems to procure Augmedix through GCP billing.

BAA Boundary

Google’s HIPAA BAA covers Vertex AI as a HIPAA-eligible service. Per the BAA, Vertex AI does not use customer data, prompts, or responses to improve or train the foundation models — the legal basis that lets identifiable PHI flow through inference.

Enterprise Data Lake Path

For enterprise customers, Augmedix publishes structured JSON to a customer-defined Pub/Sub topic; the customer’s encounter_service on GKE or Cloud Run subscribes and lands data in their own Spanner / BigQuery. Available within 30 minutes of EHR transmission. 99.0% uptime SLA.

Chapter V

EHR Integrations

Augmedix supports bidirectional integration with four named EHR platforms via two protocols: HL7 v2 (for schedule pulls) and FHIR R4 (for encounter context and DocumentReference write-back). Specifics vary by EHR. By the time of the Vizient contract (January 2025), Augmedix advertised “50+” EHR integrations; the named four below are the canonical relationships with public certification.

Diagram 5 — EHR Integration Topology
graph LR
    AUGX["Augmedix
encounter_service"] HL7["HL7 v2
(schedule pulls,
ADT/SIU feeds)"] FHIR["FHIR R4
(encounter context,
DocumentReference)"] EPIC["Epic
(Showroom Connection Hub,
listing 1103)"] ORACLE["Oracle Health
(Validated Integration,
Dec 2024)"] ATHENA["athenahealth
(athenaConnect,
section write-back)"] MEDI["MEDITECH
(Expanse Now embedded,
Sept 2025)"] AUGX --> HL7 AUGX --> FHIR HL7 --> EPIC FHIR --> EPIC FHIR --> ORACLE FHIR --> ATHENA FHIR --> MEDI style AUGX fill:#5a4633,stroke:#2a1f1a,color:#f4e4c1 style HL7 fill:#c9b87a,stroke:#704214,color:#2a1f1a style FHIR fill:#c9b87a,stroke:#704214,color:#2a1f1a style EPIC fill:#8b6914,stroke:#4a3728,color:#f4e4c1 style ORACLE fill:#8b6914,stroke:#4a3728,color:#f4e4c1 style ATHENA fill:#8b6914,stroke:#4a3728,color:#f4e4c1 style MEDI fill:#8b6914,stroke:#4a3728,color:#f4e4c1
EHR Pull Protocol Write-Back Mechanism Certification
Epic HL7 v2 schedules + FHIR R4 encounter context FHIR DocumentReference / SmartText / Smart Data Elements inside Hyperspace/Hyperdrive Epic Showroom Connection Hub tier — listing id 1103 ($500/yr)
Oracle Health (Cerner) HL7 + FHIR Validated cross all three Augmedix product tiers (Go, Assist, Live) Oracle Validated Integration — December 2024; 10+ year relationship
athenahealth FHIR APIs (schedule + encounter) Section-level write-back: HPI / ROS / PE / A&P into encounter template fields Listed in athenaConnect marketplace
MEDITECH Bidirectional; specific protocol not publicly documented (FHIR R4 inferred) Commure Ambient embedded natively in MEDITECH Expanse Now (mobile) Commure joined MEDITECH Alliance Mar 2025; Expanse Now embed live Sept 2025
The SMART on FHIR launch path

When the Augmedix app launches inside Epic Hyperspace/Hyperdrive, it receives the encounter ID, patient ID, and clinician ID via the Epic-issued SMART launch token, along with OAuth 2.0 authorization scopes. The clinician authenticates once to Epic (commonly through Imprivata-backed IdP); Augmedix inherits the authenticated context. No separate Augmedix login required in this flow.

EHR-agnostic fallback

When EHR integration cannot be configured (for organizational or technical reasons), clinicians can manually copy-paste notes from the Augmedix Customer Portal into their EHR. This degrades the feature set: no automatic structured data upload, no pre-population of pre-visit context. The Go product schedule names this explicitly as a supported degraded mode.

Chapter VI

The Multi-Pass, Multi-LLM Pipeline

Augmedix’s key architectural differentiator from single-LLM ambient-scribe competitors is its multi-model, multi-pass design: section-specific specialized models drafting different parts of the note (HPI, PE, A&P), with cross-checking between models to instill clinical accuracy and physician trust. The system was an early-tester of MedLM at its December 2023 launch and integrates Augmedix-fine-tuned models on the same Vertex AI substrate.

Diagram 6 — Multi-Pass Generation
graph TD
    AUDIO["Encounter Audio"]
    STT["Google STT
(diarized transcript)"] CTX["Context Retrieval
(prior notes, problem list,
meds via FHIR)"] subgraph PASS["Section-Specific LLM Passes"] HPI["HPI Pass
(history of
present illness)"] PE["PE Pass
(physical exam)"] AP["A&P Pass
(assessment + plan)"] ROS["ROS Pass
(review of systems)"] end XCHECK["Cross-Check Pass
(model-vs-model
quality assurance)"] STRUCT["Structured Data
Extraction
(ICD-10, meds, labs)"] ASSEMBLY["Note Assembly
(section stitching)"] QA["NoteBuilder QA
(human or AI tier)"] AUDIO --> STT STT --> CTX CTX --> HPI CTX --> PE CTX --> AP CTX --> ROS HPI --> XCHECK PE --> XCHECK AP --> XCHECK ROS --> XCHECK XCHECK --> ASSEMBLY XCHECK --> STRUCT ASSEMBLY --> QA style AUDIO fill:#6b8e6f,stroke:#2d4a3e,color:#2a1f1a style STT fill:#4a6b8a,stroke:#1e3a5f,color:#f4e4c1 style CTX fill:#4a6b8a,stroke:#1e3a5f,color:#f4e4c1 style HPI fill:#8b3a4a,stroke:#722f37,color:#f4e4c1 style PE fill:#8b3a4a,stroke:#722f37,color:#f4e4c1 style AP fill:#8b3a4a,stroke:#722f37,color:#f4e4c1 style ROS fill:#8b3a4a,stroke:#722f37,color:#f4e4c1 style XCHECK fill:#722f37,stroke:#2a1f1a,color:#f4e4c1 style STRUCT fill:#c9b87a,stroke:#704214,color:#2a1f1a style ASSEMBLY fill:#704214,stroke:#4a3728,color:#f4e4c1 style QA fill:#a8533a,stroke:#722f37,color:#f4e4c1
Diagram 7 — Fine-Tuning Feedback Loop
graph LR
    NOTES["70K Notes/Week
Production Corpus"] SME["Specialty SMEs
(rubric scoring,
50+ specialties)"] QUAL["Quality Scoring
(internal rubrics)"] TUNE["Vertex AI
Fine-Tuning Job"] BASE["MedLM Base /
MedGemma / Gemini"] SPEC["Specialty-Specific
Fine-Tuned Models"] DEPLOY["Vertex AI
Endpoint Deploy"] PROD["Production Inference
(Go / Assist / Live)"] NOTES --> SME SME --> QUAL QUAL --> TUNE BASE --> TUNE TUNE --> SPEC SPEC --> DEPLOY DEPLOY --> PROD PROD --> NOTES style NOTES fill:#c9b87a,stroke:#704214,color:#2a1f1a style SME fill:#a8533a,stroke:#722f37,color:#f4e4c1 style QUAL fill:#a8533a,stroke:#722f37,color:#f4e4c1 style TUNE fill:#8b3a4a,stroke:#722f37,color:#f4e4c1 style BASE fill:#8b3a4a,stroke:#722f37,color:#f4e4c1 style SPEC fill:#722f37,stroke:#2a1f1a,color:#f4e4c1 style DEPLOY fill:#4a6b8a,stroke:#1e3a5f,color:#f4e4c1 style PROD fill:#5a4633,stroke:#2a1f1a,color:#f4e4c1

MedLM · The early bet

Announced December 11, 2023 by Google. Built on Med-PaLM 2 (which scored 86.5% on US medical licensing exam questions). Augmedix was an early tester alongside HCA Healthcare and BenchSci. Now integrated across all Augmedix products via Vertex AI inference.

MedGemma · The next generation

Google’s subsequent medical foundation model, based on Gemini rather than PaLM. Confirmed in the June 2024 Google Dev article as actively handling healthcare-specific language understanding in the Augmedix pipeline.

Gemini · The multimodal layer

Underpins multimodal processing in the stack. Where audio characteristics, room context, or other non-text signals become relevant, Gemini provides the multimodal LLM backbone.

Augmedix proprietary models

Specialty-specific fine-tuned models trained on Augmedix’s 70K-notes/week corpus across 50+ specialties. Quality is scored against internal rubrics with feedback from human SMEs. This is the loop that turns operational scale into model quality.

Cross-check pass

Augmedix’s own framing: “By utilizing multiple models that cross-check each other, Augmedix is delivering high quality notes that instill trust among clinicians.” A second model reviews a first model’s output before assembly.

Structured-data extraction

A separate NLP path generates structured outputs — ICD-10 codes, medications with dosage/frequency, labs, A&P entries — in parallel with the narrative note. These feed both EHR discrete fields and the BigQuery analytics warehouse.

Accuracy benchmark

The HCA Healthcare pilot reported a clinical documentation F1 score > 0.8 with 1,400+ physicians piloting across 50+ hospitals. This indicates clinical note accuracy is competitive but not yet signature-ready for every specialty — which is why the Assist and Live human-review tiers continue to exist alongside fully autonomous Go.

Chapter VII

Human in the Loop

The Medical Documentation Specialist (MDS) workforce — centred in Dhaka, with additional vendor capacity in India — is the operational story behind Augmedix’s reputation for quality. At peak (pre-acquisition) the Bangladesh subsidiary employed 800+ staff including 500+ scribes, operating from HIPAA-secure rooms with three-month paid training programs. The architectural question is not “is there a human?” but “where in the workflow does the human sit?” — and that answer differs by product tier.

Diagram 8 — Workflow Router by Tier
flowchart LR
    ENC["Encounter Audio
+ AI Draft"] ROUTER{"Workflow Router
(by product tier)"} LIVEPATH["Live Tier
100% Human Review
(real-time synchronous)"] NOTESPATH["Notes / Assist Tier
AI Drafts +
MDS Polish"] GOPATH["Go / Go ED Tier
AI Only +
Physician Sign-Off"] MDSLIVE["MDS in HIPAA Room
(watches encounter live,
NoteBuilder)"] MDSNOTES["MDS Reviews Draft
(NoteBuilder + coding,
within 1 hour)"] PHYS["Physician
Review + Sign"] EHRPEND["EHR
Pending Status"] EHRSIGN["EHR
Final Signed"] ENC --> ROUTER ROUTER --> LIVEPATH ROUTER --> NOTESPATH ROUTER --> GOPATH LIVEPATH --> MDSLIVE NOTESPATH --> MDSNOTES GOPATH --> PHYS MDSLIVE --> EHRPEND MDSNOTES --> EHRPEND PHYS --> EHRSIGN EHRPEND --> EHRSIGN style ENC fill:#6b8e6f,stroke:#2d4a3e,color:#2a1f1a style ROUTER fill:#704214,stroke:#4a3728,color:#f4e4c1 style LIVEPATH fill:#a8533a,stroke:#722f37,color:#f4e4c1 style NOTESPATH fill:#a8533a,stroke:#722f37,color:#f4e4c1 style GOPATH fill:#8b3a4a,stroke:#722f37,color:#f4e4c1 style MDSLIVE fill:#a8533a,stroke:#722f37,color:#f4e4c1 style MDSNOTES fill:#a8533a,stroke:#722f37,color:#f4e4c1 style PHYS fill:#5a4633,stroke:#2a1f1a,color:#f4e4c1 style EHRPEND fill:#8b6914,stroke:#4a3728,color:#f4e4c1 style EHRSIGN fill:#8b6914,stroke:#4a3728,color:#f4e4c1
The offshore PHI question

HIPAA does not prohibit offshore PHI access, provided the Business Associate Agreement covers the subcontractor and appropriate technical and physical safeguards are in place. Augmedix’s BAA with each hospital covers its scribe operations as a downstream subcontractor relationship; MDSs operate from HIPAA-secure rooms with controlled access, no personal devices, and clean-desk policies. MDSs receive credentials directly into the customer’s EHR — meaning they upload pending notes under the same credential governance as the host health system’s own staff.

NoteBuilder · The MDS workstation

The internal tool every human-tier scribe uses. Dynamic templates filtered by specialty, visit type, and complaint. Intelligent auto-completion narrows options as the scribe types. Integrated medication datasets with dosages and frequencies. Real-time AI assistance generating HPI, PE, A&P sections.

Bangladesh primary site

Augmedix Bangladesh: 800+ total staff at peak, 500+ scribes, three-month paid MDS training program. Company leadership signaled plans for a 7,000-person Bangladesh workforce over 5 years (as of ~2022).

India vendor capacity

India scribes operate through vendor relationships rather than a direct subsidiary. At one point approximately 70% of Remote Documentation Specialists resided in India alongside the Bangladesh subsidiary’s direct employees.

Five-country footprint

US (San Francisco HQ), Bangladesh (Dhaka, primary), India, Sri Lanka, Dominican Republic. The geographic spread provides 24-hour follow-the-sun coverage for Live encounters across US time zones.

Downstream BAA chain

The legal structure: Hospital signs BAA with Augmedix → Augmedix flows the BAA terms down to its offshore subsidiary and vendor scribes. MDSs are bound by the same PHI handling rules as on-shore Business Associate staff. Augmedix’s HITRUST r2 covers this chain.

Two-way clinician chat

In Live, the MDS and clinician communicate via text during the encounter — the MDS can ask for a clarification, the clinician can flag an order to add. This is the operational mechanism that lets a remote scribe stay in sync with a fast-moving in-room visit.

The transition story

The deliberate reduction in per-note human labour is the central architectural story of Augmedix’s ten years:

  • 2013–2021 · Google Glass + full human scribe. Scribes watched the entire encounter through the clinician’s point-of-view stream and wrote the note from scratch. AI minimal. Value was human labour arbitrage.
  • 2021–2023 · NoteBuilder AI assistance + human scribe. AI pre-drafts sections; scribes correct and finalize. The S-1 explicitly called out NoteBuilder as central to the plan to “reduce the level of effort required by remote documentation specialists.”
  • 2023–present · Go (pure AI) + Assist (human polish on AI draft) + Live (real-time human). The Go tier represents the shift from human-first to AI-first documentation. Partly accelerated by COVID-era constraints that made in-person scribing infeasible.

Per Commure’s acquisition press release, all three tiers continue to receive investment post-acquisition — the architecture isn’t “replace humans”, it’s “tier humans by where they add the most value relative to the AI baseline.”

The Glass-to-smartphone transition is worth dwelling on. Google discontinued Glass Enterprise on 15 March 2023 (support ended 15 September 2023). Augmedix had already diversified to smartphone capture — the Go mobile app launched 27 September 2023 — so the Glass sunset did not materially disrupt operations. The “Content Capture Device” in the Live product schedule is now a smartphone, not a wearable. The wearable era of Augmedix is over.
Chapter VIII

Scribe Operations — A Closer Look

The scribe operation is the human substrate that has carried Augmedix for a decade. Encounter audio leaves the clinician’s phone, gets routed by tier, and lands in a queue watched by a Medical Documentation Specialist (MDS) sitting in a HIPAA-secure room — mostly in Dhaka, with vendor capacity in India and follow-the-sun coverage from Sri Lanka and the Dominican Republic. NoteBuilder is the workstation. The strategic arc has been steady: more AI in the loop, less labour per note — without dismantling the human review tier, because Live + Assist are precisely what differentiate Augmedix from pure-AI competitors.

500+
Bangladesh Scribes at Peak
800+
Bangladesh Staff (Total)
~70%
RDS Resident in India (Vendor)
5
Country Footprint
3 mo
Paid MDS Training Program
7,000
Stated Bangladesh Target (5-yr)
Diagram 9 — Scribe Routing & Queue
graph TD
    AUDIO["Encounter Audio
(iOS app / Bluetooth mic)"] ROUTE{"Tier Router"} LIVEQ["Live Queue
(synchronous,
follow-the-sun)"] ASSISTQ["Assist Queue
(post-encounter,
1-hour SLA)"] GOSKIP["Go · No Human"] subgraph SCRIBE_GEO["Scribe Geographies (HIPAA-secure rooms)"] BD["Bangladesh
(Dhaka subsidiary,
500+ MDSs)"] IN["India
(vendor partners,
~70% of RDSs)"] LK["Sri Lanka
(coverage)"] DR["Dominican Republic
(coverage)"] end NB["NoteBuilder
(MDS workstation)"] QA["QA · Quality Review
(internal rubric)"] EHR_WRITE["EHR Write-Back
(MDS-credentialed,
pending status)"] PHYS_SIGN["Physician Review & Sign"] AUDIO --> ROUTE ROUTE --> LIVEQ ROUTE --> ASSISTQ ROUTE --> GOSKIP LIVEQ --> BD LIVEQ --> IN LIVEQ --> LK LIVEQ --> DR ASSISTQ --> BD ASSISTQ --> IN BD --> NB IN --> NB LK --> NB DR --> NB NB --> QA QA --> EHR_WRITE GOSKIP --> EHR_WRITE EHR_WRITE --> PHYS_SIGN style AUDIO fill:#6b8e6f,stroke:#2d4a3e,color:#2a1f1a style ROUTE fill:#704214,stroke:#4a3728,color:#f4e4c1 style LIVEQ fill:#a8533a,stroke:#722f37,color:#f4e4c1 style ASSISTQ fill:#a8533a,stroke:#722f37,color:#f4e4c1 style GOSKIP fill:#8b3a4a,stroke:#722f37,color:#f4e4c1 style BD fill:#c9b87a,stroke:#704214,color:#2a1f1a style IN fill:#c9b87a,stroke:#704214,color:#2a1f1a style LK fill:#c9b87a,stroke:#704214,color:#2a1f1a style DR fill:#c9b87a,stroke:#704214,color:#2a1f1a style NB fill:#6b4423,stroke:#4a3728,color:#f4e4c1 style QA fill:#6b4423,stroke:#4a3728,color:#f4e4c1 style EHR_WRITE fill:#8b6914,stroke:#4a3728,color:#f4e4c1 style PHYS_SIGN fill:#5a4633,stroke:#2a1f1a,color:#f4e4c1

HIPAA-secure rooms

MDSs operate from dedicated facilities with controlled access, no personal devices, and clean-desk policies. The room is the physical-safeguard layer that lets PHI flow across borders under the BAA chain.

NoteBuilder · The MDS IDE

Dynamic templates filtered by specialty, visit type, and complaint. Intelligent auto-completion narrows options as the scribe types. Integrated medication datasets (dosage, frequency, side effects). Real-time AI assistance drafting HPI, ROS, PE, A&P.

Downstream BAA chain

Hospital signs BAA with Augmedix → Augmedix flows BAA terms down to its Dhaka subsidiary and India vendor MDSs. Direct-EHR-access scribes are bound by the same PHI-handling rules as on-shore Business Associates.

COVID pivot to remote

In-person scribing became infeasible in 2020. The Bangladesh secure-room model was already remote by design; the constraint accelerated the smartphone capture push that ultimately replaced Google Glass and laid the groundwork for Go (pure-AI) two years later.

QA & calibration loop

MDS output feeds a continuous quality rubric: discrepancies between AI drafts and human polish become training signal for the next fine-tune. The S-1 explicitly framed NoteBuilder as “reducing the level of effort required by remote documentation specialists.”

From labour arbitrage to AI augmentation

2013–2021 the value was wage arbitrage (US physicians, Bangladesh scribes). 2021–2023 the value was hybrid AI+human. 2023+ the value is tiered: pure AI (Go), AI + human polish (Assist), AI + real-time human (Live). Same workforce, different leverage.

The Bangladesh subsidiary is the only tier-1 Augmedix scribe footprint that is direct-employed. India, Sri Lanka, and the Dominican Republic operate through vendor relationships. Post-Commure, Bangladesh remains the centre of mass — the 7,000-person workforce target announced before the acquisition has not been publicly retracted.
Chapter IX

Compliance & PHI Handling

Augmedix’s compliance posture rests on three layered BAAs: Google Cloud’s BAA covering Vertex AI, Speech-to-Text, GKE, Spanner, BigQuery, Pub/Sub, and the Healthcare API; Augmedix’s BAA with each covered-entity customer; and downstream BAAs covering the offshore MDS subcontractors. HITRUST r2 (the most rigorous HITRUST tier) was earned August 2023 and explicitly scopes Augmedix Live, Augmedix Notes, and the supporting cloud services. There are no documented data breaches, no OCR enforcement actions, and no FTC matters. SOC 2 is not publicly confirmed.

Certification / Posture Date / Scope Notes
HITRUST r2 CertifiedAugust 2023 · 2-yearScope: Augmedix Live, Augmedix Notes, supporting cloud services. The risk-based, two-year tier — HITRUST’s most rigorous, with independent assessor verification.
HIPAA BAA · Google CloudOngoingCovers Vertex AI, Speech-to-Text, GKE, Cloud Spanner, BigQuery, Pub/Sub, Cloud Storage, Healthcare API. The legal basis for identifiable PHI flowing through MedLM/MedGemma inference.
HIPAA BAA · CustomerPer-customerEach hospital / health-system customer signs a BAA with Augmedix. Standard covered-entity ↔ business-associate structure.
Downstream BAAs · MDSsPer-relationshipAugmedix Bangladesh subsidiary and India vendors flow down the BAA terms. Direct-EHR-access scribes are bound under the same credential governance.
ISO 27001 / 27017 / 27018GCP infra layerApplies to Google Cloud substrate. Augmedix application-layer ISO posture not publicly documented separately from HITRUST.
SOC 2[unconfirmed]Not mentioned in any public Augmedix source. HITRUST r2 is positioned as the company’s primary third-party assurance.
GDPRNot applicable (US-only)No evidence of EU deployments. Customer base is US health systems.
Documented breachesNone foundNo reported Augmedix breach in HIPAA Journal databases or public sources as of May 2026.
FTC actionsNone foundNo FTC enforcement matters or consent orders documented.
Diagram 10 — The BAA Boundary
flowchart LR
    CLIN["Clinician
(Covered Entity staff)"] APP["Augmedix App
(iOS / Bluetooth mic)"] subgraph BAA_GCP["GCP HIPAA BAA"] STT2["Speech-to-Text"] VAI2["Vertex AI
(MedLM, MedGemma,
Gemini)"] STORE["Spanner / BigQuery /
Healthcare API"] end subgraph BAA_AUG["Augmedix-MDS BAA (downstream)"] MDS2["MDS
(NoteBuilder,
HIPAA-secure room)"] end subgraph BAA_HOSP["Hospital BAA · EHR Tenant"] EHR2["Epic / Oracle Health /
athenahealth / MEDITECH"] end CLIN --> APP APP --> STT2 STT2 --> VAI2 VAI2 --> STORE STORE --> MDS2 MDS2 --> EHR2 VAI2 --> EHR2 style CLIN fill:#5a4633,stroke:#2a1f1a,color:#f4e4c1 style APP fill:#6b8e6f,stroke:#2d4a3e,color:#2a1f1a style STT2 fill:#4a6b8a,stroke:#1e3a5f,color:#f4e4c1 style VAI2 fill:#8b3a4a,stroke:#722f37,color:#f4e4c1 style STORE fill:#4a6b8a,stroke:#1e3a5f,color:#f4e4c1 style MDS2 fill:#a8533a,stroke:#722f37,color:#f4e4c1 style EHR2 fill:#8b6914,stroke:#4a3728,color:#f4e4c1
Why Vertex AI is usable on PHI

Google’s HIPAA BAA covers Vertex AI as a HIPAA-eligible service — the contract commits that Vertex AI does not use customer data, prompts, or responses to improve or train Google’s foundation models. This is what lets identifiable patient audio, transcripts, and notes flow through MedLM and MedGemma inference inside Augmedix’s pipeline without violating the minimum-necessary rule. Without that BAA scope, the same architecture would require de-identification before inference.

Patient consent

Customers are contractually required to obtain patient consent for audio recording “to the extent required under applicable law” — state wiretapping laws vary (one-party vs. two-party consent). HCA ED pilot reported 99% patient consent acceptance.

Audio retention

Per the Google Developer article: encounter audio and note data retained one week after EMR transmission, then permanently deleted. The general privacy policy speaks to “up to six years after account closure” for user account data — the audio window is narrower. [the one-week figure is from a blog, not a formal data-retention policy; treat as unconfirmed for audio specifically]

PHI de-identification path

Optional de-identified structured-data export for analytics customers whose data-lake use cases don’t require PHI. Mechanism (Cloud DLP, custom NER, or manual review) is not specified [unconfirmed].

Encryption

HITRUST r2 baseline implies AES-256 at rest, TLS 1.2+ in transit. The privacy policy speaks to “authentication, encryption, backups, and access controls” without naming protocols. CMEK / Confidential VMs / VPC Service Controls usage not publicly documented [unconfirmed].

Chapter X

Multi-Tenant Isolation on GCP

Augmedix’s multi-tenancy story is partly documented and partly inferred. The enterprise data-integration tier is unambiguous: Augmedix publishes encounter JSON into a customer-owned Pub/Sub topic that lands in a customer-owned encounter_service on the customer’s GCP project — a bring-your-own-project pattern. The inference pipeline (STT, Vertex AI, audio buffers) runs in Augmedix’s own GCP estate, where logical isolation is via IAM, encounter identifiers, and dataset-level permissions in BigQuery. The HITRUST r2 assessor reviewed the access-control domain, which means tenant separation in shared infrastructure was independently validated — but the specific control surface (per-tenant CMEK, VPC SC, project structure) is not published.

Diagram 11 — Tenant Isolation Pattern
graph TB
    subgraph AUG_PROJ["Augmedix-owned GCP Project (shared inference plane)"]
        STT3["Speech-to-Text"]
        VAI3["Vertex AI
(MedLM / MedGemma)"] ENC_SVC["encounter_service
(per-tenant IAM)"] SPAN3["Spanner
(encounter table,
tenant key in row)"] end subgraph TEN_A["Customer A · Their GCP Project"] PSA["Pub/Sub topic A"] SVCA["encounter_service A
(GKE / Cloud Run)"] BQA["BigQuery A
(customer-owned)"] end subgraph TEN_B["Customer B · Their GCP Project"] PSB["Pub/Sub topic B"] SVCB["encounter_service B
(GKE / Cloud Run)"] BQB["BigQuery B
(customer-owned)"] end IAM["Cloud IAM · per-tenant
service accounts &
dataset permissions"] AUDIT["Cloud Audit Logs
[unconfirmed scope]"] STT3 --> VAI3 VAI3 --> ENC_SVC ENC_SVC --> SPAN3 ENC_SVC --> PSA ENC_SVC --> PSB PSA --> SVCA PSB --> SVCB SVCA --> BQA SVCB --> BQB IAM -.-> ENC_SVC IAM -.-> PSA IAM -.-> PSB AUDIT -.-> ENC_SVC style STT3 fill:#4a6b8a,stroke:#1e3a5f,color:#f4e4c1 style VAI3 fill:#8b3a4a,stroke:#722f37,color:#f4e4c1 style ENC_SVC fill:#704214,stroke:#4a3728,color:#f4e4c1 style SPAN3 fill:#4a6b8a,stroke:#1e3a5f,color:#f4e4c1 style PSA fill:#c9b87a,stroke:#704214,color:#2a1f1a style PSB fill:#c9b87a,stroke:#704214,color:#2a1f1a style SVCA fill:#704214,stroke:#4a3728,color:#f4e4c1 style SVCB fill:#704214,stroke:#4a3728,color:#f4e4c1 style BQA fill:#4a6b8a,stroke:#1e3a5f,color:#f4e4c1 style BQB fill:#4a6b8a,stroke:#1e3a5f,color:#f4e4c1 style IAM fill:#8b6914,stroke:#4a3728,color:#f4e4c1 style AUDIT fill:#6b4423,stroke:#4a3728,color:#f4e4c1
Aspect Confirmed / Inferred Detail
Tenant unitConfirmedHealth system / hospital. Each tenant is one BAA-signed customer; tenants are not per-clinician or per-encounter.
Enterprise data pathConfirmed (Google Dev blog)Bring-your-own-project: customer-owned Pub/Sub topic, encounter_service, Spanner/BigQuery in the customer’s GCP estate. Augmedix publishes; customer owns the lake.
Inference pipeline isolationInferredShared Augmedix-owned GCP infrastructure (STT, Vertex AI, audio buffers) with logical isolation via IAM service accounts, encounter IDs (_encounterID, _noteID), and dataset-level BigQuery permissions. [unconfirmed specifics]
Spanner partitioningInferredEncounter rows carry tenant identifiers (encounter ID, note ID, timestamps); per-tenant IAM enforces read scope. [unconfirmed schema]
CMEK (per-tenant keys)UnconfirmedNot publicly documented. Standard GCP-healthcare best practice supports CMEK via Cloud KMS; whether Augmedix offers per-tenant CMEK is not disclosed.
VPC Service ControlsUnconfirmedGCP Healthcare best-practice pattern: wrap Vertex AI, Healthcare API, BigQuery in a service perimeter. Augmedix has not published its perimeter posture.
Audit loggingImpliedHIPAA requires; HITRUST r2 covers. Specific architecture (Cloud Logging exports, SIEM integration) not disclosed [unconfirmed].
Tenant credential modelConfirmedEach customer onboarded with separate EHR credentials and Augmedix account credentials. MDSs with EHR write access operate under the host health system’s own credential governance.
A substrate-comparison observation

The most architecturally interesting feature is the hybrid tenancy at the data-lake boundary: the inference plane is multi-tenant under Augmedix’s IAM, but the analytics plane is single-tenant inside each customer’s GCP project. This pushes the “data exfiltration boundary” out of Augmedix’s estate and into the customer’s — reducing Augmedix’s blast radius and giving each customer full control over their longitudinal patient record store. It also means BigQuery scale, costs, and governance live with the customer, not Augmedix.

The bring-your-own-project pattern is plausible only at health-system scale — smaller customers presumably consume a shared Augmedix-hosted analytics surface. The 99.0% uptime SLA on the Pub/Sub export service applies to Augmedix’s publish side; the customer’s subscribe side is on them.
Chapter XI

The Commure Era

In July 2024 Commure — the General-Catalyst-backed healthcare-AI roll-up — signed a definitive agreement to take Augmedix private at $2.35/share, a $139M all-cash deal at a 169% premium to the 30-day VWAP. Stockholders approved September 27; closing was October 2, 2024. Augmedix delisted from Nasdaq (ticker AUGX) and became a wholly-owned Commure subsidiary. By mid-2025 Commure had raised $200M from General Catalyst’s Customer Value Fund; in May 2026 a $70M round closed at a $7B post-money valuation. The Augmedix ambient AI is now the front end of Commure’s end-to-end clinical platform — rebranded “Commure Ambient.”

Diagram 12 — Commure Acquisition Timeline
timeline
    title Augmedix → Commure · The Acquisition Arc
    2012 : Founded at Stanford
         : Glass + remote scribes thesis
    2014 : Commercial launch
         : Dignity / Sutter / CHI customers
    2020 : Reverse-merger IPO
         : Nasdaq AUGX · $40M raise
    2023 : Glass deprecation
         : MedLM early-tester (Dec)
         : Augmedix Go launches (Sept)
    2024 : Commure agreement (Jul)
         : Take-private closes (Oct 2)
         : $139M · 169% premium
    2025 : Vizient contract (Jan)
         : MEDITECH Alliance (Mar)
         : Expanse Now embed (Sept)
    2026 : Commure $7B valuation (May)
         : Brand — Commure Ambient
                

What Commure brings to the architecture

Commure’s strategic thesis is vertical integration of the clinical revenue cycle: ambient documentation generates the note, autonomous coding codes it, CDI improves it, RCM bills it, and an agentic layer automates referrals, scheduling, and prior auth from the same clinical context. Augmedix supplies the documentation layer at the front of that stack. The post-acquisition footprint is materially larger than Augmedix achieved as a standalone public company.

HCA Healthcare · Exclusive partner

October 2024 press release named HCA as exclusive development partner for the multi-specialty ambient platform. 1,400+ physicians piloting across 50+ hospitals, clinical documentation F1 score > 0.8. Ian Shakil leads the HCA integration as Chief Strategy Officer at Commure.

Vizient contract

Awarded January 2025. Vizient is the healthcare GPO serving 65%+ of US acute-care providers, 97% of academic medical centers, 35% of non-acute, and $140B+ annual purchasing. Covers Go, Assist, and Live tiers — a distribution lever Augmedix never had alone.

MEDITECH Expanse Now embed

September 2025: Commure Ambient AI embedded directly in MEDITECH’s Expanse Now mobile EHR. Notes flow into discrete EHR sections across ED, inpatient, and ambulatory contexts — no copy-paste. Commure joined the MEDITECH Alliance in March 2025.

The integration thesis

Commure CEO Tanay Tandon framed it as “language models that transcribe appointments, autonomously code them, and supercharge back-office operations for billing teams.” Augmedix is the transcribe step; Athelas (acquired Oct 2023) is the code step; Commure Agents (launched June 2025) is the automation layer.

Commure platform scale

By 2025: 130+ health systems, 350,000+ clinicians, 60+ EHR integrations, $25B+ annual claims processed, 3,000+ sites of care, 40M+ ambient appointments / year. KLAS Research named Commure 2025 Market Leader in Ambient AI.

Capital trajectory

Oct 2023: Commure+Athelas merger, $70M from General Catalyst at ~$6B combined valuation. Oct 2024: Augmedix acquisition $139M. Dec 2024: Memora Health acquisition. Mid-2025: $200M from GC’s Customer Value Fund. May 2026: $70M round — $7B post-money with Sequoia, Morgan Stanley, Kirkland & Ellis participating.

Why the 169% premium

Augmedix was growing (~40% YoY revenue in Q1 2024) but had never reached profitability and had cut its full-year 2024 revenue guidance to $52–55M, sending shares down 40%. The 169% premium over the 30-day VWAP indicates Commure valued the technology, the GCP/MedLM partnership, the offshore scribe workforce, and the HCA + customer relationships well above the depressed public-market price. From Commure’s side, owning the ambient documentation layer was the precondition for the vertically-integrated thesis — you cannot autonomously code a note you don’t own.

Chapter XII

AI Model Lineage · The Scribe-to-LLM Arc

Chapter XI traced the corporate timeline. This one cuts the same period along a different axis: how the AI substrate evolved from minimal NLP layered onto Google Glass + human scribes, through proprietary specialty fine-tunes inside NoteBuilder, to a multi-LLM Vertex AI pipeline running MedLM, MedGemma, and Gemini in parallel passes. The architectural inflection is 2022–2023: capable LLMs make Go (pure-AI) viable, and the scribe role shifts from primary author to QA reviewer.

Diagram 13 — AI / Capture Lineage
timeline
    title Capture Modality + Model Lineage
    2012 : Founding · Google Glass + remote scribes
         : Minimal NLP, human-authored notes
    2018 : Smartphone capture added
         : Bluetooth lab-coat mic option
    2020 : IPO (Nasdaq AUGX)
         : Audio + ASR mature
    2021 : NoteBuilder launches
         : ASR + NLP + suggestion templates
    2022 : Audio AI work scaled
         : Specialty fine-tunes
         : Augmedix Prep launches (Sept)
    2023 : Glass sunset (Mar)
         : MedLM early-tester (Dec)
         : Augmedix Go launches (Sept)
    2024 : MedGemma + Gemini in pipeline
         : Go ED launches (Apr)
         : Multi-LLM cross-check live
    2025 : MEDITECH Expanse Now embed (Sept)
         : Commure Agents adjacent
    2026 : Commure Ambient brand
         : Vertex AI / Healthcare API substrate
                

Three architectural inflections

  • 2021 · NoteBuilder. First “AI in the loop” stage. ASR + NLP surfaced suggestions; humans still authored. The S-1 framed it as a labour-reduction tool, not a replacement.
  • 2023 · The LLM inflection. Capable LLMs (Med-PaLM 2 lineage → MedLM) made Go viable. The role of the human shifted from primary author to QA reviewer (Assist) or was removed entirely (Go). HCA’s $12M equity investment in April 2023 funded this pipeline.
  • 2024 · Multi-LLM cross-checking. The Google Developer article confirms MedLM + MedGemma + Gemini running together, with specialty fine-tunes and an explicit cross-check pass. This is Augmedix’s differentiator versus single-LLM ambient scribe competitors (Nuance DAX, Suki, Nabla, Abridge, Freed).
A subtle architectural point: the offshore scribe workforce wasn’t replaced as AI improved — it was tiered. Same MDSs, same NoteBuilder, but now distributed across three product price-points (Live, Assist, Go). The 500+ Bangladesh scribes are still there; they just touch fewer notes per encounter.
Chapter XIII

Acronyms & Sources

A glossary of every acronym used in this map — from the clinical (HPI, ROS, A&P) through the regulatory (HIPAA, HITRUST, BAA, PHI) to the GCP-architectural (GKE, FHIR, HL7). Followed by a consolidated source list and a brief provenance note.

AI

Artificial Intelligence — here, the umbrella for ASR, NLP, and LLM components in Augmedix’s pipeline.

A&P

Assessment & Plan — the section of a clinical note that records the clinician’s diagnostic reasoning and next steps.

ASR

Automatic Speech Recognition — converts spoken audio into text. Augmedix uses Google Cloud Speech-to-Text.

AVS

After-Visit Summary — the patient-facing document summarizing the encounter, instructions, and follow-ups.

BAA

Business Associate Agreement — the HIPAA contract between a Covered Entity (hospital) and a Business Associate (Augmedix, Google Cloud, MDSs) covering PHI handling.

BQ · BigQuery

Google Cloud’s serverless analytics data warehouse. Used here for longitudinal patient analytics combining Augmedix encounter data with EHR data.

CDI

Clinical Documentation Improvement — the process of refining clinical notes to support accurate coding and reimbursement.

CMEK

Customer-Managed Encryption Keys — GCP feature allowing customers to control encryption keys via Cloud KMS. Usage at Augmedix is unconfirmed.

CPT

Current Procedural Terminology — the AMA-maintained code set for medical procedures used in billing.

DLP

Data Loss Prevention — Google Cloud DLP is the de-identification / sensitive-data-detection service referenced as a complementary PHI scrubbing layer.

E&M

Evaluation & Management — CPT code category for clinician cognitive work (the most common physician billing codes).

ED

Emergency Department — the use case behind Augmedix Go ED (launched April 2024), purpose-built for high-acuity, multi-touch ED visits.

EHR

Electronic Health Record — Epic, Oracle Health, athenahealth, MEDITECH are the main targets for Augmedix write-back.

F1

F1 score — the harmonic mean of precision and recall. HCA pilot reported clinical documentation F1 > 0.8 on Augmedix output.

FHIR

Fast Healthcare Interoperability Resources — HL7’s modern API standard for healthcare data. Used in Augmedix’s EHR integrations and the Google Healthcare API.

FTC

Federal Trade Commission — the US consumer-protection regulator. No FTC actions documented against Augmedix.

GCP

Google Cloud Platform — the cloud substrate underneath every Augmedix component from STT through analytics.

GKE

Google Kubernetes Engine — the managed Kubernetes service hosting the encounter_service that subscribes to Pub/Sub.

GPO

Group Purchasing Organization — e.g. Vizient, which negotiates pricing on behalf of member health systems.

HCA

HCA Healthcare — the US hospital operator (1,400+ physicians piloting, 50+ hospitals). Augmedix’s exclusive ambient AI development partner post-Commure-acquisition.

HCC

Hierarchical Condition Categories — CMS risk-adjustment coding methodology supported by Augmedix Live/Assist coding workflows.

HDEv2

Healthcare Data Engine v2 — Google’s FHIR R4 harmonization product, in private preview as of June 2024. Aggregates multi-source healthcare data.

HIPAA

Health Insurance Portability and Accountability Act — US law governing PHI handling. Underlies every BAA in Augmedix’s architecture.

HITRUST

Health Information Trust Alliance — the framework. Augmedix holds HITRUST r2 (the rigorous risk-based, 2-year tier) since August 2023.

HL7

Health Level Seven — the standards organization and message format (HL7 v2) used by Augmedix to pull schedules from Epic (ADT/SIU feeds).

HPI

History of Present Illness — the narrative section of a clinical note describing the patient’s current complaint and its evolution.

ICD-10

International Classification of Diseases, 10th rev. — the diagnosis code set generated by Augmedix Go’s structured-data extraction.

IdP

Identity Provider — the SSO authority (often Imprivata in hospital settings) backing EHR-embedded SMART-on-FHIR launches.

IPO

Initial Public Offering — Augmedix uplisted to Nasdaq via reverse merger in December 2020 (ticker AUGX, $40M raised).

JSON

JavaScript Object Notation — the structured payload format Augmedix publishes to customer Pub/Sub topics (transcript, note, structured clinical data).

KPI

Key Performance Indicator — e.g. F1 score > 0.8 on HCA clinical documentation pilot, 99% patient consent acceptance in ED pilot.

LLM

Large Language Model — the model class behind MedLM, MedGemma, and Gemini in Augmedix’s multi-pass pipeline.

LOS

Length of Stay — an inpatient metric supported by Assist/Live coding workflows.

M&A

Mergers & Acquisitions — the Commure–Augmedix transaction (Oct 2024) is the relevant one here.

MDS

Medical Documentation Specialist — Augmedix’s job title for the trained remote scribe (often used interchangeably with “scribe” or RDS).

MedGemma

Google’s medical-domain Gemini-derivative LLM. Confirmed active in Augmedix’s Clinical Notes Engine pipeline.

MedLM

Google’s medical-tuned foundation-model suite (Med-PaLM 2 lineage). Augmedix was an early tester (Dec 2023); MedLM is fine-tuned on Augmedix’s 70K notes/week corpus.

MFA

Multi-Factor Authentication — Augmedix’s public product schedules do not specify MFA requirements; hospital IdPs typically enforce.

NLP

Natural Language Processing — the pipeline step between ASR and LLM generation. Augmedix runs proprietary NLP for entity extraction and section assignment.

OCR

Office for Civil Rights (HHS) — the HIPAA enforcement body. No OCR actions documented against Augmedix.

PE

Physical Exam — the section of a clinical note recording the clinician’s physical examination findings.

PHI

Protected Health Information — the HIPAA-defined identifiable patient data category. Every Augmedix BAA exists to govern PHI flow.

Pub/Sub

Google Cloud Pub/Sub — asynchronous messaging service used to publish structured encounter JSON to customer-owned topics.

QA

Quality Assurance — the human review step inside NoteBuilder where MDSs polish AI drafts and the discrepancies become fine-tune training signal.

RCM

Revenue Cycle Management — the billing and reimbursement workflow Commure layers on top of Augmedix documentation via Athelas.

RDS

Remote Documentation Specialist — alternate term for MDS, used in some Augmedix materials. ~70% of RDSs were resident in India at one point.

ROS

Review of Systems — the section of a clinical note inventorying patient symptoms by organ system.

SLA

Service Level Agreement — Augmedix commits 99.0% uptime on its Pub/Sub enterprise-data-export service.

SMART on FHIR

The OAuth-based EHR-embedded app launch standard used when the Augmedix app launches inside Epic Hyperspace/Hyperdrive.

SOC 2

System and Organization Controls 2 — AICPA assurance framework. Not publicly confirmed for Augmedix; HITRUST r2 is the primary third-party posture.

SPAC

Special Purpose Acquisition Company — the vehicle (Malo Holdings) used in Augmedix’s 2020 reverse-merger IPO.

SSO

Single Sign-On — the hospital IdP authentication pattern (often Imprivata-backed) inherited by Augmedix in SMART-on-FHIR EHR-embedded launches.

STT

Speech-to-Text — Google Cloud’s diarized ASR service. Synonym for ASR in this context.

VPC SC

VPC Service Controls — GCP perimeter-isolation feature commonly used to wrap Vertex AI / Healthcare API in healthcare deployments. Usage at Augmedix unconfirmed.

VWAP

Volume-Weighted Average Price — the 30-day VWAP baseline against which Commure’s 169% acquisition premium was measured.

YoY

Year-over-Year — the growth metric used in Augmedix financial reporting (e.g. +45% YoY revenue growth in 2023, +44% YoY clinician growth).

Sources

Synthesized from publicly available sources, prioritized for technical and architectural detail:

  • Google Developer Community — “The future of clinical data: building an intelligence engine with Augmedix & Google Cloud” (June 2024). The richest single technical source — pipeline diagram, GCP service inventory, retention figures.
  • Augmedix S-1 IPO Prospectus (SEC EDGAR, 2020) — corporate structure, customer concentration, NoteBuilder framing.
  • Augmedix Form 10-K FY2023 and Q4 2023 earnings release (GlobeNewswire) — financial scale, clinicians-in-service, ARR.
  • Augmedix Go and Live product schedules (augmedix.com) — the closest public document to a technical spec.
  • Augmedix Privacy Policy (augmedix.com/privacy-policy/current) — subprocessors, data retention general terms.
  • HITRUST r2 announcement (GlobeNewswire, August 2023) — certification scope.
  • Commure acquisition press releases (July–October 2024) and Hospitalogy analysis (October 2025) — deal terms, integration thesis.
  • HCA partnership (GlobeNewswire, April 2023) and Hospitalogy interview with Ian Shakil — ED pilot details, the “conversational graph” framing.
  • MedLM launch announcement (Google Cloud Blog and Fierce Healthcare, December 2023) — Augmedix as early tester.
  • Oracle Validated Integration announcement (GlobeNewswire, December 2024); MEDITECH Expanse Now embed (Commure press release, September 2025); Vizient contract (Commure press release, January 2025).
  • TBS News Bangladesh — on the Dhaka scribe workforce.
  • Augmedix Epic Showroom listing (id=1103) and Oracle Marketplace listing — marketplace presence.
Provenance

Built for GG Prompts using research from public sources. Augmedix is a real company (now part of Commure, marketed as “Commure Ambient”); this map summarizes publicly available information as of May 2026. Items marked [unconfirmed] are inferred from typical GCP-Healthcare patterns and labelled explicitly. Not affiliated with Augmedix, Commure, Google Cloud, or any health system named on this page.

Diagram
100%
Scroll to zoom · Drag to pan · Esc to close